Method and system for on-line detection of moisture of coal into a furnace based on microwave attenuation phase analysis

Abstract

The application relates to an online detection method and system for coal moisture based on microwave attenuation phase analysis. After obtaining the attenuation and phase shift through double-frequency microwave transmission, the method first constructs a density-independent parameter through the attenuation phase ratio, then estimates the density and coarse moisture through a joint mapping model, and compensates to obtain the density-normalized attenuation and phase. The moisture fluctuation online accumulates the normalized data, locally fits to extract the curvature on the complex plane, forms the coal type dielectric dispersion geometric fingerprint, matches and identifies the coal type with the pre-built library, calls the special model, reconstructs the density-independent feature vector based on the original attenuation phase, and outputs the total moisture. Through the closed loop of density compensation, curvature perception and coal type adaptive switching, the method does not increase the hardware, realizes high-precision online detection at low pre-calibration cost under the condition of multiple coal types.

Description

Technical Field

[0001] This invention belongs to the field of material physical property testing and online analysis technology, and in particular relates to an online detection method and system for moisture content of coal fed into the furnace based on microwave attenuation phase analysis. Background Technology

[0002] In thermal power generation and coal chemical production processes, the total moisture content of the coal fed into the furnace is a key indicator affecting combustion efficiency, boiler heat balance, and fuel cost accounting. To meet the requirements of continuous production for real-time and non-contact operation, microwave transmission method, due to its high sensitivity and penetration for moisture, has been gradually introduced into online moisture detection of coal conveyor belts. The principle is that the amplitude attenuation and phase shift of microwaves after passing through the coal seam are simultaneously affected by factors such as moisture content, coal seam thickness, and bulk density. By establishing a calibration model between these observations and moisture content, moisture inversion can be achieved under certain conditions.

[0003] However, in coal conveyor belt operations, the coal seam is not a homogeneous medium. Belt speed, coal dropping methods, and hopper switching cause rapid fluctuations in coal seam thickness and bulk density over a wide range. Changes in thickness and density contribute to attenuation and phase similar to changes in moisture content. If moisture is directly extrapolated from a single observation, the error will significantly deviate from the allowable range for industrial control. To mitigate the interference of thickness and density, existing technologies have proposed using the ratio of attenuation to phase to form a density-independent function. Taking advantage of the fact that both are approximately proportional to thickness, the thickness factor is eliminated by the ratio, and a moisture calibration curve is constructed based on the weak dependence of the loss tangent on density. However, this ratio method only partially suppresses the influence of density and does not truly remove density fluctuations from the observation. The residual density effect still causes a shift in the complex plane calibration curve, limiting detection accuracy.

[0004] A more prominent problem is that coals of different metamorphic degrees, such as anthracite, bituminous coal, and lignite, exhibit fundamental differences in their matrix adsorption capacity for water molecules, pore structure, and dielectric properties of minerals. These differences are reflected in the curvature of the complex plane calibration curve through the dispersion behavior of the dielectric constant. Existing technologies often treat the complex plane calibration curve simply as a fitting tool between moisture content and measured values, using polynomials or power functions for global fitting, without considering the dielectric dispersion information of the coal type carried by the curve curvature. This leads to the failure of the original calibration model once the coal type is changed due to the mismatch in dielectric response characteristics. To cover multiple coal types, operators have to conduct numerous independent pre-calibration experiments for each possible coal type, collecting hundreds of sets of data on a rotary drum test bench for each coal type, moisture level, and density point. This process is extremely time-consuming and difficult to adapt to frequent changes in coal sources and coal quality drift. In addition, the resonant interference caused by multiple reflections from the conveyor belt's metal components and the coal flow, as well as the drift in the water's dielectric properties caused by changes in ambient temperature, further deteriorate the phase measurement quality, making adaptive measurement for multiple coal types even more difficult.

[0005] In summary, existing microwave transmission moisture detection technologies, when handling online measurements of multiple coal types, lack the interpretation and utilization of the geometric features of complex plane curves. This prevents them from sensing changes in coal type online and automatically switching to the corresponding precise calibration model. They heavily rely on full pre-calibration, making it difficult to balance high accuracy and low maintenance costs in industrial scenarios involving blending and variable coal sources. Therefore, there is an urgent need for a method that, without significantly increasing hardware, extracts feature quantities that can identify the dielectric dispersion characteristics of coal types from dual-frequency microwave transmission signals. This would enable online self-identification of coal types and dynamic matching of calibration models, thereby meeting the accuracy and real-time requirements of online full moisture detection of coal entering the furnace with extremely low pre-calibration costs. Summary of the Invention

[0006] Therefore, it is necessary to provide a method and system for online detection of moisture in coal fed into the furnace based on microwave attenuation phase analysis to address the above-mentioned technical problems.

[0007] In a first aspect, this application provides an online method for detecting the moisture content of coal fed into the furnace based on microwave attenuation phase analysis, including:

[0008] S1. Transmit at least two microwave signals of different frequencies through the coal flow on the coal conveyor belt, receive the microwave signal after penetration through the coal flow, perform orthogonal demodulation and reference comparison on the microwave signal after penetration, and obtain the relative attenuation and relative phase shift at each frequency.

[0009] S2. Based on the relative attenuation and relative phase shift, the density-independent ratio parameter is obtained by calculating the ratio of attenuation to phase at each frequency. The density-independent ratio parameter is input into the preset density-moisture joint mapping model to obtain the current bulk density estimate and the coarse moisture estimate. Based on the deviation between the current bulk density estimate and the standard bulk density, the relative attenuation and relative phase shift are linearly compensated to obtain the density-normalized attenuation and density-normalized phase.

[0010] S3. Accumulate data points of density normalized decay and density normalized phase online to obtain accumulated data points and monitor the range of change of the rough estimated moisture content; when the range of change of the rough estimated moisture content exceeds the set threshold and the number of accumulated data points reaches the preset number, perform local curve fitting on the calibration curve segment formed by the accumulated data points on the complex plane to obtain the fitted curve; construct the curvature feature vector based on the local curvature of the fitted curve.

[0011] S4. By calculating the statistical distance between the curvature feature vector and the feature centers of each coal type in the pre-constructed coal type dielectric dispersion fingerprint database, the curvature feature vector is matched in the coal type dielectric dispersion fingerprint database to identify the coal type category of the current coal flow.

[0012] S5. Call the corresponding moisture calibration sub-model according to the coal type, and construct a density-independent feature vector based on the ratio of attenuation to phase at each frequency; input the density-independent feature vector into the moisture calibration sub-model to obtain the total moisture content of the coal fed into the furnace.

[0013] Secondly, this application also provides an online detection system for the moisture content of coal entering the furnace based on microwave attenuation phase analysis, used to implement the method described in the first aspect, the system comprising:

[0014] The multi-frequency microwave sensing module is used to transmit microwave signals of at least two different frequencies through the coal flow on the coal conveyor belt, receive the microwave signal after penetrating the coal flow, perform orthogonal demodulation and reference comparison on the microwave signal after penetration, and obtain the relative attenuation and relative phase shift at each frequency.

[0015] The density compensation and coarse estimation module is used to obtain density-independent ratio parameters by calculating the ratio of attenuation to phase at each frequency based on relative attenuation and relative phase shift. The density-independent ratio parameters are then input into a preset density-moisture joint mapping model to obtain the current bulk density estimate and coarse moisture estimate. Based on the deviation between the current bulk density estimate and the standard bulk density, the relative attenuation and relative phase shift are linearly compensated to obtain density-normalized attenuation and density-normalized phase.

[0016] The dynamic feature extraction module is used to accumulate data points of density normalized decay and density normalized phase online to obtain accumulated data points and monitor the range of change of the rough estimated moisture content. When the range of change of the rough estimated moisture content exceeds the set threshold and the number of accumulated data points reaches the preset number, the module performs local curve fitting on the calibration curve segment formed by the accumulated data points on the complex plane to obtain the fitted curve. Based on the local curvature of the fitted curve, a curvature feature vector is constructed.

[0017] The coal type adaptive identification module is used to identify the coal type of the current coal flow by calculating the curvature feature vector and matching it with the statistical distance of the feature centers of each coal type in the pre-built coal type dielectric dispersion fingerprint database.

[0018] The precise moisture inversion module is used to call the corresponding moisture calibration sub-model according to the coal type, and construct density-independent feature vectors based on the ratio of attenuation to phase at each frequency; the density-independent feature vectors are input into the moisture calibration sub-model to obtain the total moisture content of the coal fed into the furnace.

