Quantitative and quantitative rapid detection method based on multi-mode AI material

Abstract

The invention discloses a multi-modal AI material quantitative and qualitative rapid detection method, which comprises the following steps: acquiring multi-modal data through a double-pulse laser excitation system, carrying out time-space alignment, and cooperatively exciting light and heavy element characteristic plasmas by using femtosecond laser and nanosecond laser; spectral decoupling and dynamic compensation are carried out based on physical rule embedding, spectral signals are separated, and interference is suppressed; a light-heavy element dual-channel deep learning model is constructed, and multi-modal data is fused, so that feature extraction and prediction capabilities are improved; and a double-source mutual verification optimization model is adopted, so that the result reliability is ensured. According to the method, the limitation of single-mode detection is overcome, multi-component high-precision real-time synchronous detection of coal is realized, the adaptability of the model to complex coal quality and environmental change is improved, and an effective and accurate solution is provided for online monitoring of coal quality of a coal-fired power plant.

Description

Technical Field

[0001] The present invention relates to the technical field of coal detection, and in particular to a rapid quantitative and qualitative detection method for materials based on multimodal AI. Background Art

[0002] Traditional coal detection technologies such as rapid gamma neutron activation analysis rely on gamma rays produced by the interaction of neutrons and atomic nuclei. Although it can perform full element analysis, it carries the risk of nuclear radiation and the equipment is bulky. Near-infrared spectroscopy is greatly affected by environmental interference, has a low signal-to-noise ratio, and requires complex preprocessing. X-ray fluorescence spectroscopy cannot detect key organic components due to the low X-ray yield of light elements. Dual-energy gamma ray transmission technology can only measure ash content and has the risk of nuclear contamination. Microwave technology is limited to moisture detection, and its accuracy is affected by the way the coal powder is stacked. Inductively coupled plasma atomic emission spectroscopy requires sample digestion and cannot isolate the water and oxygen environment, making it difficult to apply online. None of these technologies can simultaneously meet the needs of coal-fired power plants for real-time synchronous detection of multiple indicators.

[0003] Laser-induced breakdown spectroscopy, as a method of atomic emission spectroscopic analysis, uses a high-energy pulsed laser to ablate the sample surface to generate plasma, and then analyzes the characteristic spectra excited during the cooling process of the plasma to achieve qualitative and quantitative detection of elements. In the field of coal testing, LIBS has become a potential solution for online monitoring of coal quality in coal-fired power plants due to its in-situ real-time performance, multi-component simultaneous analysis capabilities, and minimally invasive detection characteristics. Although LIBS technology has significant advantages in coal testing, its practical application still faces the following drawbacks: 1. Coal composition is complex and its physical structure is variable, resulting in significant differences in the spectral signals of the same element in different coal samples. Metal oxides in the ash will suppress the intensity of carbon atomic spectral lines, while hydrocarbons in the volatile matter will affect the stability of CN molecular spectral lines. 2. Traditional methods rely on empirical screening of elemental characteristic spectral lines as input variables, which is a time-consuming and highly subjective process. Especially for mixed coal samples, fluctuations in coal quality require frequent reselection of spectral lines, resulting in poor model generalization.

[0004] Therefore, it is necessary to improve a multimodal AI-based rapid quantitative and qualitative material detection method in the existing technology to solve the above problems. Summary of the Invention

[0005] The present invention overcomes the shortcomings of the existing technology and provides a rapid quantitative and qualitative detection method for materials based on multimodal AI, aiming to solve the problems in the existing technology that it cannot dynamically compensate for coal quality fluctuations and is greatly affected by environmental interference.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a rapid quantitative and qualitative detection method for materials based on multimodal AI, comprising:

[0007] 2.S1, using a dual-pulse laser excitation system to synchronously acquire multimodal data and align the data in time and space;

[0008] S2, decouple the light-heavy element spectrum of the original spectrum based on physical rule embedding, and perform interference suppression and feature enhancement through dynamic compensation algorithm;

[0009] S3. Build a light-heavy element dual-channel deep learning model and build a fusion model by combining morphological features and environmental parameters;

[0010] S4, using dual-source mutual verification for model adaptive optimization;

[0011] S5. Output the corresponding coal quantitative and qualitative results.

