A resin level intelligent sensing and early warning method for impregnators

Abstract

The present application relates to a kind of impregnator resin level intelligent sensing and early warning method, including the multiple sampling positions of resin level area in impregnator equipment;Noise suppression and signal normalization processing are carried out to time domain dielectric signal data set;Feature classification is carried out to data sample under different working conditions;Adaptive baseline model under multiple conditions is constructed using historical liquid level response database;The standardized time domain dielectric signal sequence obtained by actual sampling is input into adaptive baseline model under multiple conditions, and local abnormal trend signal of liquid level is obtained;Local abnormal trend signal of liquid level is processed by multi-window space-time distribution dynamic imaging;The amplitude, persistence and spatial diffusion characteristic parameters of trend change are input into historical false alarm and artifact event feature library;After excluding artifact, the abnormal trend signal of liquid level, dynamically generates adaptive early warning threshold.The present application can realize the high sensitivity collection and dynamic amplification of local electromagnetic response caused by the slight fluctuation of resin level.

Description

Technical Field

[0001] This invention relates to the field of impregnation machine technology, and specifically to a method for intelligent sensing and early warning of resin liquid level in an impregnation machine. Background Technology

[0002] Currently, key manufacturing equipment such as industrial impregnation equipment, such as transformer impregnation machines and composite material prepreg machines, have extremely high requirements for real-time monitoring and early warning of internal resin liquid levels. With the improvement of intelligent manufacturing and equipment health management, intelligent liquid level sensing and early warning systems have become core technologies for improving the safe operation of equipment, assisting intelligent operation and maintenance, and reducing the risk of sudden failures.

[0003] Existing technical solutions still face many bottlenecks in the early detection of liquid level anomalies in complex factory environments and high-value equipment. Current mainstream broadband dielectric liquid level sensing systems generally have the following shortcomings:

[0004] Firstly, the sensitivity of weak signal detection is limited. Initial abnormalities or slight trend fluctuations in liquid level are often masked by stray noise or environmental interference, leading to missed reports or delayed early warnings.

[0005] Secondly, most abnormal trend thresholds are set statically and lack the ability to adapt to actual working conditions such as resin type, temperature fluctuation and impurity changes, making it difficult to efficiently cope with the ever-changing and complex environment.

[0006] Third, the means of identifying and eliminating false events caused by periodic disturbances, external noise, or equipment start-up and shutdown are limited, which easily leads to a high false alarm rate and interferes with the manual operation and maintenance system.

[0007] Fourth, existing systems often only belong to single-point measurement or low-dimensional data discrimination, lacking the ability to analyze spatial distribution, multi-point imaging and trend evolution, making it difficult to capture early escape and local micro-drift changes in the liquid level area.

[0008] Fifth, the traditional abnormal signal capture and alarm process is a unidirectional linear structure with insufficient data closure and model self-evolution, making it impossible to continuously optimize the discrimination threshold and sensitivity parameters of weak signals during long-term operation. Summary of the Invention

[0009] To address the problems existing in the prior art, the present invention aims to provide an intelligent sensing and early warning method for resin level in an impregnation machine. This method employs an integrated approach combining broadband time-domain dielectric spectroscopy imaging, spatially distributed multi-point array detection, and multi-scale time-frequency analysis, enabling highly sensitive acquisition and dynamic amplification of local electromagnetic responses caused by minute fluctuations in resin level.

[0010] The present invention provides an intelligent sensing and early warning method for resin level in an impregnation machine, comprising the following steps:

[0011] S1. At multiple sampling locations in the resin level area within the impregnation machine, the radio frequency pulse response signals collected by the broadband dielectric sensing array are obtained to form a time-domain dielectric signal dataset with spatial distribution attributes.

[0012] S2. Perform noise suppression and signal normalization processing on the time-domain dielectric signal dataset to obtain a standardized time-domain dielectric signal sequence;

[0013] S3. Based on the standardized time-domain dielectric signal sequence, extract the dielectric constant variation characteristics of each sampling point at multiple time-frequency scales, and classify the data samples under different working conditions by feature;

[0014] S4. Based on the classified dielectric constant variation characteristics and their operating condition labels, a multi-operating condition adaptive baseline model is constructed using the historical liquid level response database.

[0015] S5. Input the standardized time-domain dielectric signal sequence obtained from actual sampling into the multi-condition adaptive baseline model, calculate the dynamic offset between the liquid level at each sampling point and the baseline response, and obtain the local abnormal trend signal of the liquid level.

[0016] S6. Perform multi-window spatiotemporal dynamic imaging processing on the local abnormal trend signal of liquid level to amplify the detection sensitivity of weak abnormal fluctuations, and calculate the amplitude, persistence and spatial diffusion characteristic parameters of the trend change.

[0017] S7. Input the magnitude, persistence and spatial diffusion characteristics of trend changes into the historical false alarm and artifact event feature library to identify and eliminate false weak anomaly signals.

[0018] S8. For the abnormal liquid level trend signal after removing artifacts, combine the composite pattern of upper and lower boundaries of liquid level, abrupt change starting point and slow drift trend to dynamically generate an adaptive early warning threshold.

[0019] Preferably, after step S8, the method further includes:

[0020] S9. Determine whether the abnormal liquid level trend signal after dynamic adaptive early warning threshold processing meets the early abnormal warning conditions. If it does, output the early abnormal resin level warning signal and start information synchronization of the remote operation and maintenance server when outputting the early abnormal resin level warning signal.

[0021] Preferably, after step S9, the method further includes:

[0022] S10: Feeds back the dielectric time-domain signal, abnormal trend parameters and warning threshold involved in the entire process of the warning to the liquid level response database. Based on the data feedback mechanism, it continuously optimizes the multi-condition adaptive baseline model and AI anomaly detection parameters to improve the sensitivity of early abnormal trends of weak signals.

[0023] Preferably, step S1 specifically includes:

[0024] Physical parameter analysis was performed on the spatial distribution of the resin level area within the impregnation machine to determine a multi-sampling location deployment scheme for the broadband dielectric sensor array.

[0025] Based on the deployed multi-sampling locations, a dielectric sensor array with wideband response characteristics is selected to apply radio frequency pulse excitation to the resin liquid level region, and the time-domain radio frequency response signal is acquired synchronously for each sensor node.

[0026] The original radio frequency response signals acquired at each sampling location are subjected to spatiotemporal synchronization and time correction processing, and a clock synchronization algorithm is used to unify the sampling timestamps;

[0027] By combining the spatial coordinate information of each sampling location, the time-domain dielectric response signal of the synchronization and time calibration is acquired and formatted in parallel through multiple channels to form a data frame with spatial distribution index.

[0028] The integrity and validity of the spatially distributed time-domain dielectric signal dataset acquired in the first round are checked. The signal quality evaluation module is used to identify sampling blind spots and abnormal noise sampling points, and suggestions for optimizing the sampling points and acquisition parameters are sent back.

[0029] Preferably, step S2 specifically includes:

[0030] The acquired raw time-domain dielectric signal dataset was processed by a multi-channel adaptive filtering algorithm to filter out power frequency interference and high-frequency environmental noise in broadband radio frequency sampling, and to obtain preliminary noise suppression results.

[0031] Using the preliminary noise suppression results as input, signal baseline correction is performed based on the endpoint detection and waveform baseline drift correction algorithm to obtain the time-domain dielectric signal result after baseline correction.

[0032] Multi-scale amplitude statistical analysis was performed on the baseline-corrected time-domain dielectric signal results to calculate the mean, variance, and extreme parameters between different sampling points and between multiple samplings of the same sampling point, and the amplitude normalization factor was dynamically estimated.

[0033] Based on the amplitude normalization factor, amplitude normalization mapping processing is performed on the baseline-corrected time-domain dielectric signal result to unify the signal amplitude of different sampling points to a standardized range, forming a preliminary output of a standardized time-domain dielectric signal sequence that eliminates the influence of environmental amplitude fluctuations.

[0034] Using the initial output of the standardized time-domain dielectric signal sequence as input, individual abnormal points or short-term missing points are detected and corrected to generate the final standardized time-domain dielectric signal sequence.

[0035] Preferably, step S3 specifically includes:

[0036] The standardized time-domain dielectric signal sequence is subjected to multi-scale time-frequency decomposition processing, and the local frequency domain response features of each sampling point at different time resolutions are extracted by wavelet transform algorithm to obtain a multi-scale dielectric spectrum parameter set.