[0019] The aforementioned online detection method and system for coal moisture in furnace feed based on microwave attenuation phase analysis, after obtaining relative attenuation and relative phase shift through multi-frequency microwave transmission of coal flow, does not directly map the measured values ​​to moisture. Instead, it first constructs a density-independent ratio parameter based on the ratio of attenuation to phase, and simultaneously estimates the current bulk density and coarse moisture using a pre-set density-moisture joint mapping model. Then, it uses the deviation between the density estimate and the standard bulk density to linearly compensate for the attenuation and phase, thereby unifying the measurement data under different density conditions to the standard density state, obtaining density-normalized attenuation and density-normalized phase that only reflect the intrinsic dielectric properties of coal. Based on this, it utilizes the natural fluctuations in moisture content of the coal flow itself... Normalized data points under different moisture conditions are accumulated. When the moisture variation range and the amount of accumulated data meet the conditions, local curve fitting is performed on the calibration curve segment formed by these data points on the complex plane. The local curvature of the curve is extracted to form a curvature feature vector. This curvature feature vector carries the geometric fingerprint information of the dielectric dispersion characteristics of the coal type. Subsequently, by calculating the curvature feature vector and matching it with the statistical distance of the feature centers of each coal type in the pre-constructed coal type dielectric dispersion fingerprint database, the coal type of the current coal flow is automatically identified. Finally, according to the identified coal type, the moisture calibration sub-model dedicated to that coal type is called. The density-independent feature vector recalculated based on the original attenuation and phase is input into the sub-model, and the total moisture of the coal fed into the furnace is output. The entire process utilizes a closed-loop technology chain of "density compensation to remove interference, curvature extraction to perceive coal type, and identification and switching of dedicated models." This transforms the curvature of the complex plane calibration curve from a previously ignored fitting byproduct into a physical fingerprint for coal type identification. This enables the detection system to achieve online self-identification of coal type and dynamic switching of calibration models without increasing hardware costs. As a result, even under industrial site conditions with frequent coal type changes and drastic fluctuations in coal seam thickness and bulk density, it can still achieve high-precision online detection of total moisture content in coal entering the furnace with extremely low pre-calibration costs. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A schematic flowchart of an online detection method for moisture content in coal fed into the furnace based on microwave attenuation phase analysis provided by the present invention;

[0022] Figure 2 This is a schematic diagram illustrating the process of constructing a curvature feature vector in one optional embodiment of the present invention;

[0023] Figure 3 This is a schematic diagram of an online detection system for moisture content in coal fed into the furnace based on microwave attenuation phase analysis, provided by the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0025] This invention proposes an online detection method for the moisture content of coal entering the furnace based on microwave attenuation phase analysis, which can be deployed in the online detection system of coal conveyor belts in thermal power plants (hereinafter referred to as the "system"). The system consists of a microwave transmitting and receiving device, a temperature sensing module, a signal processing and computing unit, and a communication interface. The microwave transmitting and receiving antennas are installed facing each other on the upper and lower sides of the coal conveyor belt, with the two antennas covering the full width of the belt. The transmitting antenna is driven by at least two frequency synthesizers, capable of simultaneously or time-division multiplexing at least two continuous microwave signals of different frequencies. The receiving antenna collects the microwave signal after it penetrates the coal flow, and then sends it to the superheterodyne receiving front end through a low-noise amplifier and a pre-selected bandpass filter. After down-conversion to an intermediate frequency, it is sampled by an analog-to-digital converter. The temperature sensing module uses a non-contact infrared scanning imaging temperature measurement device, installed near the transmitting antenna, to acquire the two-dimensional temperature distribution on the upper surface of the coal flow. The signal processing and computing unit consists of a digital signal processor (DSP) and an industrial control computer. The DSP performs quadrature demodulation, time-domain gating, and amplitude-phase extraction, while the industrial control computer performs density compensation, curvature extraction, coal type identification, and moisture inversion calculations. The communication interface transmits the final moisture value to the distributed control system via a current loop and industrial Ethernet for real-time adjustment of the coal quality fed into the furnace. The technical solution of the present invention will be described in detail below with reference to specific embodiments.

[0026] Example 1:

[0027] refer to Figure 1 The document presents a flowchart illustrating an online detection method for moisture content in coal fed into a furnace based on microwave attenuation phase analysis, as provided in this application. The method includes the following steps:

[0028] S1. Transmit at least two microwave signals of different frequencies through the coal flow on the coal conveyor belt, receive the microwave signal after penetration through the coal flow, perform orthogonal demodulation and reference comparison on the microwave signal after penetration, and obtain the relative attenuation and relative phase shift at each frequency.

[0029] Specifically, the transmitting antenna is driven by at least two independent frequency synthesizers. Taking two independent frequency synthesizers as an example, each generates a center frequency of... and Two continuous microwave signals, both at frequencies within the industrial, scientific, and medical bands with sufficient frequency spacing, are used to distinguish the dispersion characteristics of the coal's dielectric constant at the two frequencies. The two microwave signals are amplified and then combined in a combiner, fed into a transmitting antenna, and radiated onto the coal flow on the conveyor belt below. The microwave signal penetrating the coal flow is received by a receiving antenna below the belt, and then passes through a low-noise amplifier and a pre-selected bandpass filter before entering a dual-channel superheterodyne receiver front-end. The dual-channel superheterodyne receiver front-end uses two local oscillator signals to... and The signals are down-converted to their respective intermediate frequencies (IFs), and the IF signals are synchronously sampled by a high-speed analog-to-digital converter at a sampling rate no less than four times that of the IF. The sampled data is then quadrature-demodulated in the digital domain to obtain the frequency. Corresponding in-phase components and orthogonal components and frequency Corresponding in-phase components and orthogonal components The amplitude and phase at each frequency are calculated using the in-phase and quadrature components respectively, according to the following formula:

[0030]

[0031]

[0032] in, Represents frequency The amplitude of the down-demodulated signal; and Frequency The corresponding in-phase and quadrature components; Represents frequency The original phase obtained directly from the arctangent has a range of . To obtain a true continuous phase, for Phase unwrapping is performed to eliminate the abrupt change in the arctangent function, resulting in the unwrapped phase. , which serves as the measurement phase at that frequency.

[0033] To eliminate resonant interference caused by reflections from conveyor belt metal joints, direct leakage signals between antennas, and multiple reflections within the coal flow, an inverse Fourier transform is performed on the demodulated complex baseband signal in the digital domain to obtain the time-domain impulse response. The arrival time of the main transmission pulse is identified on the time-domain waveform, and a time gate of preset width is set to retain only the signal component within the time gate, while all responses outside the time gate caused by multiple reflections and structural clutter are set to zero. Then, the gated time-domain data is Fourier transformed back to the frequency domain to reconstruct the pure complex transmission coefficient. ,in For frequency index, Indicates the first Each frequency. Specifically, let the discrete sampling sequence of the gated time-domain signal be... The sampling interval is The frequency domain complex transmission coefficient is obtained through discrete Fourier transform: ,in The number of points in the Discrete Fourier Transform. For each sampling point in the discrete sampling sequence, at the center frequency Extracting the corresponding complex value from the given location constitutes the... The complex transmission coefficient fully describes the changes in amplitude and phase of a microwave signal after it passes through a coal flow; its mode... Reflects the amplitude and argument of the transmitted signal. It reflects the phase delay of the transmitted signal relative to the transmitted signal.

[0034] Using the microwave signal when the belt is unloaded as a reference, the relative attenuation and relative phase shift at each frequency are obtained by the following formula:

[0035]

[0036]

[0037] in, Represents frequency The relative attenuation of microwaves under these conditions is expressed in decibels. Represents frequency The relative phase shift of microwaves, in radians; This indicates the frequency measured after penetrating the coal flow. Complex transmission coefficient at the specified depth; Indicates the frequency measured when the belt is unloaded. Complex transmission coefficient at the specified depth; Indicates taking the modulus of a complex number; This represents the argument of a complex number. The system is set to perform automatic no-load calibration at preset time intervals, taking a no-load measurement at the coal-free section of the conveyor belt every preset time interval to update the reference amplitude and phase, ensuring the long-term stability of the benchmark.

[0038] After obtaining the relative attenuation and relative phase shift, a two-dimensional temperature distribution on the upper surface of the coal flow is acquired using a non-contact infrared scanning imaging temperature measurement device. The equivalent temperature of the coal flow is then calculated using the gray-scale weighted average method. Based on the linear temperature compensation model calibrated beforehand using standard coal samples in a constant temperature chamber, temperature corrections are applied to the relative attenuation and relative phase shift, expressed as follows:

[0039]

[0040]

[0041] In the formula, Indicates the frequency after temperature correction. The relative decay below; Indicates the frequency after temperature correction. The relative phase shift below; Preset reference temperature; For frequency The temperature coefficient of attenuation represents the amount of attenuation caused by a unit temperature change near the reference temperature. For frequency The temperature coefficient of the lower phase represents the amount of phase change caused by a unit temperature change near a reference temperature. The temperature coefficient is expressed through the reference temperature... A variable-temperature microwave transmission experiment was conducted on the same coal sample nearby, and the measurement data were obtained through linear regression. The specific calibration process was as follows: the coal sample temperature was controlled at multiple different levels, and its relative attenuation and phase shift were measured at each temperature level. Attenuation-temperature curves and phase-temperature curves were established, and the slope of the curves determined the... and After this step, the temperature-corrected relative attenuation and relative phase shift are used as the final relative attenuation and relative phase shift at the corresponding frequency for all subsequent processing. It can be considered that the influence of temperature fluctuations has been largely eliminated, and they have become observations that are only related to the physical properties of coal and its moisture content.

[0042] S2. Based on the relative attenuation and relative phase shift, the density-independent ratio parameter is obtained by calculating the ratio of attenuation to phase at each frequency. The density-independent ratio parameter is input into the preset density-moisture joint mapping model to obtain the current bulk density estimate and coarse moisture estimate. Based on the deviation between the current bulk density estimate and the standard bulk density, the relative attenuation and relative phase shift are linearly compensated to obtain the density-normalized attenuation and density-normalized phase.

[0043] Specifically, firstly, using temperature-corrected relative attenuation and relative phase shift, density-independent ratio parameters that are thickness-insensitive and robust to density changes are constructed for each frequency. Under the condition of plane wave perpendicular incidence in a homogeneous lossy medium, the attenuation... and phase shift All are related to the thickness of the coal seam Proportional, that is , ,in The attenuation constant of the medium (unit: Np / m). This is the phase constant of the medium (unit: rad / m). Therefore, the ratio... With thickness Completely unrelated, and at the same time because and Both are related to bulk density They are approximately proportional, and their ratio is right The dependence is much smaller than that of a single entity. or Based on this principle, the ratio of attenuation to phase is calculated for each frequency:

[0044]

[0045] in, Represents frequency Density-independent ratio parameter, Represents frequency Density-independent ratio parameters; and Frequency Relative attenuation and phase shift after temperature correction; and Frequency The relative attenuation and phase shift after temperature correction are illustrated using two frequencies as examples.