[0012] In a preferred embodiment of the present invention, in step S1, the dual-pulse laser excitation system includes a femtosecond laser and a nanosecond laser;

[0013] Femtosecond laser parameters: wavelength 1020–1040 nm, pulse width 100–500 fs, energy 1–5 mJ, repetition rate 10–100 Hz, focused spot 40–50 μm;

[0014] Nanosecond laser parameters: wavelength 1060–1070 nm, pulse width 5–10 ns, energy 50–150 mJ, repetition rate 10–100 Hz, focused spot 90–100 μm;

[0015] The femtosecond laser was used as a pilot pulse for the first excitation, and the nanosecond laser was used for the second excitation after a delay of 1–5 μs.

[0016] In a preferred embodiment of the present invention, in step S1, the multimodal data includes:

[0017] The multimodal data includes:

[0018] Spectral signals: collected by a dual-channel spectrometer, with a light element band of 200-500nm and a heavy element band of 500-900nm;

[0019] Morphological parameters: 3D point cloud of coal powder was collected by a high-speed visible light camera at ≥1000 fps;

[0020] Thermodynamic parameters: The plasma temperature gradient is monitored by an infrared thermal imager with a sampling frequency of ≥1 kHz, with a detection range of 400-2000°C;

[0021] Environmental parameters: collected through temperature and humidity sensors and dust sensors, with sampling frequencies of 1-10Hz and ≥50Hz respectively.

[0022] In a preferred embodiment of the present invention, in step S1, the spatiotemporal alignment is achieved by:

[0023] The laser pulse signal is the main trigger source, and the FPGA controller sends synchronization signals to all sensors with a time jitter of <10ns.

[0024] Using the timing compensation formula t all =t tri +Δt dev , where t all is the unified timestamp after alignment, t tri is the laser pulse triggering moment, Δt dev is the inherent delay compensation value of each sensor, which is determined by pre-selection in experiments;

[0025] Calculate the actual time deviation δt=argmax∫S based on the cross-correlation function spe (t)·I cam (t+τ)dt, where S spe is the spectral intensity curve, I cam is the camera brightness curve.

[0026] In a preferred embodiment of the present invention, the spectral decoupling of step S2 includes: separating the femtosecond spectrum f(λ)=BPF(S(λ), 200nm, 500nm) and the nanosecond spectrum g(λ)=S(λ)-f(λ)-WD(S(λ)), where BPF is bandpass filtering and WD is background noise elimination; when the ash characteristic peak is detected, the ash interference index is calculated. Among them, I is the spectral intensity of the characteristic peak of ash content, I back is the current spectral line intensity, A pla is the plasma area.

[0027] In a preferred embodiment of the present invention, the dynamic compensation in step S2 includes:

[0028] Morphology Compensation S corr =g(λ)×(1-αΔρ), where Δρ is the deviation between the current bulk density and the calibration reference, and α is the attenuation coefficient;

[0029] Thermodynamic correction R adj =R means *e -β▽T , R means is the measured ion and atomic line intensity ratio, β is the temperature sensitivity coefficient, and ▽T comes from the plasma temperature field monitored by the infrared thermal imager.

[0030] In a preferred embodiment of the present invention, step S3 includes:

[0031] Light element channel: 1D-CNN with 5 convolution layers and 2 fully connected layers is used to process femtosecond spectra and output C, H, O, N content and volatile matter;

[0032] Heavy element channel: 1D-CNN with residual connection is used to process nanosecond spectra and output Fe, Ca, and Al feature vectors;

[0033] The fusion model integrates light element features, heavy element features, morphological parameters, and thermodynamic parameters through Transformer. The attention mechanism formula is: Among them, Q is the characteristic of light elements, K, V=[F h ; Mor; ▽T], Mor is the morphological feature, αΔρ is the physical rule embedding term, and the packing density deviation compensation.

[0034] In a preferred embodiment of the present invention, in step S4, the dual-source mutual verification mechanism includes: comparing the carbon content predicted by the femtosecond channel with the carbon content inferred by the nanosecond channel. If the relative error is greater than 5%, three re-inspections of the same point are triggered, and the average value is taken as the final result.

[0035] In a preferred embodiment of the present invention, when the model confidence is <90%, the femtosecond energy ±0.5mJ or the nanosecond delay ±0.5μs is automatically adjusted, and abnormal data supplementary sampling is triggered; a standard coal sample is inserted into each batch of detection, and the model bias term is dynamically updated through moving window weighted regression.