[0037] Based on the multi-scale dielectric spectrum parameter set, the main dielectric constant variation characteristic parameters are generated by applying the extracted characteristic power spectral density and principal component analysis algorithm to each sampling point.

[0038] The dielectric constant variation characteristic parameters of each sampling point are aligned with the resin type, temperature and impurity condition labels stored in the device to generate a feature fusion sample set of related operating conditions.

[0039] Feature selection and normalization are performed on the feature fusion sample set to retain the core dielectric constant variation parameters that are sensitive to operating conditions and have discriminative power, and feature dimension compression is performed.

[0040] Based on the core dielectric constant variation parameter after dimensional compression and the operating condition label, the data samples of all sampling points are grouped and classified according to different operating condition categories such as resin type, temperature range and impurity concentration to obtain a standard feature category set under multiple operating conditions.

[0041] Preferably, step S4 specifically includes:

[0042] A standardized vector mapping process is performed on the dielectric constant variation characteristics and their operating condition labels to generate a multidimensional feature matrix of dielectric constant under cross-operating conditions;

[0043] Based on the dielectric constant multidimensional feature matrix, multi-type working condition sample data from the historical liquid level response database are called, and the historical sequence of liquid level electromagnetic response under the associated working condition is obtained through feature solidification and tag indexing.

[0044] The liquid level electromagnetic response history sequence is grouped according to resin type, temperature, and impurity concentration parameters to obtain the characteristic output response mode under each working condition.

[0045] By utilizing the characteristic output response patterns under various operating conditions, adaptive parameter estimation under multiple operating conditions is performed, and a set of adaptive baseline parameters under multiple operating conditions is output.

[0046] The set of multi-condition adaptive baseline parameters is input into the model optimization module to generate a multi-condition adaptive baseline model that can be automatically adjusted according to the operating conditions.

[0047] Preferably, step S5 specifically includes:

[0048] Using the standardized time-domain dielectric signal sequence as input, the standardized time-domain dielectric signal of each sampling point is sequentially input into the multi-condition adaptive baseline model to obtain the standardized baseline response signal corresponding to each sampling point.

[0049] Based on the standardized baseline response signal, a point-by-point dynamic offset calculation is performed on the standardized time-domain dielectric signal sequence of each sampling point and its standardized baseline response signal to generate a liquid level dynamic offset sequence.

[0050] Using the dynamic offset sequence as a factor, multi-scale spatiotemporal feature extraction is performed to refine the dynamic offset evolution characteristics of each sampling point at different time scales and regions, thereby obtaining the basic signal of local anomaly trend in the liquid level area.

[0051] Adaptive feature normalization processing is performed on the basic signal of the local anomaly trend to output the local anomaly trend signal of the liquid level normalized to the operating condition.

[0052] The normalized local anomaly trend signals of liquid level are summarized into a spatial distribution anomaly trend map, forming the basic input for dynamic imaging analysis.

[0053] Preferably, step S6 specifically includes:

[0054] For local abnormal trend signals of liquid level, a multi-scale time-domain sliding window signal sequence is generated based on an adaptive window partitioning strategy, and the sliding window signal set is output.

[0055] For the sliding window signal set, a multi-window spatiotemporal distribution mapping map is established for the abnormal trend response data of different sampling points within each time domain window;

[0056] Weak signal enhancement and noise suppression processing are performed on the generated multi-window spatiotemporal distribution map. Then, for the noise-suppressed spatiotemporal distribution signal template, the amplitude, persistence, and spatial diffusion characteristic parameters of the trend change within each window are calculated.

[0057] For all calculated trend change magnitude, persistence, and spatial diffusion characteristic parameters, a globally standardized trend feature vector is generated based on multi-scale aggregation processing.

[0058] Preferably, step S7 specifically includes:

[0059] Based on the trend change magnitude, trend change persistence, and spatial diffusion characteristic parameters, feature vector assembly is performed to form a trend feature parameter set;

[0060] Based on the trend feature parameter set, retrieve and call the historical false alarm and artifact event feature library, perform similarity judgment between the input feature parameter set and the artifact event features in the library, and obtain preliminary false weak anomaly signal matching results;

[0061] The preliminary pseudo-weak anomaly signal matching results are divided into multidimensional clusters and classified into specific category groups based on the distribution relationship of the trend feature parameter set, thereby identifying potential periodic disturbances and external noise patterns.

[0062] Based on the clustering results, the statistical confidence of the category to which the trend feature parameter set belongs is evaluated, and feature removal processing is performed on the trend feature parameter set that is identified as a high probability of false weak anomaly signal, and the removed liquid level anomaly trend feature parameters are output.

[0063] The intelligent sensing and early warning method for resin level in an impregnation machine described in this invention has the following advantages:

[0064] 1. This invention employs an integrated method combining broadband time-domain dielectric spectroscopy imaging, spatially distributed multi-point array detection, and multi-scale time-frequency analysis. This method enables highly sensitive acquisition and dynamic amplification of local electromagnetic responses caused by minute fluctuations in resin levels. Compared to conventional single-point level sensors or methods that rely solely on threshold fluctuations, the system, through multi-dimensional feature extraction and spatial imaging, can effectively amplify and separate weak abnormal signals in the initial stages of level changes.

[0065] 2. This invention innovatively introduces a historical false alarm / spurious event feature library, and combines unsupervised clustering discrimination and discriminant analysis to characterize and accurately eliminate false weak abnormal signals caused by non-liquid level reasons such as periodic disturbances and external noise of equipment, effectively suppressing the occurrence of false alarms.

[0066] 3. This invention utilizes feature normalization based on multiple tags such as dielectric parameters, resin type, temperature, and impurities, along with adaptive baseline modeling using a multi-condition database. This allows for online dynamic adjustment of the electromagnetic response baseline, enabling real-time adaptation of the model to different material properties and external environments. Compared to traditional static reference lines or rule bases, it maintains high accuracy and robustness in anomaly detection even under changing actual operating conditions (such as raw material batch changes or production temperature fluctuations), effectively extending the equipment maintenance prediction window and reducing unexpected downtime.

[0067] 4. This invention employs an AI-enhanced multi-target anomaly trend detection algorithm, which jointly considers various temporal anomaly modes such as the upper and lower boundary breaches of liquid level, abrupt change initiation, and slow drift, and combines them with a dynamic threshold adaptive adjustment algorithm to achieve dynamic early warning under weak signal conditions.

[0068] 5. This invention utilizes structured data feedback of all parameters throughout the early warning process to continuously optimize the adaptive baseline model and AI decision network parameters. The system can iteratively improve detection sensitivity and false alarm suppression performance based on actual operational feedback, forming a positive feedback loop of data-model-parameter self-evolution. With long-term use, it can maintain or even improve the system's accuracy in responding to weak anomalies, solving the performance degradation problem caused by environmental factors and aging in conventional systems. Attached Figure Description

[0069] Figure 1 This is a flowchart of a method for intelligent sensing and early warning of resin level in an impregnation machine, as described in this invention. Detailed Implementation

[0070] like Figure 1 As shown, the present invention provides an intelligent sensing and early warning method for resin level in an impregnation machine, comprising the following steps:

[0071] S1. At multiple sampling locations in the resin level area within the impregnation machine, the radio frequency pulse response signals collected by the broadband dielectric sensing array are obtained to form a time-domain dielectric signal dataset with spatial distribution attributes.

[0072] S2. Perform noise suppression and signal normalization processing on the time-domain dielectric signal dataset to obtain a standardized time-domain dielectric signal sequence;

[0073] S3. Based on the standardized time-domain dielectric signal sequence, extract the dielectric constant variation characteristics of each sampling point at multiple time-frequency scales, and classify the data samples under different working conditions by feature;

[0074] S4. Based on the classified dielectric constant variation characteristics and their operating condition labels, a multi-operating condition adaptive baseline model is constructed using the historical liquid level response database.

[0075] S5. Input the standardized time-domain dielectric signal sequence obtained from actual sampling into the multi-condition adaptive baseline model, calculate the dynamic offset between the liquid level at each sampling point and the baseline response, and obtain the local abnormal trend signal of the liquid level.

[0076] S6. Perform multi-window spatiotemporal dynamic imaging processing on the local abnormal trend signal of liquid level to amplify the detection sensitivity of weak abnormal fluctuations, and calculate the amplitude, persistence and spatial diffusion characteristic parameters of the trend change.