[0046] Will and A pre-defined density-moisture joint mapping model is input. In this embodiment, the model employs a mapping relationship constructed based on multiple linear regression. This mapping relationship is established by collecting dual-frequency microwave measurement data on coal samples from multiple mines across four major coal types—anthracite, lean coal, bituminous coal, and lignite—on a laboratory rotary drum test bench, covering typical moisture and bulk density ranges. and Using the known bulk density and moisture content as independent variables, regression coefficients are obtained through least squares fitting. During online measurement, the currently calculated values ​​are used... and Substituting this regression relationship, we obtain the current packing density estimate. and rough estimate of moisture Because the multiple linear regression model has a simple structure and low computational cost, it can meet the needs of online real-time processing. Its regression model can be expressed as:

[0047]

[0048]

[0049] in, , , and , , These are the regression coefficients, determined through offline training.

[0050] Obtain the packing density estimate Then, using a pre-calibrated density sensitivity coefficient, density linear compensation is performed on the temperature-corrected relative attenuation and relative phase shift, uniformly mapping them to the standard packing density. The corresponding state, the process is represented as:

[0051]

[0052]

[0053] in, Represents frequency Density normalization attenuation after density compensation; Represents frequency Density-normalized phase after density compensation; For frequency The density sensitivity coefficient with lower attenuation represents the amount of attenuation caused by a one-unit change in bulk density under roughly the same moisture conditions. For frequency The density sensitivity coefficient of the lower phase represents the amount of phase change caused by a one-unit change in bulk density under roughly the same moisture conditions. Estimate the current packing density; This represents the standard bulk density. The density sensitivity coefficient is obtained by linear regression of measurement data of mixed coal samples at different density levels under laboratory conditions: This involves measuring the same type of coal at different density levels... and Keep the moisture content constant and fit. and The slope of the straight line is the sensitivity coefficient. The superscript "norm" indicates that the variable has undergone density normalization. After density normalization, the point pair... It can be approximated as the measured value when the density is uniformly set to the standard bulk density. The trajectory formed by its change with moisture will reflect the dielectric properties of the coal itself more, providing a clean input for subsequent curvature analysis.

[0054] S3. Accumulate data points of density normalized decay and density normalized phase online to obtain accumulated data points and monitor the range of change of the rough estimate of moisture. When the range of change of the rough estimate of moisture exceeds the set threshold and the number of accumulated data points reaches the preset number, perform local curve fitting on the calibration curve segment formed by the accumulated data points on the complex plane to obtain the fitted curve. Construct a curvature feature vector based on the local curvature of the fitted curve.

[0055] Specifically, after obtaining the density-normalized attenuation and density-normalized phase of each frame of measurement data, the system stores them together with the corresponding coarsely estimated moisture content in a data buffer established for each frequency. In this embodiment, the data buffer adopts a first-in-first-out queue structure, with each buffer having a preset maximum capacity, and each data point containing density-normalized attenuation. Density-normalized phase and rough estimate of moisture Three fields. The system continuously monitors the difference between the maximum and minimum estimated moisture values ​​in the buffer. When the range of variation in the estimated moisture exceeds a preset moisture fluctuation threshold and the total number of data points in the buffer reaches a preset minimum cumulative frame count, it is considered that sufficient diverse data covering a certain moisture range has been accumulated, and curvature extraction can be initiated. This design is based on an important understanding: only when the natural fluctuations in moisture cover a sufficiently wide range can the data points on the complex plane depict a complete calibration curve, and its curvature can be reliably calculated. If the moisture fluctuation is insufficient, the fitted quadratic curve will heavily rely on local noise, resulting in inaccurate curvature estimation. Therefore, the dual conditions of moisture range and minimum frame count constitute the "gating" of curvature extraction, ensuring the robustness of coal type fingerprints.

[0056] Curvature extraction is performed on the attenuation-phase complex plane, with density-normalized attenuation as the x-axis and density-normalized phase as the y-axis. For frequency... The corresponding buffer zone is used to project all data points onto the complex plane. The data points are distributed along a curved trajectory, which represents the complex plane calibration curve segment for the current coal type within this moisture range. A quadratic polynomial fit is performed on the buffer zone data points to establish a local curve model of phase decay.

[0057]

[0058] in, This indicates the normalized decay of the independent variable density; Indicates the normalized phase of the dependent variable density; , , These represent the constant term, the coefficient of the first term, and the coefficient of the second term obtained from the fitting process. In this embodiment, ordinary least squares method is used for fitting, that is, minimizing the sum of squares of the fitting residuals for each data point to obtain the coefficients. , , The estimated value. In the specific solution, assume the queue contains... Points Construct the design matrix , of which Behavior Then the coefficient vector The least squares solution is ,in ; , indicating the index of the data point This indicates the transpose operation.

[0059] Calculate the mean of the density normalized decay of all data points within the buffer, denoted as . The local curvature at this mean is calculated using the fitting coefficients:

[0060]

[0061] in, Indicates frequency Local curvature extracted from below; and These are the coefficients of the first term and the coefficients of the second term obtained from fitting the above quadratic polynomial, respectively. This represents the arithmetic mean of the normalized decay of the density of all data points within the buffer. (The numerator contains...) To fit the curve in The absolute value of the second derivative reflects the degree to which the curve deviates from a straight line; in the denominator... The first derivative of the fitted curve at that point. As a normalization factor, the curvature value depends only on the shape of the curve rather than the coordinate scale. Since the density effect has been normalized in S2, the curvature mainly reflects the geometric representation of the dispersion characteristics of the coal's inherent complex permittivity as a function of frequency on the complex plane, and is less affected by the absolute value of moisture and density fluctuations. This leap from "pure numerical fitting" to "geometric feature extraction" endows this method with the key capability of upgrading from ordinary moisture measurement to intelligent coal type sensing. Its innovation lies in the fact that existing technologies all regard the calibration curve as a mathematical mapping tool and have never realized that its curvature itself carries identifiable coal type identity information. However, this scheme, by means of dual-frequency measurement and density normalization, creatively elevates the curvature from a fitting byproduct into a stable physical fingerprint.

[0062] Perform the curvature extraction process described above on the two frequencies respectively to obtain... and , forming a two-dimensional curvature eigenvector This vector represents the dielectric dispersion geometric fingerprint of the current coal type under dual-frequency microwave detection.

[0063] S4. By calculating the statistical distance between the curvature feature vector and the feature centers of each coal type in the pre-constructed coal type dielectric dispersion fingerprint database, the curvature feature vector is matched in the coal type dielectric dispersion fingerprint database to identify the coal type category of the current coal flow.

[0064] Specifically, a dielectric dispersion fingerprint database for each coal type was pre-constructed in a laboratory environment. The construction process involved: for each standard sample of the coal type to be covered, simulating the working conditions of a coal conveyor belt on a rotary drum test bench, covering typical moisture and density ranges, and collecting dual-frequency attenuation and phase data; using the same density normalization and curvature extraction process as in the online process, obtaining multiple sets of curvature feature vector samples for each coal type under different moisture ranges; and calculating the mean vector for multiple sets of curvature feature vectors of the same coal type. subscript For coal type category index, This represents the total number of different coal types in the storage facility. For the first The fingerprint database contains the mean vector of the curvature feature vector samples for each type of coal, with each component representing the average curvature of that coal type at two frequencies. The database also stores the calibration sub-model number and model parameters corresponding to each type of coal.

[0065] In this embodiment, Euclidean distance is used as the statistical distance for coal type matching. For the currently extracted curvature feature vector... Calculate its relationship with the fingerprint in the database. Coal type mean vector Euclidean distance between them:

[0066]

[0067] in, Indicates the current coal flow and the first Euclidean distance between coal types; and These are the two components of the current curvature eigenvector; and The first Mean curvature vector of coal The two components.

[0068] The minimum value among all calculated Euclidean distances is selected, and the coal type corresponding to this minimum value is the matched coal type. An acceptance threshold is set. When the minimum Euclidean distance is less than the acceptance threshold, the coal type is considered successfully identified, and the identified coal type is used as the coal type of the current coal flow. When the minimum Euclidean distance is greater than or equal to the acceptance threshold, it indicates that the current curvature feature vector is significantly different from any known coal type in the fingerprint database. The current coal flow is marked as an unknown coal type, and the general model is continued to calculate moisture content. At the same time, the current curvature feature vector and the corresponding original data are stored in a temporary storage area for subsequent offline analysis and database expansion.

[0069] S5. Call the corresponding moisture calibration sub-model according to the coal type, and construct a density-independent feature vector based on the ratio of attenuation to phase at each frequency; input the density-independent feature vector into the moisture calibration sub-model to obtain the total moisture content of the coal fed into the furnace.

[0070] Specifically, after coal type identification, the moisture calibration sub-model corresponding to the identified coal type is retrieved from the model storage unit. This moisture calibration sub-model is a dedicated model built for this single coal type under laboratory conditions. Its input features are density-independent feature vectors composed of the ratio of attenuation to phase at various frequencies. However, the training data comes only from calibration experiments of this coal type under different moisture and density conditions. Therefore, the moisture inversion accuracy for this coal type is much higher than that of the general model for mixed coal types. The sub-model adopts a linear regression structure with bias, and defines the density-independent feature vector as an extended vector. ,in and The current frame frequency and The ratio of the decrease in attenuation to the phase indicates the model's prediction of moisture content. The output is ,in This is the weight vector, containing the slope coefficients and intercept of the linear regression. The model parameters are determined through least-squares fitting of offline calibration data for this coal type.