[0036] In a preferred embodiment of the present invention, the output result of step S5 includes: the content of organic components C, H, O, and N, volatile matter, percentage of ash elements Fe, Ca, and Al, coal type classification, and calorific value range, and is transmitted to the control system via the industrial bus.

[0037] The present invention solves the defects existing in the background technology and has the following beneficial effects:

[0038] (1) The present invention proposes a rapid quantitative and qualitative detection method for materials based on multimodal AI. By adopting a dual-pulse laser excitation system to synchronously collect multimodal data and perform spatiotemporal alignment, the synergistic effect of femtosecond laser and nanosecond laser is used to excite the characteristic plasma of light and heavy elements respectively, providing a data basis for subsequent accurate detection. Then, based on physical rule embedding, the original spectrum is decoupled from the light and heavy element spectrum to separate the spectral signals of different elements. At the same time, the dynamic compensation algorithm is used to suppress the interference caused by factors such as morphology and thermodynamics and enhance the characteristics, solving the problem of spectral signals being interfered with by coal quality and environment in traditional methods. Then, a light-heavy element dual-channel deep learning model is constructed, and a fusion model is constructed by combining morphological characteristics and environmental parameters to give full play to the detection advantages of each channel for specific elements, and to achieve deep fusion of multimodal data, thereby improving feature extraction and prediction capabilities. Finally, dual-source mutual verification is used to perform model adaptive optimization to ensure the reliability and stability of the detection results. Compared with the existing technology, the present invention overcomes the limitations of single-modal detection, realizes high-precision, real-time synchronous detection of multiple components of coal, and improves the adaptability of the model to complex coal quality and environmental changes.

[0039] (2) The present invention uses a dual-pulse laser excitation system, combining femtosecond laser and nanosecond laser with the synchronous acquisition and spatiotemporal alignment of multimodal data. The femtosecond laser first excites to produce a low-temperature, high-density plasma to avoid the volatilization of light elements and retain light elements; the nanosecond laser then excites to enhance the ionization of heavy metal elements and increase the intensity of heavy element spectral lines. At the same time, the laser pulse signal is used as the main trigger source, and the FPGA controller is used to send synchronization signals to all sensors. The timing compensation formula and cross-correlation function are used to calculate the actual time deviation to achieve spatiotemporal alignment of multimodal data. This combination effect enables the system to obtain plasma excitation signals with different characteristics of light and heavy elements respectively, and ensures that the collected multimodal data accurately corresponds to the state at the same physical moment, thereby achieving comprehensive and accurate information acquisition of coal samples. Compared with the existing technology, it further achieves the effect of overcoming the problem that single laser excitation is difficult to simultaneously take into account the detection accuracy of light and heavy elements and the problem of data asynchrony.

[0040] (3) The present invention uses a spectral decoupling and dynamic compensation algorithm embedded in physical rules, including morphological compensation and thermodynamic correction combined with a dual-channel deep learning model. First, the femtosecond spectrum and nanosecond spectrum are separated, and the light and heavy element spectra are decoupled by bandpass filtering and background noise elimination. When the ash characteristic peak is detected, the ash interference index is calculated. Then, morphological compensation and thermodynamic correction are performed based on physical rules, that is, the optical attenuation is corrected according to the bulk density deviation, and the element ionization degree offset is corrected according to the plasma temperature field. The decoupled and compensated spectral data is then input into the dual-channel deep learning model for processing. This combined effect achieves the extraction of pure and enhanced characteristic signals from the original spectrum and converts them into inputs that can be effectively learned by the deep learning model. Compared with the existing technology, it further achieves the effect of improving the model's detection accuracy and generalization ability for coal elements, and solves the problems of time-consuming, subjective and poor model generalization ability caused by the traditional method relying on experience to screen characteristic spectral lines.

[0041] (4) The present invention combines a light-heavy element dual-channel deep learning model (1D-CNN architecture) with a Transformer fusion model. The light element channel uses a 5-layer convolution + 2-layer fully connected 1D-CNN to process femtosecond spectra and output C, H, O, N content and volatile matter; the heavy element channel uses a 1D-CNN with residual connection to process nanosecond spectra and output Fe, Ca, Al feature vectors. Afterwards, the Transformer fusion model fuses light element features, heavy element features, morphological parameters and thermodynamic parameters, and fully explores the correlation and importance between each modal data through the attention mechanism. This combination effect realizes the targeted processing and deep fusion of different modal data. Compared with the existing technology, it further achieves the effect of improving feature extraction capabilities and prediction accuracy, and can more comprehensively and accurately perform quantitative and qualitative analysis of coal materials.