[0077] S7. Input the magnitude, persistence and spatial diffusion characteristics of trend changes into the historical false alarm and artifact event feature library to identify and eliminate false weak anomaly signals.

[0078] S8. For the abnormal liquid level trend signal after removing artifacts, combine the composite pattern of upper and lower boundaries of liquid level, abrupt change starting point and slow drift trend to dynamically generate an adaptive early warning threshold.

[0079] Furthermore, in this embodiment, step S8 is followed by:

[0080] S9. Determine whether the abnormal liquid level trend signal after dynamic adaptive early warning threshold processing meets the early abnormal warning conditions. If it does, output the early abnormal resin level warning signal and start information synchronization of the remote operation and maintenance server when outputting the early abnormal resin level warning signal.

[0081] Furthermore, in this embodiment, step S9 is followed by:

[0082] S10: Feeds back the dielectric time-domain signal, abnormal trend parameters and warning threshold involved in the entire process of the warning to the liquid level response database. Based on the data feedback mechanism, it continuously optimizes the multi-condition adaptive baseline model and AI anomaly detection parameters to improve the sensitivity of early abnormal trends of weak signals.

[0083] Furthermore, in this embodiment, step S1 specifically includes:

[0084] Physical parameter analysis was performed on the spatial distribution of the resin liquid level area in the impregnation machine to determine the multi-sampling location deployment scheme of the broadband dielectric sensor array, so as to cover all key measurement points in the liquid level area and provide a basis for the subsequent formation of a time-domain dielectric signal dataset with spatial distribution attributes.

[0085] Using the three-dimensional physical structure parameters and process constraints of the resin liquid level area in the impregnation machine as input, a multi-dimensional physical parameter analysis is conducted on the spatial distribution characteristics of the overall liquid level area and key sub-regions.

[0086] The finite element spatial modeling method (parameters: equipment cavity size, expected resin fluctuation height range, potential liquid level anomaly points, structural obstruction, etc.) is used to achieve accurate division of the measurable area boundary of the entire liquid level region.

[0087] By using statistical analysis and fluid dynamic simulation methods (parameters: resin properties, flow distribution, temperature field, historical leakage trend), sensitive zoning data of areas prone to triggering abnormal signals are obtained, and the results of priority point placement area division are obtained.

[0088] By using a statistical coverage optimization algorithm (maximum coverage and minimum point distribution principle), the redundancy and accessibility constraints of spatial distribution are comprehensively evaluated to form a preliminary set of multi-sampling locations and generate coverage and blind spot risk analysis indicators.

[0089] For the above set of multiple sampling locations, a process safety isolation and actual installation feasibility analysis are introduced (parameters: equipment safety distance, electromagnetic interference intensity, sensor maintainability). A geometric misalignment and safety buffer review mechanism is adopted to screen out undeployable or high-risk areas and output an implementable multi-sampling location layout scheme.

[0090] The above-mentioned deployment scheme provides a spatial distribution basis covering all key measurement points for the subsequent precise deployment of broadband dielectric sensing arrays and acquisition of radio frequency excitation signals, and realizes the requirement for time-domain dielectric signal dataset modeling of the spatial distribution attributes of liquid level area.

[0091] For example, in a medium-sized impregnation machine with an inner cavity length of 1200mm, width of 600mm, and height of 400mm, when performing physical parameter analysis on the liquid level area inside the resin tank, a finite element three-dimensional model is first introduced, and the liquid level fluctuation sensitive area is set to be 50mm to 350mm from the bottom.

[0092] Fluid dynamics simulation shows that the two ends and the middle of the equipment are the leakage-prone points and the low-velocity flow field, respectively. After adopting the maximum coverage minimum sampling point algorithm, 9 spatial sampling points are selected (including 3 end corner points, 3 middle height difference points and 3 random redundancy compensation points), and the coverage efficiency reaches 98%.

[0093] Further considering the equipment wall thickness, electrical safety distance of at least 20mm, and sensor maintainability, one point near the high-voltage interface and two points obstructed by the structure in the initial layout plan were eliminated. Finally, the coordinates of the seven layout points were confirmed as (100,50,60), (100,50,340), (600,50,200), (600,350,200), (1150,50,60), (1150,50,340), and (600,200,370).

[0094] Under this deployment scheme, after on-site deployment and sampling signal quality testing, the noise level generated at each point is lower than -65dBm, and the resin fluctuation response in all key areas meets the spatial distribution criteria. Finally, a spatially distributed sampling coordinate set is output for dynamic liquid level anomaly trend modeling, realizing spatial coverage of all key measurement points of resin liquid level and safe deployment under physical parameter constraints.

[0095] Based on the deployed multi-sampling locations, a dielectric sensor array with wideband response characteristics is selected to apply radio frequency pulse excitation to the resin liquid level region, and the time-domain radio frequency response signal is acquired synchronously for each sensor node.

[0096] Using a spatially optimized set of multi-sampling position coordinates as input, a wideband response dielectric sensor array is deployed for each specific coordinate point, wherein each sensor is configured with a corresponding radio frequency pulse excitation interface.

[0097] The radio frequency pulse excitation method (parameters: excitation pulse width 5ns~20ns, center frequency 200MHz-3GHz, excitation amplitude 20Vpp) is used to synchronously excite the dielectric sensor at each sampling position, so as to achieve multi-point parallel excitation of the local resin dielectric response while covering the key measurement areas of liquid level in the equipment.

[0098] By using a synchronous high sampling rate ADC data acquisition module (parameters: sampling rate not less than 1GS / s, quantization accuracy more than 12 bits), the time-domain dielectric radio frequency response signal of each sensor node after excitation is acquired to obtain the initial response voltage signal sequence at the spatial distribution point.

[0099] A time-domain windowing and pre-filtering algorithm (parameters: active pre-charge amplifier bandwidth 1GHz, bandpass filter cutoff frequency 100MHz-2GHz) is used to perform time-domain truncation and noise suppression processing on the acquired raw response voltage signal, suppressing broadband interference components in the radio frequency environment and ensuring the identifiability of the small liquid level change signal.

[0100] By using an integrated excitation-response synchronous triggering mechanism, the radio frequency response signal of each sampling node is time-base aligned with the excitation pulse timing to obtain the original time-domain dielectric response data corresponding to each sensor node, ensuring the synchronization and comparability of signal data between different spatial measurement points.

[0101] Through the above acquisition and processing process, the spatially distributed broadband radio frequency response signal is converted into raw dielectric response data carrying information on minute changes in resin level. This establishes a high-quality data foundation for subsequent spatiotemporal synchronization, multi-channel encapsulation, and dynamic anomaly extraction, achieving a synergistic improvement in liquid level sensing sensitivity and spatial coverage.

[0102] For example, in the above-described immersion machine equipment, a broadband dielectric sensor array of model DBE1000 is deployed at seven spatial sampling positions. The excitation pulse width of a single sensor is set to 8ns, the center excitation frequency is 900MHz, the excitation timing adopts board-level synchronous distribution, and the sampling period is 2ms.

[0103] The RF response of all sensor nodes is acquired in parallel by a 16-bit 1GS / s high-speed ADC, the bandwidth of the pre-charge amplifier is set to 1.2GHz, and the bandpass filter bandwidth is limited to 150MHz to 950MHz.

[0104] Through synchronous excitation triggering, the response time deviation of the seven sensors is controlled within 2ns, and the original response waveform data are all acquired at full amplitude.

[0105] In real-world conditions, the amplitude fluctuation of the original radio frequency response signal caused by liquid level changes is 0.15–0.35V, and the signal-to-noise ratio is higher than 45dB. After pre-filtering and synchronization timing alignment, the original dielectric response curves collected from different spatial points maintain high time base consistency within a 200ns window. Through this step, a high-quality, synchronized, and spatially well-distributed original radio frequency dielectric response dataset is formed, laying a solid technical foundation for high-sensitivity sensing of minute fluctuations in resin liquid level.

[0106] The original radio frequency response signals obtained from each sampling location are subjected to spatiotemporal synchronization and time correction processing. A clock synchronization algorithm is used to unify the sampling timestamps to ensure high comparability of multi-node signal data and to establish a unified time reference for spatial distribution correlation analysis.

[0107] By combining the spatial coordinate information of each sampling location, the time-domain dielectric response signal of the synchronization and time calibration is acquired and formatted in parallel through multiple channels to form a data frame with spatial distribution index, ensuring that the signal attributes correspond one-to-one with the physical spatial points, and realizing the preliminary modeling of the time-domain dielectric signal dataset with spatial distribution attributes.