[0071] After each coal type identification, the system needs to recalculate the ratio of attenuation to phase at each frequency based on the latest acquired and temperature-corrected relative attenuation and relative phase shift of the current frame, forming a density-independent feature vector. The purpose of this recalculation is to ensure that the final moisture value reflects the actual coal flow state at the current moment, rather than the moisture state corresponding to the historical data used in the curvature extraction stage. The recalculated density-independent feature vector is input into the moisture calibration sub-model corresponding to the coal type, and the model outputs the total moisture content of the coal entering the furnace for the current coal flow. The final moisture value is then low-pass filtered and transmitted to the distributed control system via a current loop and an industrial Ethernet interface.

[0072] The aforementioned online detection method for coal moisture in furnace feed based on microwave attenuation phase analysis, after obtaining relative attenuation and relative phase shift through multi-frequency microwave transmission of coal flow, does not directly map the measured values ​​to moisture. Instead, it first constructs a density-independent ratio parameter based on the ratio of attenuation to phase, and simultaneously estimates the current bulk density and coarse moisture using a pre-set density-moisture joint mapping model. Then, it uses the deviation between the density estimate and the standard bulk density to linearly compensate for the attenuation and phase, thereby unifying the measurement data under different density conditions to the standard density state, obtaining density-normalized attenuation and density-normalized phase that only reflect the intrinsic dielectric properties of coal. On this basis, it utilizes the natural fluctuation of the coal flow's own moisture content online... Normalized data points under different moisture conditions are accumulated. When the range of moisture variation and the amount of accumulated data meet the conditions, local curve fitting is performed on the calibration curve segment formed by these data points on the complex plane. The local curvature of the curve is extracted to form a curvature feature vector. This curvature feature vector carries the geometric fingerprint information of the dielectric dispersion characteristics of the coal type. Subsequently, by calculating the curvature feature vector and matching it with the statistical distance of the feature centers of each coal type in the pre-constructed coal type dielectric dispersion fingerprint database, the coal type of the current coal flow is automatically identified. Finally, according to the identified coal type, the moisture calibration sub-model dedicated to that coal type is called. The density-independent feature vector recalculated based on the original attenuation and phase is input into the sub-model, and the total moisture of the coal fed into the furnace is output. The entire process utilizes a closed-loop technology chain of "density compensation to remove interference, curvature extraction to perceive coal type, and identification and switching of dedicated models." This transforms the curvature of the complex plane calibration curve from a previously ignored fitting byproduct into a physical fingerprint for coal type identification. This enables the detection system to achieve online self-identification of coal type and dynamic switching of calibration models without increasing hardware costs. As a result, even under industrial site conditions with frequent coal type changes and drastic fluctuations in coal seam thickness and bulk density, it can still achieve high-precision online detection of total moisture content in coal entering the furnace with extremely low pre-calibration costs.

[0073] Example 2:

[0074] Based on relative attenuation and relative phase shift, density-independent ratio parameters are obtained by calculating the ratio of attenuation to phase at each frequency. These parameters are then input into a pre-set density-moisture joint mapping model to obtain an estimate of the current bulk density and a coarse estimate of moisture content, including the following steps:

[0075] S11. For each frequency, calculate the ratio of relative attenuation to relative phase shift to obtain at least two density-independent ratio parameters.

[0076] Specifically, the calculation process in this step is the same as the ratio calculation process for S2 in Example 1, that is, based on the temperature-corrected relative attenuation and relative phase shift, the density-independent ratio parameter at each frequency is calculated: ,in Represents frequency Density-independent ratio parameter, Frequency after temperature correction The relative decay below, Frequency after temperature correction The relative phase shift below, subscript For frequency indexing. For dual-frequency systems, obtain and Two density-independent ratio parameters; for multi-frequency systems with more than two frequencies, the same density-independent ratio parameter as the number of frequencies can be obtained.

[0077] S12. Based on at least two density-independent ratio parameters, construct joint input features that include at least two density-independent ratio parameters and their pairwise products.

[0078] Specifically, when using only the ratio of attenuation to phase at each frequency as input features, the model can obtain information about the loss tangent from each individual ratio, but it cannot capture the coupling relationship between dielectric dispersion behavior at different frequencies. The product terms between density-independent ratio parameters at different frequencies contain important information about the nonlinear coupling between the packing density effect and the moisture effect. This is because in the Cole-Cole model of the complex permittivity, the ratio of the loss tangent at two frequencies has a deterministic functional relationship with the relaxation time and broadening coefficient, and the pairwise product can precisely approximate this relationship in low-order polynomial form. For example, It can approximately reflect ,in, and Frequency and Dielectric loss angle of coal feeding, The loss tangent is the ratio of the imaginary to the real part of the dielectric constant. The cross-sensitivity of this product term with moisture and density differs from that of linear combinations, thus providing an additional independent dimension of information for decoupling density and moisture. This approach of using the product of ratios as model input yields more accurate density estimation compared to the conventional practice of using only the ratios themselves or linear combinations.

[0079] Assume there is a total Each frequency, to obtain A density-independent ratio parameter, denoted as ,in Constructing a joint input feature vector It contains all Ratio parameter ( ), and all product terms of pairwise combinations ( ).when At that time, the joint input feature vector By introducing a product term, the model can extract more high-order information about dielectric dispersion from a limited number of measurement frequencies without increasing hardware costs.

[0080] S13. Input the joint input features into the pre-trained dual-output nonlinear mapping model, and obtain the current packing density estimate and coarse moisture estimate through nonlinear transformation.

[0081] Specifically, the dual-output nonlinear mapping model in this embodiment can be implemented using a shallow artificial neural network. The network structure is as follows: the number of nodes in the input layer is the same as the dimension of the joint input feature vector; a hidden layer is set, and the nodes in the hidden layer use the hyperbolic tangent activation function. This function exhibits smooth nonlinear transformation characteristics, with an output range of (−1,1), and can approximate complex input-output mapping relationships with fewer hidden nodes. The output layer has two linear nodes, which output the current packing density estimate respectively. and rough estimate of moisture The advantage of the dual-output design lies in the fact that packing density and moisture are coupled in the microwave response. The single-output model requires training two independent models separately and cannot take advantage of parameter sharing when jointly estimating density and moisture. However, the feature representation learned in the hidden layer of the dual-output model is useful for both outputs, which helps to improve estimation accuracy and generalization ability.

[0082] The training process of this model was completed offline in the laboratory. Dual-frequency attenuation and phase data of mixed coal samples at different moisture and bulk density levels were collected on a rotary drum test bench. For each sample, a joint input feature vector was calculated using S11 and S12 methods, while the known moisture content and bulk density were recorded as supervision labels. All samples were divided into training and validation sets. The network parameters were trained using backpropagation combined with the Levenberg-Marquardt optimization algorithm, with the optimization objective being to minimize the sum of the mean square errors of the two output nodes. After training, the network weights and biases were fixed and loaded into the signal processing and computation unit of the online measurement system.

[0083] Compared to the multiple linear regression model used in Example 1, the dual-output nonlinear mapping model used in this example, through the nonlinear transformation of the hidden layer and the introduction of the product term in the joint input features, can more accurately capture the complex nonlinear relationship between the density-independent ratio parameter and the bulk density and moisture. Under the condition of mixed coal types, the estimation deviation of bulk density is significantly reduced, thereby improving the accuracy of subsequent density compensation and curvature extraction.

[0084] Example 3:

[0085] refer to Figure 2The system accumulates data points of density-normalized decay and density-normalized phase online to obtain accumulated data points and monitors the range of changes in the coarsely estimated moisture content. When the range of changes in the coarsely estimated moisture content exceeds a set threshold and the number of accumulated data points reaches a preset number, it performs local curve fitting on the calibration curve segment formed by the accumulated data points on the complex plane to obtain a fitted curve. Based on the local curvature of the fitted curve, it constructs a curvature feature vector, including the following steps:

[0086] S21. For each frequency, establish a first-in-first-out data queue corresponding to the frequency. Record the density normalized attenuation, density normalized phase, and corresponding coarse moisture estimate of each measurement frame as a data record and store the data record in the first-in-first-out data queue.

[0087] Specifically, for each frequency Establish an independent first-in-first-out data queue The queue has a preset maximum capacity. Each frame of measurement data is processed by S2 to obtain density-normalized attenuation. Density-normalized phase And a rough estimate of moisture content Combine the three into a single data record. subscript Indicates the first The sequence number of the frame measurement data in the queue. New data records enter the queue from the tail. When the queue is full, the oldest record at the head of the queue is automatically removed, ensuring that the queue always contains valid data from the most recent period.

[0088] S22. Monitor the difference between the maximum and minimum values ​​of the rough estimate of moisture in the first-in-first-out (FIFO) data queue. When the difference between the maximum and minimum values ​​exceeds the preset moisture fluctuation threshold and the total number of data records in the FIFO data queue reaches the preset minimum cumulative frame number, use density-normalized attenuation as the independent variable and density-normalized phase as the dependent variable to perform weighted quadratic polynomial fitting on all data records in the FIFO data queue to obtain the quadratic term coefficients, linear term coefficients, and constant term.