[0042] (5) The present invention combines a dual-source mutual verification mechanism with a model adaptive optimization step. When the relative error between the carbon content predicted by the femtosecond channel and the carbon content inferred by the nanosecond channel is greater than 5%, three re-inspections are triggered at the same point, and the average value is taken as the final result. At the same time, when the model confidence is lower than 90%, the femtosecond energy or nanosecond delay is automatically adjusted, and abnormal data re-collection is triggered. Standard coal samples are also inserted into each batch of detection, and the model bias term is dynamically updated through moving window weighted regression. This combined effect realizes real-time verification and optimization of the model prediction results. Compared with the existing technology, it further achieves the effect of improving the reliability, stability and model adaptability of the detection results, and effectively solves the problems of uncertainty in single-mode prediction and poor applicability of the model under different coal quality conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments described in the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive efforts.

[0044] Figure 1 is a flow chart of a preferred embodiment of the present invention;

[0045] Figure 2 is a fusion model diagram of a preferred embodiment of the present invention;

[0046] Figure 3 This is a multi-spectral synchronous acquisition diagram of a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0047] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0048] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0049] Application Overview:

[0050] Traditional single-modal detection technology is limited by fundamental bottlenecks, making it difficult to overcome the contradiction between accuracy and adaptability in the simultaneous detection of multiple components in coal. Taking laser-induced breakdown spectroscopy (LIBS) as an example: On the one hand, the spectral responses of ash (such as Fe and Ca oxides) and organic matter (hydrocarbons) in coal are nonlinearly coupled. The enhanced spectral lines of metal elements suppress the intensity of carbon atomic spectral lines, while the spectral lines of CN molecules released by volatiles are affected by combustion kinetic parameters, making it difficult to establish a stable quantitative model using a single spectral modality. On the other hand, physical morphological parameters such as coal bulk density and particle size distribution can change the plasma excitation efficiency, making traditional methods unable to dynamically compensate for matrix effects caused by coal quality fluctuations. In addition, external interference such as ambient temperature, humidity, and dust concentration can further exacerbate signal drift by changing the plasma temperature field distribution and optical transmission characteristics.

[0051] The present invention proposes a method for quantitative and qualitative material detection based on multimodal AI. It breaks through the detection limit of a single modality through dual-pulse laser excitation and multi-source data fusion, develops an interference suppression algorithm embedded with physical rules, and realizes the deep integration of physical mechanism and data-driven through light / heavy metal spectrum decoupling, element dynamic compensation and plasma temperature gradient correction.

[0052] Exemplary methods:

[0053] like Figure 1 As shown, a rapid quantitative and qualitative material detection method based on multimodal AI includes:

[0054] S1, using a dual-pulse laser excitation system to synchronously acquire multimodal data and align the data in time and space;

[0055] S2, decouple the light-heavy element spectrum of the original spectrum based on physical rule embedding, and perform interference suppression and feature enhancement through dynamic compensation algorithm;

[0056] S3. Build a light-heavy element dual-channel deep learning model and build a fusion model by combining morphological features and environmental parameters;

[0057] S4, using dual-source mutual verification for model adaptive optimization;

[0058] S5. Output the corresponding coal quantitative and qualitative results.

[0059] like Figure 3 As shown, in step S1, the dual-pulse laser system consists of a femtosecond laser and a nanosecond laser, which act on the same sample point in sequence through precise timing control to excite plasmas with different characteristics respectively;

[0060] Femtosecond lasers ablate the sample surface at low temperatures to prevent the volatilization of light elements, and the excited plasma is low-temperature, high-density plasma. Nanosecond lasers excite deep through thermal ablation, enhancing the ionization of heavy metal elements, and the excited plasma is high-temperature plasma.