[0108] The integrity and validity of the spatially distributed time-domain dielectric signal dataset acquired in the first round were checked. The signal quality evaluation module was used to identify sampling blind spots and abnormal noise sampling points, and suggestions for optimizing the sampling points and acquisition parameters were sent back to provide a basis for continuous optimization for subsequent large-scale data acquisition and dynamic signal tracking.

[0109] Furthermore, in this embodiment, step S2 specifically includes:

[0110] The acquired raw time-domain dielectric signal dataset is processed by a multi-channel adaptive filtering algorithm to filter out power frequency interference and high-frequency environmental noise in broadband radio frequency sampling, obtain preliminary noise suppression results, and significantly enhance the true value components of the time-domain dielectric signal at the sampling points.

[0111] Using the preliminary results of noise suppression as input, and based on the endpoint detection and waveform baseline drift correction algorithm, signal baseline correction is performed to obtain the time-domain dielectric signal result after baseline correction; it can compensate for low-frequency time-varying offsets introduced by factors such as sampling circuit or external temperature drift;

[0112] Multi-scale amplitude statistical analysis was performed on the baseline-corrected time-domain dielectric signal results to calculate the mean, variance, and extreme parameters between different sampling points and between multiple samplings of the same sampling point, and the amplitude normalization factor was dynamically estimated.

[0113] Based on the amplitude normalization factor, amplitude normalization mapping processing is performed on the baseline-corrected time-domain dielectric signal results to unify the signal amplitudes of different sampling points into a standardized range, forming a preliminary output of a standardized time-domain dielectric signal sequence that eliminates the influence of environmental amplitude fluctuations.

[0114] Using the initial output of the standardized time-domain dielectric signal sequence as input, individual abnormal points or short-term missing points are detected and corrected to generate the final standardized time-domain dielectric signal sequence.

[0115] Furthermore, in this embodiment, step S3 specifically includes:

[0116] A multi-scale time-frequency decomposition process is performed on the standardized time-domain dielectric signal sequence, and the local frequency domain response features of each sampling point at different time resolutions are extracted by wavelet transform algorithm to obtain a multi-scale dielectric spectrum parameter set.

[0117] Using a standardized time-domain dielectric signal sequence as input, the signals at each sampling point have undergone baseline correction, amplitude normalization, and outlier repair, and are comparable across time and space.

[0118] A multi-scale wavelet transform algorithm (using compactly supported orthogonal wavelet bases such as Daubechies and Symlet, with the number of decomposition levels determined based on the signal sampling rate and target bandwidth) is employed to decompose the standardized time-domain dielectric signal at each sampling point, thereby enabling the unfolding of local features of the signal at different time resolutions.

[0119] By calculating the frequency domain response characteristics of the signal at multiple scales using the wavelet component sequences obtained at each scale (i.e., different wavelet decomposition layers), the multidimensional feature parameters of each sampling point at the local detail scale and the global trend scale are obtained.

[0120] The wavelet energy spectrum statistical method is used to calculate the scale energy characteristics of the wavelet components at each scale for each sampling point:

[0121]

[0122] Among them, E j Let E be the wavelet energy at the j-th scale, and N be the number of coefficients at that scale. j Energy distribution reflects the energy concentration properties of the liquid level dielectric signal in different frequency bands;

[0123] For each wavelet component at each scale, morphological statistical feature parameters are extracted, including maximum amplitude, root mean square value, kurtosis, skewness, and energy normalization distribution, to provide a rich representation of the local frequency domain response.

[0124] Through the above wavelet multi-scale time-frequency decomposition process, the standardized time-domain dielectric signal sequence is mapped to a multi-dimensional dielectric spectrum parameter set at multiple time-frequency scales for each sampling point, thereby improving the resolution and characterizing the dynamic characteristics of the weak signal changes in resin level.

[0125] For example, in the scenario of resin level monitoring in an impregnation machine, a standardized time-domain dielectric signal is acquired at a sampling rate of 1 MSps, with a sampling length of 4096 points per point. For each sampling point, a 5-level wavelet decomposition is performed using the Daubechies-4 wavelet basis to obtain wavelet components and approximate components at scales 1 to 5. For each wavelet component, nine statistical features, including scale energy, root mean square value, skewness, and kurtosis, are calculated to form a 450-dimensional multi-scale dielectric spectrum parameter set of 10 points × (5 levels × 9 features). Under different resin, temperature, and impurity concentration conditions, The wavelet energy features at scale 2 (corresponding to the signal frequency range of 2-8kHz) and scale 4 (frequency range of 32-128kHz) are particularly sensitive to minute changes in resin level. After wavelet decomposition, the high-frequency pseudo-variables caused by noise are manifested as discontinuous energy or abnormal skewness in the scale distribution. Related interference can be significantly eliminated through feature screening. The final output multi-scale dielectric spectrum parameter set comprehensively covers different time and frequency ranges of abnormal liquid level fluctuations, which strongly supports subsequent principal component feature extraction and operating condition classification operations. In actual testing, it improves the sensitivity and robustness of weak signal trend capture.

[0126] Based on the multi-scale dielectric spectrum parameter set, the characteristic power spectral density extraction and principal component analysis algorithm are applied to each sampling point to generate its main dielectric constant variation characteristic parameters, thereby realizing a quantitative description of the time-frequency bidirectional spatial variation.

[0127] Using the multi-scale dielectric spectrum parameter set output above as the input object, it covers the frequency domain characteristic parameters of each sampling point at different frequencies and time scales obtained after wavelet transform and other time-frequency decomposition algorithms;

[0128] The Power Spectral Density (PSD) analysis method was used to perform Fast Fourier Transform (FFT) on the time-domain dielectric signals at multiple scales at each sampling point, and the PSD value was calculated using the following formula:

[0129]

[0130] Where x(n) is the discrete sequence of time-domain dielectric signal after baseline correction and normalization at the sampling point, N is the number of sampling points, and f is the analysis frequency point;

[0131] By calculating the PSD, the main power distribution parameters of each sampling point at a preset frequency resolution are obtained, thereby realizing the extraction of local energy features;

[0132] For power spectral parameters at multiple frequencies and time scales, the Principal Component Analysis (PCA) algorithm is applied to map the high-dimensional spectral parameter matrices at all scales to a standardized low-dimensional feature space.

[0133] The PCA processing chain includes: data centralization, covariance matrix construction, eigenvalue and eigenvector decomposition, determination of the number of principal components (usually based on a cumulative variance contribution rate greater than 95% or reconstruction error less than a threshold), and obtaining the principal component features using the following formula:

[0134] Z = X × W

[0135] Where X is the m-dimensional power spectrum parameter matrix of the n sampling points to be processed, W is the projection matrix composed of the first k principal component vectors selected from the m-dimensional eigenvectors, and Z is the set of main dielectric constant variation characteristic parameters after dimensionality reduction;

[0136] During PCA, redundant information is compressed and the sensitivity to minute changes in liquid level is enhanced by quantifying the variance and correlation of spectral parameters in each dimension.

[0137] By standardizing the principal component features obtained by PCA to a unified range, the amplitude benchmark differences between different sampling points are eliminated, ensuring a unified discrimination standard for subsequent classification and comparison.

[0138] By combining the above-mentioned characteristic power spectral density extraction and principal component analysis algorithm, the high-dimensional, multi-time-frequency scale parameters of the previous sub-step are transformed into the main dielectric constant variation characteristic parameters of each sampling point, forming a feature vector that can comprehensively reflect the minute changes in liquid level and has discriminative properties, thereby realizing a quantitative description of the time-frequency bidirectional change of the resin level in the impregnation machine.

[0139] For example, in the process of analyzing the dielectric signal of the resin level in the impregnation machine, a radio frequency pulse excitation with a sampling rate of 1MSps is used to acquire the standardized time-domain dielectric signal of 10 sampling points, with each point having a sampling length of 4096 points;

[0140] After applying FFT, the 0-500kHz bandwidth was divided into 200 frequency segments, and the power spectrum distribution of each sampling point was obtained by PSD calculation.

[0141] Using a 10-point × 200-dimensional parameter matrix as PCA input, after centering and covariance matrix eigenvalue decomposition, the top 7 principal components (k = 7) with a cumulative variance contribution rate > 95% are selected. The original 200-dimensional power spectrum features of each sampling point are reduced to 7-dimensional main dielectric constant variation feature parameters. Subsequent standardization normalizes the 7-dimensional features to the [0,1] interval.