[0089] Specifically, curvature extraction requires a sufficient range of moisture fluctuations to ensure that the data points on the complex plane cover a calibration curve segment with adequate curvature. If the moisture fluctuations are too small, the data points will be concentrated in local areas of the complex plane, making the fitted quadratic polynomial susceptible to noise and resulting in unstable curvature estimation. The system continuously monitors the queue. The rough estimate of moisture content for all data records is used to calculate the maximum value. and minimum value difference Curvature extraction is triggered when both of the following conditions are met simultaneously: Firstly... Exceeding the preset moisture fluctuation threshold Secondly, the total number of data records in the queue. Reach the preset minimum cumulative frame rate If both conditions are met, it indicates that a sufficient number and variety of data points have accumulated in the queue.

[0090] This embodiment uses robust locally weighted quadratic polynomial regression instead of ordinary least squares fitting in Embodiment 1. The fitting model remains in quadratic polynomial form. However, the weights of the data points are not uniform. A local weighting strategy is adopted, using the current estimated points... Centered on a point, a triangular kernel function is used to assign weights to each data point. The form of the triangular kernel function is:

[0091]

[0092] in, To give the first The weight of each data point The bandwidth parameter is a preset proportion of the density-normalized attenuation range among the data points, typically 70%. This weighting function ensures that points closer to the center point contribute more to the fitting, while points outside the bandwidth have zero weight, completely eliminating their interference with local curvature estimation.

[0093] To further enhance robustness to outliers, an iterative reweighted least squares method is introduced based on local weighting. The specific procedure is as follows: First, an initial weighted least squares fitting is performed using the local triangular kernel weights to obtain the initial fitting coefficients. , , Then calculate the fitting residuals for each data point. The residuals are then reweighted according to the Huber weighting function. The Huber weighting function is defined as follows:

[0094]

[0095] in, To preset a residual threshold, residuals smaller than the threshold retain full weight, while residuals larger than the threshold are assigned weights inversely proportional to the threshold, effectively suppressing the influence of large residual anomalies. The new weights are multiplied by the original triangular kernel weights to obtain composite weights, which are then used for weighted least squares fitting to update the coefficients. This process is iterated a preset number of times, typically three times, until the coefficients converge. The final output is the coefficient of the first-order term of the robust estimate. coefficient of quadratic term and constant term Thanks to the use of a two-layer weighting mechanism, this method can maintain the stability of curvature estimation even when there is abnormal data in the buffer caused by electromagnetic interference or belt vibration.

[0096] S23. Calculate the mean of the density-normalized attenuation of all data records in the first-in-first-out (FIFO) data queue. Based on the quadratic term coefficients, linear term coefficients, and the mean, calculate the local curvature at the corresponding frequency using the curvature formula. The formula for calculating the local curvature is:

[0097]

[0098] in, Indicates the first Frequency Local curvature below; This represents the coefficient of the first-order term obtained by fitting a weighted quadratic polynomial. This represents the coefficients of the quadratic term obtained by fitting a weighted quadratic polynomial. This represents the mean of the density normalized decay of all data records in the first-in-first-out (FIFO) data queue.

[0099] Specifically, compute queue All The arithmetic mean of the density-normalized decay of each data record:

[0100]

[0101] in, Represents frequency The mean of density-normalized decay in the corresponding queue. For the first in the queue The density-normalized decay value of each record. This represents the total number of data records in the queue.

[0102] Mean and fitting coefficients , Substituting into the curvature formula, we get This formula is the standard definition of the curvature of a plane curve in differential geometry, applied to a specific form of a quadratic polynomial curve. In the formula, The coefficients of the first term obtained by fitting a weighted quadratic polynomial represent the slope of the tangent line to the fitted curve at the mean point. The coefficients of the quadratic term are obtained by fitting a weighted quadratic polynomial. This is the second derivative of the fitted curve, reflecting the degree to which the curve deviates from a straight line. At the mean point, the first derivative is... The sum of the squares of the first derivatives plus one in the denominator, raised to the power of the cube of 2, forms a normalization factor, making the curvature measure independent of the parameterization method. A positive absolute value ensures the non-negativity of the curvature, because in moisture calibration, calibration curves typically bend in the same direction, and the sign is not significant.

[0103] S24. Combine the local curvatures corresponding to at least two different frequencies to form a curvature feature vector.

[0104] Specifically, frequency The result obtained and frequency The result obtained Combined into a two-dimensional vector If the system uses more than two frequencies, the dimension of the curvature eigenvector increases accordingly, providing richer dispersion information.

[0105] This embodiment, by introducing triangular kernel local weighting and Huber iterative reweighted least squares, significantly enhances the robustness of curvature extraction against outliers and local noise compared to the ordinary least squares fitting in Embodiment 1. Under conditions of conveyor belt vibration, coal seam discontinuity, and sporadic electromagnetic interference in industrial settings, robust fitting ensures the stability of the curvature feature vector, thereby improving the consistency and accuracy of coal type identification.

[0106] Example 4:

[0107] By calculating the statistical distance between the curvature feature vector and the feature centers of each coal type in a pre-constructed coal type dielectric dispersion fingerprint database, the curvature feature vector is matched in the coal type dielectric dispersion fingerprint database to identify the coal type category of the current coal flow. This includes the following steps:

[0108] S31. Obtain the mean curvature vector and curvature covariance matrix of various coal types from the pre-constructed coal type dielectric dispersion fingerprint database; wherein, the curvature covariance matrix is ​​the covariance matrix obtained by statistically analyzing the curvature feature vector samples extracted from the corresponding coal type under multiple moisture ranges.

[0109] Specifically, in the offline construction of the coal type dielectric dispersion fingerprint database, not only is the mean curvature vector of each type of coal stored, but also its curvature covariance matrix. For the th For a given type of coal, multiple different moisture levels were selected within its typical moisture range on a laboratory rotary drum test bench. At each moisture level, the operating conditions of a coal conveyor belt were simulated, and multiple sets of dual-frequency attenuation and phase data were collected. After density normalization and curvature extraction, multiple sets of curvature feature vector samples were obtained. Let the total number of samples obtained be... A set of effective curvature feature vector samples, denoted as Each sample is a two-dimensional vector. subscript The sample number. For the first Each sample at frequency The curvature below, For the first Each sample at frequency The curvature below.

[0110] Curvature mean vector Calculated from all samples, the expression is: .in, For the first The mean curvature vector of the coal type has the following components: and , respectively, represent the frequency of this type of coal. and The average value of the lower curvature.

[0111] curvature covariance matrix The distribution characteristics and related structure of the curvature feature vector samples of this coal type around the mean vector are described by the following formula:

[0112]

[0113] in, For the first The curvature covariance matrix of the coal is a two-dimensional symmetric positive definite matrix, whose diagonal elements and They are respectively and The sample variance of the components reflects the fluctuation range of the curvature of the coal type at two frequencies; off-diagonal elements The sample covariance between the two curvature components reflects the degree of linear correlation between them. In the fingerprint database, each coal type stores its... and This is used for online matching.

[0114] S32. For each type of coal, calculate the Mahalanobis distance between the curvature eigenvector and the curvature mean vector based on the curvature covariance matrix; where the formula for calculating the Mahalanobis distance is:

[0115]

[0116] in, This indicates the current coal flow and coal type dielectric dispersion fingerprint in the database. The Mahalanobis distance between different types of coal; Represents the curvature eigenvector; This indicates the first [coal type] dielectric dispersion fingerprint in the database. The mean curvature vector of the coal type; This indicates the first [coal type] dielectric dispersion fingerprint in the database. The inverse matrix of the curvature covariance matrix of the coal type; This represents the transpose operation of a vector.

[0117] Specifically, for the curvature feature vector extracted online... and the fingerprint database Mean curvature vector of coal curvature covariance matrix Calculate the Mahalanobis distance between the two. .in, Indicates the current coal flow and the first The Mahalanobis distance between different types of coal; This is the current curvature feature vector; For the first The mean curvature vector of the coal type; For the first The inverse matrix of the curvature covariance matrix of coal; superscript This represents the transpose operation of a vector. The calculation process of Mahalanobis distance can be understood as follows: First, the difference vector between the current curvature vector and the center of the coal type is calculated. A linear transformation is performed using the inverse covariance matrix. This transformation eliminates the influence of correlation between components and normalizes the variance in each direction, making the deviation tolerance higher in directions with large curvature fluctuations and more sensitive to deviations in directions with small curvature fluctuations. Then, the Euclidean length of the transformed vector is calculated, which is the Mahalanobis distance.

[0118] Compared to the Euclidean distance used in Example 1, the Mahalanobis distance has the advantage of fully considering the statistical correlation between the components of the curvature eigenvector. In actual measurements, and There is often a certain degree of positive correlation between different moisture levels of the same coal type. This means that the curvature eigenvectors are not isotropic spherical in the complex plane, but rather elongated ellipsoids along a certain direction. If Euclidean distance is used, the equidistant lines are circular, which cannot reflect the directionality and dispersion of the actual data distribution. This may lead to points far from the mean along the principal axis of the ellipsoid being misclassified as too large and rejected, or outliers along the minor axis being improperly accepted. Mahalanobis distance uses an affine transformation of the distance metric through the covariance matrix, transforming the equidistant lines into ellipsoids consistent with the shape of the data distribution, enabling more accurate classification decisions. When the covariance matrix is ​​an identity matrix, Mahalanobis distance degenerates into Euclidean distance.

[0119] S33. Select the minimum Mahalanobis distance from all Mahalanobis distances, and use the coal type corresponding to the minimum Mahalanobis distance as the coal type of the current coal flow.