[0061] Femtosecond laser is used as the pilot pulse for excitation, optimizing the breakage of CH and NO molecular bonds, improving the accuracy of volatile detection, and retaining light elements, including C, H, O, and N. The wavelength is 1020-1040nm, the pulse width is 100-500fs, the energy is 1-5mJ, and the low energy prevents sample splashing. The repetition rate is 10-100Hz, and the focused spot is 40-50μm for precise ablation. Nanosecond pulse delayed excitation, with a delay time of 1-5μs, the nanosecond laser is the main pulse, with a wavelength of 1060-1070nm, a pulse width of 5-10ns, and an energy of 50-150mJ. It has high energy deep excitation, a repetition rate of 10-100Hz, and a focused spot of 90-100μm to expand the excitation area. The nanosecond laser uses the pilot plasma to lower the breakdown threshold and enhance the spectral line intensity of heavy elements, including Fe, Ca, and Al. In addition, the dual-pulse timing can evaporate surface moisture and reduce spectral interference.

[0062] The reason for the femtosecond laser to proceed first is to avoid the aerosol generated by the nanosecond laser from shielding the subsequent ablation, and the delay time of 1-5μs is to wait for the initial expansion of the femtosecond plasma and the secondary excitation of the nanosecond laser.

[0063] In step S1, the multimodal data includes: spectral signals, morphological parameters, thermodynamic parameters and environmental parameters;

[0064] A dual-channel spectrometer is used to collect spectral signals for quantitative analysis of light and heavy elements, with light elements ranging from 200-500nm and heavy elements ranging from 500-900nm. A high-speed visible light camera with a speed of more than 1000fps is used to collect the morphological parameters of coal powder to form a 3D point cloud to correct for optical attenuation caused by bulk density. An infrared thermal imager is used to collect thermodynamic parameters, monitor the plasma temperature gradient, and correct the element ionization offset. The detection temperature range is 400-2000℃, and the sampling frequency is greater than 1kHz. Environmental parameters are obtained through temperature and humidity sensors and dust sensors, with sampling frequencies of 1-10Hz and above 50Hz, respectively.

[0065] In step S1, the data is aligned in time and space. The laser pulse signal is used as the main trigger source, and a synchronization signal is sent to all sensors through the FPGA controller with a time jitter of <10ns to ensure that the spectrometer, camera, and thermal imager start collecting data at the same physical moment. The sampling frequency and trigger delay of different sensors vary. If they are not aligned, the data collected at the same time will correspond to different physical states.

[0066] Timing compensation is t all =t tri +Δt dev , where t all is the unified timestamp after alignment, t tri is the laser pulse triggering moment, Δt devis the inherent delay compensation value of each sensor, which is determined by pre-selection in experiments;

[0067] Calculate the actual time deviation δt=argmax∫S based on the cross-correlation function spe (t)·I cam (t+τ)dt, where S spe is the spectral intensity curve, I cam is the camera brightness curve.

[0068] In step S1, light element spectra, heavy element spectra, coal powder morphology point cloud, plasma temperature field and environmental parameters are synchronously collected through time-sharing excitation of femtosecond laser and nanosecond laser, and μs-level spatiotemporal alignment is achieved based on hardware triggering and software cross-correlation calibration, providing a multimodal correlation data foundation for interference suppression embedded in physical rules.

[0069] The multimodal data after S1 alignment still have core interferences such as ash suppressing carbon spectral lines and volatiles interfering with molecular spectral lines, which require feature decoupling and enhancement through physical rule embedding; metal elements such as Fe and Ca in the ash will absorb C247.8nm spectral line photons, resulting in low carbon detection values, and the volatiles of coal powder will break under high-temperature excitation, causing fluctuations in the N and O element spectral lines.

[0070] In step S2, decoupling is performed to separate the spectral signals of light elements (C / H / O / N) and heavy elements (Fe / Ca / Al) to avoid mutual interference.

[0071] The original spectrum S(λ) = α·f(λ) + β·g(λ) + γ·N(λ), where f(λ) and g(λ) are the light element spectra excited by femtosecond laser and the heavy element spectra excited by nanosecond laser; α and β are channel weight coefficients determined by the laser energy, and γ·N(λ) is the noise term including environmental interference.

[0072] The decoupling method is:

[0073] First, perform channel separation: femtosecond spectrum extraction: f(λ) = BPF(S(λ), 200 nm, 500 nm), and nanosecond spectrum extraction: g(λ) = S(λ) - f(λ) - WD(S(λ)). BPF is a 200–500 nm bandpass filter that matches the femtosecond channel, and WD is used to eliminate background noise in the nanosecond spectrum.