[0142] In practical applications, for four different resins, two temperatures, and three impurity concentrations, the changes in the principal component characteristic parameters were compared, and it was found that the correlation coefficient between the first principal component of PCA and the minor anomaly in liquid level was as high as 0.92, which is significantly better than the discrimination of the original high-dimensional spectral parameters.

[0143] The main variable characteristic parameters output serve as important inputs for multi-condition feature model classification and abnormal trend identification, verifying the excellent effect of this step in improving the sensitivity of weak signal detection, suppressing condition-related noise, and reducing the risk of false alarms.

[0144] The dielectric constant variation characteristic parameters of each sampling point are aligned with the resin type, temperature and impurity condition labels stored in the device to generate a feature fusion sample set of related operating conditions.

[0145] Using the main dielectric constant variation characteristic parameters of each sampling point (the feature vector formed after principal component analysis and standardization) as input objects, the system retrieves the operating condition label dataset stored on the device, including multi-label environmental information such as resin type, temperature, and impurity conditions.

[0146] A data alignment algorithm (parameter settings: dual indexing of unique sampling point ID and collection timestamp) is adopted to achieve one-to-one pairing of feature parameters of each sampling point with corresponding working condition labels, ensuring that the physical state and environmental description between samples correspond accurately.

[0147] By using multi-label data annotation methods (such as One-Hot encoding and multi-label mapping matrix method), each feature parameter record is assigned a full set of working condition labels, thereby achieving data fusion mapping of multi-dimensional labels.

[0148] By applying feature fusion methods (such as feature-level splicing and tag feature co-coding), the dielectric constant variation feature parameter and the working condition label are embedded into a unified feature fusion sample vector to form a multi-working-condition, full-label feature fusion sample set.

[0149] By verifying feature consistency and repairing missing labels (such as using a sample neighborhood average interpolation algorithm), the label information of all samples is ensured to be complete, thus achieving full label integrity encapsulation of the fused sample set.

[0150] Through the above multi-label alignment and feature fusion algorithm, the standardized dielectric constant variation feature dataset output in the previous step is transformed into a multi-label, strongly condition-related feature fusion sample set, thereby achieving the technical goal of expressing the dielectric properties of resin level under different working conditions in a unified feature space.

[0151] For example, in the resin level monitoring application of an impregnation machine, after PCA dimensionality reduction, 10 sampling points and 7 key dielectric constant variation characteristic parameters are obtained. The equipment's historical configuration information database is called to clearly define the operating condition label for each sampling point as "Resin A, Temperature 40℃, Impurity Concentration 0.1%". The sampling point ID and time index are hash-mapped and aligned with the operating condition table to form a joint data table. One-Hot encoding is used to map 4 resin types to 4-bit labels, 2 temperatures to 2-bit labels, and 3 impurity concentrations to 3-bit labels. The labels are merged into the features of the sampling points to form a 16-dimensional feature fusion vector. Through the feature fusion module, 10 sampling points output 10 16-dimensional feature fusion sample records, and all record label fields are complete. In practical applications, statistical analysis of sample sets under different working conditions by label shows that sampling points with labels of high impurity concentration (0.3%) have lower PCA principal component amplitudes. Combined with subsequent clustering and classification operations, this provides a rich working condition basis for judging the abnormal trend of resin liquid level, realizing adaptive labeling and high sensitivity discrimination of weak abnormal signals of resin liquid level under multiple working conditions.

[0152] Feature selection and normalization are performed on the feature fusion sample set to retain the core dielectric constant variation parameters that are sensitive to operating conditions and have discriminative power, and feature dimension compression is performed.

[0153] Based on the core dielectric constant variation parameter after dimensional compression and the operating condition label, the data samples of all sampling points are grouped and classified according to different operating condition categories such as resin type, temperature range and impurity concentration to obtain a standard feature category set under multiple operating conditions.

[0154] Furthermore, in this embodiment, step S4 specifically includes:

[0155] Normalized vector mapping is performed on the dielectric constant variation characteristics and their operating condition labels to generate a multi-dimensional feature matrix of dielectric constant under cross-operating conditions. This feature matrix serves as the data input for constructing a multi-operating condition adaptive baseline model.

[0156] Using the compressed core dielectric constant variation parameter and operating condition labels as input, a standardized vector mapping method is employed to generate a multi-dimensional feature matrix across operating conditions. The Z-Score standardization algorithm (formula: z) is selected. i =(x i -μ) / σ, where x i For each feature component, μ is the mean of the feature in the sample set, and σ is the standard deviation. The mean-reduction and standardization transformations are applied to all dielectric constant variation parameters to ensure that each feature dimension has a common amplitude reference and zero mean attribute.

[0157] By using One-Hot encoding or multi-label embedding (parameter settings: resin type, temperature, and impurity concentration are encoded independently), the feature space is expanded according to the category field, and the working condition label information is added to the standardized feature vector in vector form to achieve deep fusion of features and working condition information.

[0158] Furthermore, a feature concatenation and assembly algorithm is employed to concatenate the compressed principal component feature vector with the One-Hot condition label vector in a prescribed order to construct a joint input vector:

[0159]

[0160] Wherein, λ is the label vector weighting factor, used to adjust the importance ratio of the working condition label and dielectric feature in the input matrix;

[0161] Assemble the joint input vector of all sampling points to generate an n×m multidimensional feature matrix F of dielectric constant, where n is the number of samples and m is the sum of the dielectric feature dimension and the label encoding dimension. Employ sparsity normalization and numerical range mapping to normalize all features of F to the [0,1] interval, improving the stability and adaptability of the dataset in subsequent model training;

[0162] Through the above-mentioned standardized vector mapping and feature fusion processing, the dielectric features of multiple working conditions and the working condition labels are uniformly encoded in the same standard feature space, and the output is used as the input feature matrix of the multi-working-condition adaptive baseline model, providing a highly consistent and highly identifiable data foundation for subsequent association with historical databases and parameter fitting modeling.

[0163] For example, the 7-dimensional dielectric constant variation characteristic parameters obtained by principal component analysis for dimensionality reduction are spliced ​​with the 4-dimensional resin type, 2-dimensional temperature, and 3-dimensional impurity concentration tags (a total of 9 dimensions) encoded by One-Hot, with a weighting factor λ = 0.5.

[0164] The standardized feature values ​​of the 10 sampling points are all in the range of [0,1]; after summarizing, a multidimensional feature matrix of 10 rows and 16 columns is formed.

[0165] During the test, the first to third dimensions of the standardized principal components were significantly lower than those under clean and high-temperature conditions under high-impurity and low-temperature conditions. The One-Hot label accurately identified all operating condition attributes.

[0166] After subsequent recursive least squares modeling and cluster analysis, this feature matrix can achieve high-resolution dynamic adjustment of the liquid level electromagnetic response baseline under multiple operating conditions. Actual verification shows that it supports a more than 10% improvement in sensitivity to weak anomalies.

[0167] Based on the multidimensional feature matrix of dielectric constant, various types of working condition sample data are called from the historical liquid level response database. Through feature solidification and tag indexing, the historical sequence of liquid level electromagnetic response under related working conditions is obtained to enrich the model sample base.

[0168] The dielectric constant multidimensional feature matrix generated by standardized vector mapping is used as input data to obtain the feature basis for multi-condition adaptive baseline modeling.

[0169] Call the historical liquid level response database interface, and perform multi-label condition search based on the working condition label parameters (such as resin type, temperature range, impurity concentration) embedded in the dielectric constant multidimensional feature matrix to filter out historical sample sets that are completely consistent with the input working condition or have a similarity higher than the preset threshold.

[0170] The feature solidification algorithm is used to perform fixed value locking on the dielectric constant variation feature parameters in the retrieved historical sample set, and convert the original response data in the historical records into structured feature vectors with amplitude and label normalized, which are consistent with the current feature space scale.

[0171] Furthermore, by using the label index mapping method, the liquid level electromagnetic response data of each sampling point in the historical samples are mapped one by one with the corresponding operating condition labels, filling in the information of missing labels, ensuring seamless alignment between the full sample features and operating condition parameters, and forming a set of historical response sample sequences with clear categories.

[0172] Using time-series data extraction and sequence reconstruction algorithms, historical samples of liquid level electromagnetic response under each working condition category are arranged in order of timestamp to generate a data package of historical sequence of liquid level electromagnetic response under multi-label constraints.