[0120] Specifically, iterate through all fingerprints in the database. Coal planting, calculation Mahalanobis distance Select the minimum value from them. Its corresponding coal type index Furthermore, set an acceptance threshold. ,when At that time, the first The coal type category is output as the coal type category of the current coal flow; when If the distance between the current curvature feature vector and the center of all coal types in the database is too large, the current coal type may be a new coal type that has not been included or the coal quality has changed significantly. At this time, the current coal flow is marked as an unknown coal type, and the system continues to output moisture using the general model, while recording new fingerprint data.

[0121] This embodiment introduces Mahalanobis distance for coal type matching, making full use of the pre-statistical curvature distribution characteristics of each coal type and considering the covariance structure between curvature components. Compared with Euclidean distance matching in Embodiment 1, it significantly reduces the probability of misclassification caused by uneven data distribution and improves the accuracy and reliability of online identification of multiple coal types.

[0122] Example 5:

[0123] The moisture calibration sub-model is a linear regression model, including a weight vector and a parameter covariance matrix. After inputting the density-independent eigenvectors into the moisture calibration sub-model to obtain the total moisture content of the coal fed into the furnace, the present invention further includes the following steps:

[0124] S41. Obtain the reference moisture value of the current coal flow, and calculate the difference between the reference moisture value and the total moisture of the coal entering the furnace as the prediction error.

[0125] Specifically, this embodiment adds an online recursive fine-tuning mechanism for the moisture calibration sub-model based on Embodiment 1. In this embodiment, the moisture calibration sub-model adopts a linear regressor structure with bias, consistent with the model form defined in S5 of Embodiment 1, that is, expanding the density-independent feature vector to... Weight vector The model prediction output includes slope and intercept. This sub-model, having obtained its initial weight vector during offline training in the laboratory, can provide high-accuracy moisture predictions for specific coal types after deployment. However, in actual operation, coal quality may slowly drift due to factors such as the advancement of mining faces, seasonal climate changes, and wind and rain exposure during open-air coal storage, causing the offline-trained model parameters to gradually deviate from the optimal values ​​for the current coal quality over time. If the model parameters remain unchanged, the measurement accuracy will gradually degrade. This embodiment introduces a recursive least squares algorithm to achieve online self-calibration of the model, enabling it to track the slow changes in coal quality and maintain high accuracy throughout its entire lifecycle. This self-learning mechanism solves the pain point of traditional methods requiring frequent shutdowns for recalibration.

[0126] Specifically, online fine-tuning requires an externally provided reference moisture value as a monitoring signal. The reference moisture value can be obtained in several ways: manually sampling the conveyor belt at regular intervals by operators, measuring the total moisture content in a laboratory using standard methods, and then inputting the result into the system; or providing redundant measurements from capacitive moisture sensors deployed on the same coal conveyor belt and operating in conjunction with a microwave detection system. Let the currently acquired reference moisture value be... The sub-model currently outputs the total moisture content of the coal entering the furnace as follows: Calculate the prediction error:

[0127]

[0128] in, This represents the prediction error, which is the difference between the reference value and the model's predicted value. For reference moisture content; This is the current weight vector; This is the current density-independent feature vector.

[0129] S42. Based on prediction error, density-independent eigenvectors, and a preset forgetting factor, the weight vector and parameter covariance matrix of the moisture calibration sub-model are corrected using a recursive least squares update formula to obtain the corrected moisture calibration sub-model; wherein, the recursive least squares update formula is:

[0130]

[0131]

[0132]

[0133] in, Represents the gain vector; The parameter covariance matrix represents the moisture calibration sub-model. Indicates the forgetting factor; Represents density-independent eigenvectors; This indicates a reference moisture value; This represents the weight vector of the moisture calibration sub-model; This represents the predicted moisture value calculated by the moisture calibration sub-model on the density-independent eigenvector, which is the total moisture content of the coal fed into the furnace. This represents the transpose operation of a vector.

[0134] Specifically, after obtaining the prediction error, the model parameters are corrected according to the recursive least squares update formula in this step. The gain vector has the same dimensions as the weight vector. The same principle determines the amount of update that the prediction error is allocated to each weight parameter; The covariance matrix is ​​a parameter that reflects the uncertainty of the current weight vector estimate. It is initialized as a diagonal matrix with a large number of diagonal elements. ,coefficient A larger positive value is chosen to make the algorithm respond faster to new data in the initial stage. It is the identity matrix; The forgetting factor has a value greater than zero and less than one. The closer the value is to one, the slower the decay of historical data, the stronger the model's ability to smooth noise, but the slower the tracking speed. The smaller the value, the faster the model responds to new data, but the more sensitive it is to noise. The density-independent feature vector; The reference moisture value; Let be the weight vector. This refers to the moisture value predicted by the model; superscript This represents the transpose operation of a vector. During the update process, First, based on the gain vector and input Perform reduction correction, then divide by the forgetting factor. It achieves exponential decay of historical information. The entire update process is executed once each time a reference moisture value is obtained, without the need to store historical data, and is completed in real time.

[0135] The mathematical foundation of the recursive least squares algorithm lies in minimizing the cost function of the sum of squared errors with exponential forgetting. After each update, the weight vector... This corresponds to the optimal estimate that minimizes the weighted cumulative sum of squared errors under the current forgetting factor. Through continuous online updates, the moisture calibration sub-model can automatically track the slow drift of coal quality, maintaining optimal measurement accuracy throughout the entire operating cycle without requiring downtime for recalibration.

[0136] Example 6:

[0137] The microwave signal after penetration is subjected to orthogonal demodulation and reference comparison to obtain the relative attenuation and relative phase shift at each frequency, including the following steps:

[0138] S51. Perform pulse compression processing on the received transmitted microwave signal to obtain the time-domain profile; wherein, the microwave signal is a linear frequency modulated continuous wave signal.

[0139] Specifically, the microwave signal transmitted by the microwave transmitting antenna adopts a linear frequency modulated continuous wave (LFM) mode. The frequency of the transmitted signal varies with time within a preset bandwidth. The scanning follows a linear pattern, with a scanning period of [number missing]. That is, time The instantaneous frequency of the signal is ,in The starting frequency is used. The microwave signal received by the receiving antenna after penetrating the coal flow is mixed with a copy of the transmitted signal and low-pass filtered to obtain a difference frequency signal, the frequency of which is proportional to the target delay. The difference frequency signal frequencies corresponding to different propagation delay paths are different; the greater the propagation delay, the higher the difference frequency frequency. Performing a Fourier transform on the difference frequency signal achieves pulse compression processing, focusing the energy of the wide pulse into a narrow peak, obtaining a time-domain profile. In the time-domain profile, the horizontal axis corresponds to the propagation delay, and the vertical axis corresponds to the response intensity of each delay path. Pulse compression separates multiple propagation path signals that were originally indistinguishable in the time domain in the time-domain profile, allowing the peaks corresponding to different delay paths to be distinguished in the time domain. The advantage of using a pulse compression system in this embodiment is that it simultaneously achieves high range resolution and high signal-to-noise ratio, and can clearly separate the main transmitted pulse from multiple reflected pulses, which is unmatched by simple time-domain gates for continuous waves.

[0140] S52. Identify the arrival time of the main transmission pulse in the time domain profile. Using the arrival time as a reference, apply a time gate to the time domain profile to extract the signal segment within the time gate and set the signal segment outside the time gate to zero to obtain the gated time domain signal.

[0141] Specifically, in the time-domain profile, different peaks correspond to different microwave propagation paths. The peak with the strongest amplitude corresponds to the main transmission path, that is, the path where microwaves directly penetrate the coal seam and propagate in a straight line from the transmitting antenna to the receiving antenna; smaller peaks with longer delay times correspond to various reflection paths, including multiple reflections of microwaves at the conveyor belt's metal frame, belt joints, and interfaces between different layers within the coal flow. The arrival time of the main transmission pulse is identified using a peak detection algorithm. ,by Set a time gate with a preset width at the center, and the gate interval is... ,in The time gate width is half the width of the time interval. The original values ​​of signal samples within the gate interval in the time-domain profile are retained, while all samples outside the gate interval are set to zero, resulting in the gated time-domain signal. The width of the time gate must be greater than the time-domain width of the main transmitted pulse to ensure complete preservation of the waveform information of the direct wave, and smaller than the time delay difference between the main transmitted pulse and the most recent reflected pulse to ensure complete isolation of reflection interference. Only the main transmitted information is retained in the gated time-domain signal, and clutter is eliminated.

[0142] S53. Perform a Fourier transform on the gated time-domain signal to obtain the frequency-domain complex transmission coefficient; based on the amplitude and phase of the frequency-domain complex transmission coefficient at each of at least two different frequencies, and taking the microwave signal when the belt is unloaded as a reference, calculate the relative attenuation and relative phase shift.

[0143] Specifically, the gated time-domain signal is returned to the frequency domain via a fast Fourier transform to obtain the frequency-domain complex transmission coefficient. This coefficient has been stripped of multiple reflection interference components, retaining only the propagation response of the main transmission path. Two center frequencies are extracted from the complex transmission coefficient in the frequency domain. and The complex value at the point is where the magnitude is the amplitude of the transmission coefficient and the argument is the phase of the transmission coefficient. Following the calculation formulas for relative attenuation and relative phase shift given in Example 1, and using the microwave signal under no-load conditions as a reference, the relative attenuation and relative phase shift at each frequency are calculated.