[0074] When the ash characteristic peak is detected, the ash interference index is calculated Among them, I is the spectral intensity of the characteristic peak of ash content, I back is the current spectral line intensity, A pla is the plasma area.

[0075] When the heavy element signal exceeds the threshold, the light element spectral line is dynamically compensated;

[0076] Morphology compensation corrects optical attenuation:

[0077] Based on physical rules: the increase in stacking density leads to a decrease in laser penetration depth and a decay in spectral line intensity.

[0078] The compensation formula is S corr =g(λ)×(1-αΔρ), where Δρ is the deviation between the current bulk density and the calibration reference, and α is the attenuation coefficient.

[0079] Thermodynamic correction suppresses ionization shifts:

[0080] Based on physical rules: the increase of plasma temperature gradient ▽T leads to an increase in element ionization degree and an imbalance in the ratio of ion spectrum and atomic spectrum.

[0081] The correction formula is R adj =R means *e -β▽T , R means is the measured ion and atomic line intensity ratio, β is the temperature sensitivity coefficient, and ▽T comes from the plasma temperature field monitored by the infrared thermal imager.

[0082] Physical rule embedding converts domain knowledge into computable mathematical constraints through light-heavy element decoupling and morphological-thermodynamic dynamic correction, fundamentally solving the interference problem in coal LIBS detection and providing physically interpretable input features for multimodal AI models.

[0083] like Figure 2 As shown, in step S3, the dual-channel deep learning architecture:

[0084] The light element channel is used to predict organic components. The input is the femtosecond spectrum after S2 decoupling compensation. A 1D-CNN model is used with 5 convolution layers and 22 full connections. The local features of the light element atomic spectrum are extracted. The activation function is LeakyReLU to alleviate the gradient disappearance. The output is the C, H, O and N content and volatile matter prediction.

[0085] The heavy element channel is used for ash content prediction. The input is the thermodynamically corrected nanosecond spectrum. The homogeneous 1D-CNN is used in combination with residual connections to output the heavy element feature vector F. h =ReLU(Conv1D(g(λ))+Conv1D(g(λ))), where Conv1D is a one-dimensional convolution operation and ReLU is a rectified linear unit activation function. The residual design avoids gradient vanishing and improves the deep feature extraction capability of Fe, Ca, and Al.

[0086] Perform cross-modal fusion of feature interaction layers;

[0087] Both light element features and heavy element features are 128-dimensional CNN output vectors, morphological parameters are 32-dimensional 3D point cloud density, and thermodynamic parameters are 16-dimensional temperature gradient ▽T;

[0088] Transformer fusion is: Among them, Q is the characteristic of light elements, K, V=[F h ; Mor; ▽T], Mor is the morphological feature, αΔρ is the physical rule embedding term, and the packing density deviation compensation.

[0089] In step S4, to resolve the uncertainty of single-mode prediction and improve the robustness of carbon element detection, dual-source mutual verification is performed:

[0090] The predicted value of the femtosecond channel is is the carbon content output by the light element CNN;

[0091] The nanosecond channel back-calculation value is Infer the carbon content from the ash content;

[0092] The trigger condition is Automatically trigger the re-inspection of the same point, perform laser re-sampling and model re-prediction three times, and then take the average value.

[0093] And dynamically expand the model capabilities to new coal types through thermal maps; thermal map generation Among them, L is the model loss function, and x, y are the feature space coordinates.

[0094] Step S4 focuses on addressing the uncertainty of single-modal predictions to improve the robustness of carbon detection. This primarily utilizes a dual-source mutual verification approach, comparing the femtosecond channel predictions with the nanosecond channel inferences. When the relative error between the two exceeds 5%, a retest of the same point is automatically triggered. Simultaneously, the model's capabilities are dynamically expanded to new coal types through a heatmap generated based on the distribution of the loss function across feature space coordinates.

[0095] Step S5 enters the final output link of coal quantitative and qualitative results. When the model output confidence is less than 90%, the system automatically adjusts the laser parameters, femtosecond energy ±0.5mJ or nanosecond delay ±0.5μs, and triggers supplementary sampling of abnormal data at the same point until the confidence threshold is met;

[0096] Standard coal samples are inserted into each batch of testing, and the model bias term is dynamically updated through moving window weighted regression.