[0173] Through the above-mentioned database interface calls, sample solidification, label indexing and time series processing, the multi-condition historical liquid level electromagnetic response dataset is transformed into a structured and standardized condition response sequence, providing a highly consistent and rich model sample foundation for subsequent condition-related cluster analysis and adaptive parameter modeling, and achieving comprehensive prior coverage of the model of liquid level electromagnetic response characteristics under different resin types and external environments.

[0174] For example, in the resin level intelligent sensing system of the impregnation machine, the input dielectric constant multidimensional feature matrix contains 7-dimensional principal component features and 9-dimensional One-Hot condition label vectors, which are evenly distributed in the [0,1] interval after normalization mapping;

[0175] The historical liquid level response database stores 15,000 liquid level status samples from the past three years, covering 10 resin types, 4 temperature gradients, and 5 impurity concentration levels. When performing a multi-label conditional search, if the target input label is "Resin A - Temperature 40℃ - Impurities 0.1%", the database retrieves 537 completely matching historical samples.

[0176] The feature solidification algorithm was used to map the historical original liquid level electromagnetic response data of these 537 samples to the current principal component space unified coordinate system, and missing labels were added to generate 537 standardized structured sample vectors. Then, they were sorted by collection time and reconstructed into a 537-point time series historical sample package.

[0177] The extracted historical sequence samples contain complete electromagnetic response curves, full label information, and feature consistency guarantees, which meet the input requirements for subsequent recursive parameter modeling and working condition clustering. The implementation of this step ensures that there are sufficient and highly confident training samples for the model to adapt to diverse working conditions, thereby improving the generalization ability of the multi-working-condition liquid level dynamic baseline model and the sensitivity of early abnormal trend detection.

[0178] The historical sequence of liquid level electromagnetic response was grouped according to resin type, temperature, and impurity concentration parameters to obtain the characteristic output response mode under each working condition.

[0179] Using historical liquid level electromagnetic response sequences as input, and combining the generated dielectric constant multidimensional feature matrix and its contained operating condition labels, feature grouping and clustering modeling are achieved.

[0180] A multi-condition indexing algorithm for operating condition labels (parameters: resin type, temperature range, impurity concentration encoding) is adopted to achieve preliminary classification of operating condition attributes of historical liquid level electromagnetic response samples, and the sample set is divided into several operating condition candidate groups according to different label combinations.

[0181] Fine-grained operating condition normalization is performed using continuous multivariate clustering analysis. K-means clustering or a Gaussian mixture model is employed (parameter: the number of clusters K is adaptively determined by the number of label combinations and the statistical silhouette coefficient). Using the principal component parameter of dielectric constant as the feature vector, the historical response curves of each candidate operating condition group are aggregated into a consistent feature cluster based on spatiotemporal characteristic distance. Its mathematical expression is as follows:

[0182]

[0183] Where J is the clustering objective function, S i For the i-th cluster, x j Belongs to S i The sample feature vector, μ i For S i The mean center of the in-sample;

[0184] By using the curve alignment normalization algorithm (parameter: Dynamic Time Warping (DTW) distance threshold), the time sequence pattern alignment and amplitude normalization of the liquid level electromagnetic response time series in each cluster are performed to eliminate the slight offset caused by sampling abnormalities or environmental disturbances in different batches of signals under the same working conditions, thereby achieving a highly consistent expression of curve features.

[0185] Furthermore, using a clustering template modeling method, the mean response curve, standard deviation, and distribution interval of the curve group corresponding to each cluster are statistically calculated, and the following characteristic output response pattern is output:

[0186]

[0187] in, Let N be the mean liquid level response curve for the i-th operating condition cluster. i R represents the number of historical samples within this cluster. ij (t) represents the time-series signal of the j-th group of historical responses in this cluster;

[0188] Through the above-mentioned algorithms such as working condition label indexing, feature clustering, curve normalization and feature template extraction, the historical liquid level electromagnetic response sequence is transformed into a clustered feature output response mode that satisfies the working condition normalization characteristics, realizing the standardized representation of the liquid level electromagnetic response baseline under each working condition, and providing a quantitative template for subsequent multi-working-condition adaptive parameter estimation.

[0189] For example, for 12 combined operating conditions such as resin type A / B / C / D, temperature 30℃ / 40℃, and impurity concentration 0.1% / 0.2% / 0.3%, a total of 4,000 samples of standardized dielectric characteristic principal components and historical liquid level response curves were extracted from the historical database.

[0190] For each candidate group of working conditions, K-means clustering with label index is used (K = 2 to 4, adaptively selecting the optimal number of clusters), and the Davies-Bouldin index < 0.5 is used as the clustering convergence criterion.

[0191] For the thousands of time series extracted from each cluster, the DTW algorithm is used to align the center curve and normalize the amplitude to the [0,1] interval. Based on 400 samples in each cluster, the mean response curve and the 95% interval standard deviation band are calculated to obtain the feature output template for each working condition class.

[0192] In practical applications, the residual of the curve mean template is <0.03, and the clustered standard deviation band covers 98% of the historical response variation, achieving effective normalization of the liquid level response under working conditions such as resin type, temperature and impurities.

[0193] The clustering matching rate is greater than 96%, providing a standard template input for high-resolution multi-condition adaptive baseline modeling;

[0194] The final output feature response pattern template can significantly improve the ability to distinguish subtle signal differences when the model is dynamically adjusted and abnormal trends are identified. It can effectively suppress false recognition of artifacts caused by environmental disturbances and support subsequent recursive least squares modeling and model sensitivity improvement.

[0195] By utilizing the characteristic output response patterns under various operating conditions, adaptive parameter estimation under multiple operating conditions is performed, and a set of adaptive baseline parameters under multiple operating conditions is output.

[0196] The set of multi-condition adaptive baseline parameters is input into the model optimization module to generate a multi-condition adaptive baseline model that can be automatically adjusted according to the operating conditions, which is used for subsequent dynamic comparison and discrimination of abnormal liquid level trends.

[0197] Furthermore, in this embodiment, step S5 specifically includes:

[0198] Using the standardized time-domain dielectric signal sequence as input, the standardized time-domain dielectric signal of each sampling point is sequentially input into the multi-condition adaptive baseline model to obtain the standardized baseline response signal corresponding to each sampling point.

[0199] Based on the standardized baseline response signal, the standardized time-domain dielectric signal sequence of each sampling point and its standardized baseline response signal are subjected to point-by-point dynamic offset calculation to generate the liquid level dynamic offset sequence.

[0200] Using the dynamic offset sequence as a factor, multi-scale spatiotemporal feature extraction is performed to refine the dynamic offset evolution characteristics of each sampling point at different time scales and regions, thereby obtaining the basic signal of local anomaly trend in the liquid level area.

[0201] Adaptive feature normalization processing is performed on the basic signal of local abnormal trend to output the local abnormal trend signal of liquid level normalized by the working condition.

[0202] The normalized local anomaly trend signal of liquid level is summarized into a spatially distributed anomaly trend map, which forms the basic input for dynamic imaging analysis. This provides a spatial-temporal feature dataset for identifying the amplitude, persistence and diffusion characteristics of anomaly trend fluctuations, and can be used to distinguish minute anomaly signals.

[0203] Furthermore, in this embodiment, step S6 specifically includes:

[0204] For local anomaly trend signals of liquid level, a multi-scale time-domain sliding window signal sequence is generated based on an adaptive window partitioning strategy to capture local liquid level fluctuation patterns at different time scales. This provides an input standard compatible with the local anomaly trend signal of liquid level in the previous step for time-domain multi-window analysis, and outputs a sliding window signal set.

[0205] For the sliding window signal set, a multi-window spatiotemporal distribution mapping map is established for the abnormal trend response data of different sampling points in each time domain window. This enables the correlation and integration of local abnormal trend signals of liquid level in the spatial distribution dimension, providing high-resolution spatiotemporal image data for subsequent feature extraction.

[0206] The generated multi-window spatiotemporal distribution map is subjected to weak signal enhancement and noise suppression processing. The amplitude, persistence, and spatial diffusion characteristic parameters of trend changes within each window are calculated for the noise-suppressed spatiotemporal distribution signal template. Through signal processing techniques such as high-pass / low-pass filtering and adaptive signal-to-noise ratio adjustment, the weak abnormal fluctuations of liquid level in the spatial and temporal dimensions are further amplified, effectively removing environmental noise artifacts related to the previous step, and realizing a more sensitive spatiotemporal trend anomaly template.