[0144] This embodiment employs linear frequency modulated continuous wave combined with pulse compression and time-domain gating techniques. Compared to the method in Embodiment 1, which directly processes continuous wave signals using inverse Fourier transform and time-domain gating, this approach offers higher distance resolution and better resistance to multiple reflection interference. The frequency domain complex transmission coefficient after time-domain gating has a higher signal-to-noise ratio and more stable phase, providing higher-quality foundational data for subsequent density-independent feature construction, curvature extraction, and moisture inversion.

[0145] Example 7:

[0146] After inputting the density-independent feature vector into the moisture calibration sub-model to obtain the total moisture content of the coal fed into the furnace, the present invention further includes the following steps:

[0147] S61. Continuously calculate the Mahalanobis distance between the curvature feature vector of each subsequent measurement frame and the mean curvature vector corresponding to the currently identified coal type.

[0148] Specifically, after successfully identifying the coal type and switching the calibration model to a dedicated sub-model for that coal type, the system does not stop the curvature extraction and fingerprint calculation process. Each new frame of measurement data, after being processed by S1 to S3, continues to calculate density normalized attenuation and density normalized phase, storing them in a buffer. When the first-in-first-out data queue corresponding to any frequency meets the trigger condition and a new curvature feature vector is extracted... Then, let the currently locked coal type be the first... Coal-like material, whose mean curvature vector is The curvature covariance matrix is Calculate the Mahalanobis distance between the newly extracted curvature feature vector and the mean vector of the locked coal type:

[0149]

[0150] in, The current curvature feature vector With locked coal type Mahalanobis distance; To lock the mean curvature vector of the coal type; This is the inverse matrix of the curvature covariance matrix of the locked coal type. The Mahalanobis distance measures the deviation between the dielectric dispersion fingerprint of the current coal flow and the central fingerprint of the locked coal type. During normal operation without coal type change, this distance will fluctuate randomly within a small range due to the natural fluctuations in coal quality. Once the coal type is changed, the curvature characteristics of the complex plane calibration curve will change because the dielectric dispersion characteristics of the new coal type differ from the original locked coal type, leading to a significant increase in the Mahalanobis distance.

[0151] S62. Monitor the number of times the Mahalanobis distance exceeds the preset switching threshold consecutively within the sliding window.

[0152] Specifically, set a length of The frame's sliding window works by storing the newly extracted curvature feature vector and its corresponding Mahalanobis distance value into the sliding window each time a new feature vector is extracted, thus recording the closest... The corresponding Mahalanobis distance value is extracted each time. A preset switching threshold is set. This threshold is set higher than the acceptance threshold used for coal type identification to avoid frequent erroneous switching due to normal coal quality fluctuations. Within the sliding window, the number of consecutive values ​​exceeding [a certain threshold] is [not specified]. The number of frames for the Mahalanobis distance, denoted as .

[0153] S63. When the number of times reaches the preset switching threshold, it is determined that the coal type has been changed. At least two first-in-first-out data queues corresponding to different frequencies are cleared, and the operation of S3 is restarted to identify the new coal type.

[0154] Specifically, a preset threshold for the number of switching operations. When the sliding window contains more than Frame count When a change in coal type is detected, it is determined that the coal type has been changed. The threshold for the number of switching events needs to balance the risk of erroneous switching with response latency: a value that is too low may cause occasional disturbances to be misinterpreted as a coal type change, triggering unnecessary model switching and buffer clearing; a value that is too high will result in excessively long response delays in identifying the new coal type, during which the system uses the incorrect coal type's sub-model for moisture calculation, affecting measurement accuracy. After determining that the coal type has changed, the following operations are performed: clear the first-in-first-out data queues corresponding to all frequencies. The system clears the curvature feature vector cache and resets the coal type lock status flag. Starting from S3, the system re-accumulates data points for density normalized attenuation and density normalized phase online. Once the moisture fluctuation trigger condition is met, the curvature feature vector is re-extracted and matched against the coal type dielectric dispersion fingerprint database to identify the new coal type. The corresponding calibration sub-model is then automatically invoked. The entire process requires no operator intervention. The system automatically completes the closed loop of detection, identification, and model switching after coal type change, achieving fully automated coal type tracking for full moisture detection of coal entering the furnace under multiple coal type conditions. This coal change identification strategy based on continuous monitoring of curvature fingerprints gives the system the ability to self-perceive and self-adjust, fundamentally solving the long-standing problem of requiring manual intervention for coal type changes in the field.

[0155] The aforementioned online detection method for coal moisture in furnace feed based on microwave attenuation phase analysis, after obtaining relative attenuation and relative phase shift through multi-frequency microwave transmission of coal flow, does not directly map the measured values ​​to moisture. Instead, it first constructs a density-independent ratio parameter based on the ratio of attenuation to phase, and simultaneously estimates the current bulk density and coarse moisture using a pre-set density-moisture joint mapping model. Then, it uses the deviation between the density estimate and the standard bulk density to linearly compensate for the attenuation and phase, thereby unifying the measurement data under different density conditions to the standard density state, obtaining density-normalized attenuation and density-normalized phase that only reflect the intrinsic dielectric properties of coal. On this basis, it utilizes the natural fluctuation of the coal flow's own moisture content online... Normalized data points under different moisture conditions are accumulated. When the range of moisture variation and the amount of accumulated data meet the conditions, local curve fitting is performed on the calibration curve segment formed by these data points on the complex plane. The local curvature of the curve is extracted to form a curvature feature vector. This curvature feature vector carries the geometric fingerprint information of the dielectric dispersion characteristics of the coal type. Subsequently, by calculating the curvature feature vector and matching it with the statistical distance of the feature centers of each coal type in the pre-constructed coal type dielectric dispersion fingerprint database, the coal type of the current coal flow is automatically identified. Finally, according to the identified coal type, the moisture calibration sub-model dedicated to that coal type is called. The density-independent feature vector recalculated based on the original attenuation and phase is input into the sub-model, and the total moisture of the coal fed into the furnace is output. The entire process utilizes a closed-loop technology chain of "density compensation to remove interference, curvature extraction to perceive coal type, and identification and switching of dedicated models." This transforms the curvature of the complex plane calibration curve from a previously ignored fitting byproduct into a physical fingerprint for coal type identification. This enables the detection system to achieve online self-identification of coal type and dynamic switching of calibration models without increasing hardware costs. As a result, even under industrial site conditions with frequent coal type changes and drastic fluctuations in coal seam thickness and bulk density, it can still achieve high-precision online detection of total moisture content in coal entering the furnace with extremely low pre-calibration costs.

[0156] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0157] Based on the same inventive concept, this application also provides a system for implementing the above-mentioned online detection method for moisture content of coal fed into the furnace based on microwave attenuation phase analysis. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more embodiments of the online detection system for moisture content of coal fed into the furnace based on microwave attenuation phase analysis provided below can be found in the limitations of the online detection method for moisture content of coal fed into the furnace based on microwave attenuation phase analysis described above, and will not be repeated here.

[0158] In one exemplary embodiment, such as Figure 3 As shown, an online moisture detection system 30 for coal entering the furnace based on microwave attenuation phase analysis is provided to implement the methods in the above embodiments. The system includes:

[0159] The multi-frequency microwave sensing module 31 is used to transmit microwave signals of at least two different frequencies through the coal flow on the coal conveyor belt, receive the microwave signal after penetrating the coal flow, perform orthogonal demodulation and reference comparison on the microwave signal after penetration, and obtain the relative attenuation and relative phase shift at each frequency.

[0160] The density compensation and coarse estimation module 32 is used to obtain density-independent ratio parameters by calculating the ratio of attenuation to phase at each frequency based on relative attenuation and relative phase shift; input the density-independent ratio parameters into a preset density-moisture joint mapping model to obtain the current bulk density estimate and coarse moisture estimate; and perform linear compensation on the relative attenuation and relative phase shift based on the deviation between the current bulk density estimate and the standard bulk density to obtain density-normalized attenuation and density-normalized phase.

[0161] The dynamic feature extraction module 33 is used to accumulate data points of density normalized decay and density normalized phase online to obtain accumulated data points and monitor the range of change of the rough estimated moisture content. When the range of change of the rough estimated moisture content exceeds the set threshold and the number of accumulated data points reaches the preset number, the local curve fitting is performed on the calibration curve segment formed by the accumulated data points on the complex plane to obtain the fitted curve. Based on the local curvature of the fitted curve, a curvature feature vector is constructed.

[0162] The coal type adaptive identification module 34 is used to identify the coal type category of the current coal flow by calculating the curvature feature vector and matching it with the statistical distance of the feature centers of each coal type in the pre-built coal type dielectric dispersion fingerprint database.

[0163] The precise moisture inversion module 35 is used to call the corresponding moisture calibration sub-model according to the coal type, and construct a density-independent feature vector based on the ratio of attenuation to phase at each frequency; the density-independent feature vector is input into the moisture calibration sub-model to obtain the total moisture content of the coal fed into the furnace.