[0207] For the spatiotemporal distribution signal template after noise suppression, based on the joint feature extraction algorithm of time domain and spatial domain (such as duration statistics, spatial correlation coefficient, maximum amplitude, root mean square amplitude, spatial diffusion metric, etc.), the amplitude, persistence and spatial diffusion characteristic parameters of trend change within each window are calculated, realizing the technical derivation with the local abnormal trend signal of liquid level as the cause and the trend characteristic parameters as the effect.

[0208] For all calculated trend change magnitude, persistence, and spatial diffusion characteristic parameters, a globally standardized trend feature vector is generated based on multi-scale aggregation processing. This provides a unified scale of feature data input for the subsequent clustering and elimination algorithm based on the artifact event feature library, and ensures connection with the aforementioned causal chain technique for local abnormal trend signals of liquid level.

[0209] Furthermore, in this embodiment, step S7 specifically includes:

[0210] Based on the trend change magnitude, trend change persistence, and spatial diffusion characteristic parameters, feature vector assembly is performed to form a trend feature parameter set;

[0211] Based on the trend feature parameter set, retrieve and call the historical false alarm and artifact event feature library, perform similarity judgment between the input feature parameter set and the artifact event features in the library, and obtain preliminary false weak anomaly signal matching results;

[0212] The preliminary pseudo-weak anomaly signal matching results are divided into multidimensional clusters and classified into specific category groups based on the distribution relationship of the trend feature parameter set, thereby identifying potential periodic disturbances and external noise patterns.

[0213] Based on the clustering results, the statistical confidence of the category to which the trend feature parameter set belongs is evaluated. For the trend feature parameter set that is identified as a high-probability pseudo-weak anomaly signal, feature removal processing is performed, and the removed liquid level anomaly trend feature parameters are output as a high-confidence real weak anomaly signal data source.

[0214] Furthermore, in this embodiment, step S8 specifically includes:

[0215] For the abnormal liquid level trend signal after artifact removal, the time-frequency domain joint feature extraction method is used to calculate the upper and lower boundary features of liquid level, the abrupt change start feature and the slow drift trend feature parameter set to obtain a multi-target trend feature matrix, which serves as the input basis for subsequent AI abnormal trend detection algorithms.

[0216] By utilizing a multi-objective trend feature matrix and applying an AI-enhanced abnormal trend detection algorithm (including time-series clustering analysis and trend change detection network), multi-mode recognition of abnormal liquid level trends is performed, and the output includes abnormal trend category labels and preliminary abnormal intensity scores, thereby achieving accurate identification of weak abnormal signals of resin liquid level.

[0217] Based on the abnormal trend category labels and abnormal intensity scores, combined with the dynamic threshold curve under historical working conditions, the adaptive boundary update algorithm is used to revise the early warning response thresholds of various abnormal trends in real time, and generate a multi-level adaptive early warning threshold matrix for modes such as upper and lower boundary breakthroughs, mutation start points, and slow drift.

[0218] The adaptive early warning threshold matrix is ​​mapped and compared with the feature sequence of the current liquid level anomaly trend signal. The early warning sensitivity control algorithm determines whether to trigger the early warning signal and dynamically adjusts the false alarm suppression parameter so that the output anomaly trend judgment result has both high foresight and low false alarm rate.

[0219] For early warning signals triggered by the above process, the warning trigger time, corresponding multi-target trend feature parameters, threshold matrix and false alarm suppression parameters are recorded to generate data samples for subsequent data feedback and threshold model self-optimization, providing support for the model to continuously improve the sensitivity and robustness of weak signal abnormal trend detection.

[0220] Furthermore, in this embodiment, step S9 specifically includes:

[0221] The characteristic parameters of the abnormal liquid level trend signal after dynamic adaptive early warning threshold processing are determined, and the threshold cross relationship and critical point timestamp of the trend signal are calculated by using a logical discrimination algorithm to obtain the early anomaly judgment parameter set.

[0222] Based on the early anomaly judgment parameter set, multi-dimensional condition triggering rules are used to compare the trend signal characteristics with the historical graded early warning model to determine whether the current liquid level anomaly trend signal meets or exceeds the early anomaly early warning triggering threshold, and generate early warning judgment results.

[0223] Based on the generated early warning judgment results, the decision module performs process bifurcation control for the two situations of "meeting early anomaly" and "not meeting early anomaly", realizing a hierarchical response to the generation of anomaly early warning signals;

[0224] When the conditions for early abnormality warning are met, an early abnormality warning signal for resin level is automatically generated. This signal is then structured and encapsulated, and core data such as trend signal characteristic parameters and threshold cross timestamps are integrated to output a standardized warning information package.

[0225] The standardized early warning information package is pushed to the operation and maintenance server module via a remote data interface protocol to realize the information synchronization and closed-loop management of abnormal equipment status, and at the same time record the full process data of the early warning event for subsequent traceability.

[0226] Furthermore, in this embodiment, step S10 specifically includes:

[0227] The dielectric time-domain signals collected throughout the entire early warning response process are archived and processed. The original dielectric time-domain signals are labeled with key information such as sampling timestamp, sampling location, and radio frequency excitation parameters, and a structured dielectric time-domain signal archive is output for unified management and retrieval in subsequent database backflow operations.

[0228] Based on the archived dielectric time-domain signal files, combined with the abnormal trend parameters (including local abnormal amplitude, persistence parameters, spatial distribution parameters, etc.) calculated during the current warning period, the abnormal trend parameters are associated, indexed and dynamically bound to generate an abnormal trend parameter set with feature labels, providing targeted feature data for model training.

[0229] The adaptive warning threshold dynamically generated during the warning period is multidimensionally labeled with the corresponding abnormal trend parameters and sampling environment characteristics. Based on the multi-condition mapping rules, it is incorporated into the liquid level response database to realize the dynamic coupling between the warning threshold and the historical liquid level feature model, providing a spatiotemporal baseline reference for subsequent modeling and optimization.

[0230] For the dielectric time-domain signal archives, abnormal trend parameter sets, and multi-dimensional early warning thresholds of the returned warehouse, the multi-condition adaptive baseline model update algorithm is used to iteratively optimize the multi-condition adaptive baseline model by fusing historical and current data, so as to achieve self-calibration of model sensitivity under different resin types and environmental conditions.

[0231] By utilizing the abnormal trend parameter set from the backflow, the characteristic early warning threshold, and the model output results, adaptive optimization processing based on data performance is performed on the AI ​​anomaly detection parameters (such as model weights, feature selection gating, and threshold adjustment factors), updating the parameter configuration within the AI ​​multi-target anomaly trend detection algorithm, and improving the sensitivity of early anomaly trend identification in weak signals.

[0232] In the description of this invention, it should be understood that the orientation or positional relationship indicated by directional terms such as "front, back, up, down, left, right", "horizontal, vertical, horizontal" and "top, bottom" is generally based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing this invention and simplifying the description. Unless otherwise stated, these directional terms do not indicate or imply that the device or element referred to must have a specific orientation or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on the scope of protection of this invention.

[0233] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

Claims

1. A method for intelligent sensing and early warning of resin level in impregnation machine, characterized in that, The method comprises the following steps: S1, obtaining the radio frequency pulse response signals collected by the broadband dielectric sensing array at multiple sampling positions in the resin liquid level area of the impregnator equipment, and forming a time domain dielectric signal dataset with spatial distribution attributes; S2, performing noise suppression and signal normalization processing on the time domain dielectric signal dataset to obtain a standardized time domain dielectric signal sequence; S3, extracting the dielectric constant change characteristics of each sampling point in multiple time-frequency scales according to the standardized time domain dielectric signal sequence, and classifying the data samples under different working conditions; S4, constructing a multi-condition adaptive baseline model using a historical liquid level response database according to the classified dielectric constant change characteristics and their working condition labels; S5, inputting the standardized time domain dielectric signal sequence obtained by actual sampling into the multi-condition adaptive baseline model, calculating the dynamic offset between each sampling point liquid level and the baseline response, and obtaining a liquid level local anomaly trend signal; S6, performing multi-window spatio-temporal distribution dynamic imaging processing on the liquid level local anomaly trend signal, amplifying the detection sensitivity of weak abnormal fluctuations, and calculating the amplitude, persistence and spatial diffusion characteristic parameters of the trend change; S7, inputting the amplitude, persistence and spatial diffusion characteristic parameters of the trend change into a historical false alarm and artifact event feature library to identify and eliminate pseudo-weak abnormal signals; S8, dynamically generating an adaptive early warning threshold for the liquid level anomaly trend signal after removing the artifacts in combination with the composite mode of the upper and lower boundaries of the liquid level, the mutation starting point and the slow drift trend.