[0164] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for online detection of moisture content in coal fed into a furnace based on microwave attenuation phase analysis, characterized in that, The method includes: S1. Transmit at least two microwave signals of different frequencies through the coal flow on the coal conveyor belt, receive the microwave signal after penetration through the coal flow, perform orthogonal demodulation and reference comparison on the microwave signal after penetration, and obtain the relative attenuation and relative phase shift at each frequency. S2. Based on the relative attenuation and the relative phase shift, the density-independent ratio parameter is obtained by calculating the ratio of attenuation to phase at each frequency. The density-independent ratio parameter is input into a preset density-moisture joint mapping model to obtain the current bulk density estimate and the coarse moisture estimate. Based on the deviation between the current bulk density estimate and the standard bulk density, the relative attenuation and the relative phase shift are linearly compensated to obtain the density-normalized attenuation and the density-normalized phase. S3. Accumulate the data points of the density normalized decay and the density normalized phase online to obtain accumulated data points, and monitor the range of change of the coarse moisture estimate; when the range of change of the coarse moisture estimate exceeds a set threshold and the number of accumulated data points reaches a preset number, perform local curve fitting on the calibration curve segment formed by the accumulated data points on the complex plane to obtain a fitted curve; construct a curvature feature vector based on the local curvature of the fitted curve. S4. By calculating the statistical distance between the curvature feature vector and the feature centers of each coal type in the pre-constructed coal type dielectric dispersion fingerprint database, the curvature feature vector is matched in the coal type dielectric dispersion fingerprint database to identify the coal type category of the current coal flow. S5. Call the corresponding moisture calibration sub-model according to the coal type, and construct a density-independent feature vector based on the ratio of attenuation to phase at each frequency; input the density-independent feature vector into the moisture calibration sub-model to obtain the total moisture content of the coal fed into the furnace.

2. The method according to claim 1, characterized in that, The density-independent ratio parameter is obtained by calculating the ratio of attenuation to phase at each frequency based on the relative attenuation and the relative phase shift. The density-independent ratio parameter is input into a preset density-moisture joint mapping model to obtain the current bulk density estimate and coarse moisture estimate, including: S11. For each frequency, calculate the ratio of the relative attenuation to the relative phase shift to obtain at least two density-independent ratio parameters; S12. Based on at least two density-independent ratio parameters, construct a joint input feature that includes at least two density-independent ratio parameters and their pairwise products; S13. Input the joint input features into a pre-trained dual-output nonlinear mapping model, and obtain the current packing density estimate and the coarse moisture estimate through nonlinear transformation.

3. The method according to claim 1, characterized in that, The system accumulates data points of the density normalized decay and the density normalized phase online to obtain accumulated data points, and monitors the range of change of the coarsely estimated moisture content. When the range of change of the coarsely estimated moisture content exceeds a set threshold and the number of accumulated data points reaches a preset number, a local curve fitting is performed on the calibration curve segment formed by the accumulated data points on the complex plane to obtain a fitted curve. Based on the local curvature of the fitted curve, a curvature feature vector is constructed, including: S21. For each frequency, establish a first-in-first-out data queue corresponding to the frequency, and store the density normalized attenuation, density normalized phase and the corresponding coarse moisture estimate of each measurement frame as a data record in the first-in-first-out data queue. S22. Monitor the difference between the maximum and minimum values ​​of the coarsely estimated moisture in the first-in-first-out data queue. When the difference between the maximum and minimum values ​​exceeds a preset moisture fluctuation threshold and the total number of data records in the first-in-first-out data queue reaches a preset minimum cumulative frame number, use the density normalized decay as the independent variable and the density normalized phase as the dependent variable to perform a weighted quadratic polynomial fitting on all the data records in the first-in-first-out data queue to obtain the quadratic term coefficients, the linear term coefficients, and the constant term. S23. Calculate the mean of the density-normalized attenuation of all data records in the first-in-first-out data queue. Based on the quadratic term coefficient, the linear term coefficient, and the mean, calculate the local curvature at the corresponding frequency using the curvature formula. The formula for calculating the local curvature is: ； in, Indicates the first Frequency The local curvature mentioned below; This represents the coefficient of the first-order term obtained by fitting the weighted quadratic polynomial; This represents the coefficients of the quadratic term obtained by fitting the weighted quadratic polynomial; This represents the mean of the density-normalized decay of all data records in the first-in-first-out data queue; S24. Combine the local curvatures corresponding to at least two different frequencies to form the curvature feature vector.

4. The method according to claim 3, characterized in that, The process involves calculating the statistical distance between the curvature feature vector and the feature centers of each coal type in a pre-constructed coal type dielectric dispersion fingerprint database, and then matching the curvature feature vector within the database to identify the coal type category of the current coal flow. This includes: S31. Obtain the mean curvature vector and curvature covariance matrix of various coal types from the pre-constructed dielectric dispersion fingerprint database of the coal types; wherein, the curvature covariance matrix is ​​the covariance matrix obtained by statistically analyzing the curvature feature vector samples extracted from the corresponding coal types under multiple moisture ranges. S32. For each type of coal, calculate the Mahalanobis distance between the curvature eigenvector and the curvature mean vector based on the curvature covariance matrix; wherein the formula for calculating the Mahalanobis distance is: ； in, This indicates that the current coal flow is related to the dielectric dispersion fingerprint of the coal type in the database. The Mahalanobis distance between the coal types; This represents the curvature feature vector; This indicates the first [coal type] in the dielectric dispersion fingerprint database. The mean curvature vector of the coal type; This indicates the first [coal type] in the dielectric dispersion fingerprint database. The inverse matrix of the curvature covariance matrix of the coal type; This represents the transpose operation of a vector; S33. Select the minimum Mahalanobis distance from all the Mahalanobis distances, and use the coal type category corresponding to the minimum Mahalanobis distance as the coal type category of the current coal flow.

5. The method according to claim 1, characterized in that, The moisture calibration sub-model is a linear regression model, including a weight vector and a parameter covariance matrix; after inputting the density-independent feature vector into the moisture calibration sub-model to obtain the total moisture content of the coal fed into the furnace, the method further includes: S41. Obtain the reference moisture value of the current coal flow, and calculate the difference between the reference moisture value and the total moisture of the coal entering the furnace as the prediction error; S42. Based on the prediction error, the density-independent feature vector, and the preset forgetting factor, the weight vector and the parameter covariance matrix of the moisture calibration sub-model are corrected using a recursive least squares update formula to obtain the corrected moisture calibration sub-model; wherein, the recursive least squares update formula is: ； ； ； in, Represents the gain vector; The parameter covariance matrix represents the moisture calibration sub-model. This represents the forgetting factor; This represents the density-independent feature vector; This indicates the reference moisture value; The weight vector represents the moisture calibration sub-model; The predicted moisture value calculated by the moisture calibration sub-model on the density-independent eigenvector is the total moisture content of the coal fed into the furnace. This represents the transpose operation of a vector.

6. The method according to claim 1, characterized in that, The process of performing orthogonal demodulation and reference comparison on the penetrated microwave signal to obtain the relative attenuation and relative phase shift at each frequency includes: S51. The received transmitted microwave signal is subjected to pulse compression processing to obtain a time-domain profile; wherein, the microwave signal is a linear frequency modulated continuous wave signal; S52. Identify the arrival time of the main transmission pulse in the time domain profile, and apply a time gate to the time domain profile based on the arrival time to extract the signal segment within the time gate, and set the signal segment outside the time gate to zero to obtain the gated time domain signal. S53. Perform a Fourier transform on the gated time-domain signal to obtain the frequency-domain complex transmission coefficient; based on the amplitude and phase of the frequency-domain complex transmission coefficient at each of the at least two different frequencies, and taking the microwave signal when the belt is unloaded as a reference, calculate the relative attenuation and the relative phase shift.

7. The method according to claim 4, characterized in that, After inputting the density-independent feature vector into the moisture calibration sub-model to obtain the total moisture content of the coal fed into the furnace, the method further includes: S61. Continuously calculate the Mahalanobis distance between the curvature feature vector of each subsequent measurement frame and the mean curvature vector corresponding to the currently identified coal type. S62. Monitor the number of times the Mahalanobis distance exceeds a preset switching threshold consecutively within the sliding window; S63. When the number of times reaches the preset switching number threshold, it is determined that the coal type has been changed, the first-in-first-out data queues corresponding to the at least two different frequencies are cleared, and the operation of S3 is restarted to identify the new coal type.

8. An online detection system for moisture content of coal fed into a furnace based on microwave attenuation phase analysis, used to implement the method according to any one of claims 1 to 7, characterized in that, The system includes: A multi-frequency microwave sensing module is used to transmit microwave signals of at least two different frequencies through a coal conveyor belt, receive the microwave signal after penetrating the coal flow, perform orthogonal demodulation and reference comparison on the microwave signal after penetration, and obtain the relative attenuation and relative phase shift at each frequency. The density compensation and coarse estimation module is used to obtain a density-independent ratio parameter by calculating the ratio of attenuation to phase at each frequency based on the relative attenuation and the relative phase shift; input the density-independent ratio parameter into a preset density-moisture joint mapping model to obtain the current bulk density estimate and coarse moisture estimate; and perform linear compensation on the relative attenuation and relative phase shift based on the deviation between the current bulk density estimate and the standard bulk density to obtain density-normalized attenuation and density-normalized phase. A dynamic feature extraction module is used to accumulate data points of the density normalized decay and the density normalized phase online to obtain accumulated data points and monitor the range of change of the coarse moisture estimate; when the range of change of the coarse moisture estimate exceeds a set threshold and the number of accumulated data points reaches a preset number, local curve fitting is performed on the calibration curve segment formed by the accumulated data points on the complex plane to obtain a fitted curve; and a curvature feature vector is constructed based on the local curvature of the fitted curve. The coal type adaptive identification module is used to identify the coal type category of the current coal flow by calculating the statistical distance between the curvature feature vector and the feature centers of each coal type in the pre-constructed coal type dielectric dispersion fingerprint database, matching the curvature feature vector in the coal type dielectric dispersion fingerprint database; The precise moisture inversion module is used to call the corresponding moisture calibration sub-model according to the coal type, and construct a density-independent feature vector based on the ratio of attenuation to phase at each frequency; the density-independent feature vector is input into the moisture calibration sub-model to obtain the total moisture content of the coal fed into the furnace.

Images