2. The resin level intelligent sensing and early warning method for impregnator according to claim 1, characterized in that, The step S8 further comprises: S9, judging whether the liquid level anomaly trend signal processed by the dynamically adaptive early warning threshold meets the early anomaly warning condition, and if so, outputting a resin liquid level early anomaly warning signal and starting information synchronization of the remote operation and maintenance server when the resin liquid level early anomaly warning signal is output.

3. The resin level intelligent sensing and early warning method for impregnator according to claim 2, characterized in that, The step S9 further comprises: S10: feeding the dielectric time domain signal, anomaly trend parameters and early warning threshold involved in the whole early warning process to the liquid level response database, continuously optimizing the multi-condition adaptive baseline model and AI anomaly detection parameters based on the data backflow mechanism, and improving the recognition sensitivity of weak signal early anomaly trend.

4. The resin level intelligent sensing and early warning method for impregnators according to claim 3, characterized in that, The step S1 specifically comprises: performing physical parameter analysis on the spatial distribution of the resin liquid level area in the impregnator equipment to determine the multi-sampling position deployment scheme of the broadband dielectric sensing array; based on the deployed multi-sampling positions, selecting a dielectric sensor array with wide frequency response characteristics to implement radio frequency pulse excitation on the resin liquid level area, and synchronously acquiring time domain radio frequency response signals of each sensor node; performing time and space synchronous time correction processing on the original radio frequency response signals obtained at each sampling position, and uniformly correcting the sampling time stamp using a clock synchronization algorithm; combined with the spatial coordinate information of each sampling position, the time domain dielectric response signals after synchronous time correction are collected in a multi-channel parallel format, forming a data frame with spatial distribution index; performing integrity and effectiveness detection on the first round of collected spatial distribution type time domain dielectric signal dataset, identifying the sampling blind area and abnormal noise sampling points using a signal quality evaluation module, and returning the deployment and collection parameter optimization suggestions.

5. The resin level intelligent sensing and early warning method for impregnator according to claim 3, characterized in that, The step S2 specifically comprises: Performing multi-channel adaptive filtering algorithm processing on the collected original time-domain dielectric signal data set, filtering out power frequency interference and high-frequency environmental noise in the wideband radio frequency sampling, and obtaining a noise suppression preliminary result; Taking the noise suppression preliminary result as input, performing signal baseline correction based on endpoint detection and waveform baseline drift correction algorithm, and obtaining a baseline-corrected time-domain dielectric signal result; Performing multi-scale amplitude statistical analysis on the baseline-corrected time-domain dielectric signal result, calculating mean, variance and extreme value parameters between different sampling points and multiple samplings of the same sampling point, and dynamically estimating an amplitude normalization factor; According to the amplitude normalization factor, performing amplitude normalization mapping processing on the baseline-corrected time-domain dielectric signal result, unifying the signal amplitudes of different sampling points to a standardized interval, and forming a standardized time-domain dielectric signal sequence preliminary output eliminating the influence of environmental amplitude fluctuation; Taking the standardized time-domain dielectric signal sequence preliminary output as input, detecting and correcting individual abnormal points or short-time missing points, and generating a final standardized time-domain dielectric signal sequence.

6. The resin level intelligent sensing and early warning method for impregnator according to claim 3, characterized in that, The step S3 specifically comprises: Performing multi-scale time-frequency decomposition processing on the standardized time-domain dielectric signal sequence, extracting local frequency domain response characteristics of each sampling point under different time resolutions by using wavelet transform algorithm, and obtaining a multi-scale dielectric spectrum parameter set; According to the multi-scale dielectric spectrum parameter set, applying extraction characteristic power spectrum density and principal component analysis algorithm to each sampling point to generate its main dielectric constant variation characteristic parameter; Aligning the dielectric constant variation characteristic parameter of each sampling point with the resin type, temperature and impurity working condition label stored by the device, and generating a feature fusion sample set associated with the working condition environment; Performing feature selection and normalization processing on the feature fusion sample set, retaining the core dielectric constant variation parameters sensitive to the working condition and having distinguishing degree, and performing feature dimension compression; According to the dimension-compressed core dielectric constant variation parameters and working condition labels, grouping the data samples of all sampling points, classifying the features according to the working condition categories with different resin types, temperature intervals and impurity concentrations, and obtaining a standard feature category set under multiple working conditions.

7. The resin level intelligent sensing and early warning method for impregnator according to claim 3, characterized in that, The step S4 specifically comprises: Performing standardized vector mapping processing on the dielectric constant variation characteristics and its working condition labels, and generating a dielectric constant multi-dimensional feature matrix under cross-working condition; According to the dielectric constant multi-dimensional feature matrix, calling multi-type working condition sample data in the historical liquid level response database, and obtaining a liquid level electromagnetic response historical sequence under the associated working condition through feature solidification and label indexing; Grouping the liquid level electromagnetic response historical sequence according to the resin type, temperature and impurity concentration parameters, and obtaining a feature output response mode under each working condition; Using the feature output response mode under each working condition, performing adaptive parameter estimation under multiple working condition conditions, and outputting a multi-working condition adaptive baseline parameter set; Taking the multi-working condition adaptive baseline parameter set as input, generating a multi-working condition adaptive baseline model that can be automatically adjusted according to the working condition.

8. The resin level intelligent sensing and early warning method for impregnator according to claim 3, characterized in that, The step S5 specifically comprises: The standardized time-domain dielectric signal sequence is taken as an input condition, and the standardized time-domain dielectric signals of each sampling point are sequentially input into the multi-condition adaptive baseline model to obtain the standardized baseline response signals corresponding to the sampling points; According to the standardized baseline response signals, the point-by-point dynamic offset calculation is performed on the standardized time-domain dielectric signal sequence and the standardized baseline response signals of each sampling point to generate a liquid level dynamic offset sequence; The multi-scale space-time feature extraction is performed by taking the dynamic offset sequence as an input, the dynamic offset evolution features of each sampling point at different time scales and regions are extracted, and a local abnormal trend basis signal of the liquid level region is obtained; The adaptive feature normalization processing is performed on the local abnormal trend basis signal to output a condition-normalized liquid level local abnormal trend signal; The condition-normalized liquid level local abnormal trend signal is summarized as a spatial distribution abnormal trend mapping to form a basic input for dynamic imaging analysis.

9. The resin level intelligent sensing and early warning method for impregnator according to claim 3, characterized in that, The step S6 specifically includes: For the liquid level local abnormal trend signal, a multi-scale time-domain sliding window signal sequence is generated based on an adaptive window division strategy, and a sliding window signal set is outputted; For the sliding window signal set, a multi-window space-time distribution mapping is established for the abnormal trend response data of different sampling points in each time-domain window; The generated multi-window space-time distribution mapping is subjected to weak signal enhancement and noise characteristic suppression processing, and the space-time distribution signal template after noise suppression is subjected to calculation of the amplitude, persistence and spatial diffusion characteristic parameters of the trend change in each window; The trend change amplitude, persistence and spatial diffusion characteristic parameters calculated are subjected to multi-scale aggregation processing to generate a global standardized trend feature vector.

10. The resin level intelligent sensing and early warning method for impregnators according to claim 3, characterized in that, The step S7 specifically includes: According to the trend change amplitude, trend change persistence and spatial diffusion characteristic parameters, the feature vector assembly processing is performed to form a trend feature parameter set; According to the trend feature parameter set, the historical false alarm and artifact event feature library is retrieved and called, the similarity between the input feature parameter set and the artifact event features in the library is judged, and a preliminary pseudo-weak abnormal signal matching result is obtained; The multi-dimensional clustering division is performed on the preliminary pseudo-weak abnormal signal matching result, the distribution relationship of the trend feature parameter set is attributed to a specific category group, and the potential periodic disturbance and external noise mode are identified; In combination with the clustering division result, the statistical reliability of the category to which the trend feature parameter set belongs is evaluated, the feature elimination processing is performed on the trend feature parameter set judged as a high-possibility pseudo-weak abnormal signal, and an eliminated liquid level abnormal trend feature parameter is outputted.

Images