[0097] The final result output integrates the multimodal model prediction value and dynamic calibration parameters to generate a quantitative detection report, including: the content of organic components C, H, O, N, volatile matter, ash elements Fe, Ca, Al percentage, as well as coal classification and calorific value range, and is transmitted to the control system in real time via the industrial bus.

[0098] For the above embodiment, experiments were conducted using femtosecond lasers and nanosecond lasers with different wavelengths and energies, while keeping other variables the same, to verify that changes in wavelength and energy affect plasma generation and spectral emission, thereby affecting the accuracy of coal detection.

[0099] Table 1 Detection results of femtosecond laser and nanosecond laser with different wavelengths and energies

[0100]

[0101]

[0102] In laser-induced breakdown spectroscopy (LIBS), the wavelength and energy of femtosecond and nanosecond lasers have a crucial impact on detection accuracy. Femtosecond lasers, with their ultrashort pulse width and high peak power, enable precise excitation of micro-regions in a material, facilitating the spectral decoupling and detection of light elements (such as carbon). As the energy of the femtosecond laser increases, more energy is coupled into the material, stimulating stronger spectral signals of light elements and improving the detection accuracy of carbon content. However, the effect of femtosecond laser wavelength on carbon content detection accuracy is relatively small, as femtosecond lasers of different wavelengths can produce similar spectral responses when exciting light elements. In contrast, nanosecond lasers are primarily used to excite the spectra of heavy elements (such as ash components). Increasing nanosecond laser energy helps improve the detection accuracy of ash components, as higher energy more effectively excites the spectral signals of heavy elements. However, excessive nanosecond laser energy can lead to spectral saturation or interference, thereby reducing detection accuracy.

[0103] Overall detection confidence is an important indicator for comprehensively evaluating the accuracy of carbon content detection and ash element detection. It reflects the detection system's ability to accurately and reliably analyze material composition under different laser wavelength and energy combinations. As can be seen from the tabular data, when the wavelength and energy combination of the femtosecond laser and nanosecond laser are appropriate, the detection accuracy of both carbon content and ash element detection reaches a high level, thereby improving the overall detection confidence. This is because appropriate laser parameters can simultaneously optimize the excitation conditions for light and heavy elements, improving the quality and stability of the spectral signal. Conversely, when a femtosecond laser or nanosecond laser is lacking, the spectral signals of light or heavy elements cannot be effectively excited, resulting in a significant decrease in the corresponding detection accuracy, and thus a significant drop in the overall detection confidence.

[0104] The above description is based on the ideal embodiment of the present invention. Based on the above description, relevant personnel can make various changes and modifications without departing from the technical scope of this invention. The technical scope of this invention is not limited to the content of the specification and must be determined according to the scope of the claims.

Claims

1. A rapid quantitative and qualitative detection method for materials based on multimodal AI, characterized in that: include: S1, using a dual-pulse laser excitation system to synchronously acquire multimodal data and align the data in time and space; S2, decouple the light-heavy element spectrum of the original spectrum based on physical rule embedding, and perform interference suppression and feature enhancement through dynamic compensation algorithm; S3. Build a light-heavy element dual-channel deep learning model and build a fusion model by combining morphological features and environmental parameters; S4, using dual-source mutual verification for model adaptive optimization; S5. Output the corresponding coal quantitative and qualitative results.

2. The method for rapid quantitative and qualitative material detection based on multimodal AI according to claim 1, characterized in that: In step S1, the dual-pulse laser excitation system includes a femtosecond laser and a nanosecond laser; Femtosecond laser parameters: wavelength 1020–1040 nm, pulse width 100–500 fs, energy 1–5 mJ, repetition rate 10–100 Hz, focused spot 40–50 μm; Nanosecond laser parameters: wavelength 1060–1070 nm, pulse width 5–10 ns, energy 50–150 mJ, repetition rate 10–100 Hz, focused spot 90–100 μm; The femtosecond laser was used as a pilot pulse for the first excitation, and the nanosecond laser was used for the second excitation after a delay of 1–5 μs.

3. The method for rapid quantitative and qualitative material detection based on multimodal AI according to claim 1, characterized in that: In step S1, the multimodal data includes: The multimodal data includes: Spectral signals: collected by a dual-channel spectrometer, with a light element band of 200-500nm and a heavy element band of 500-900nm; Morphological parameters: 3D point cloud of coal powder was collected by a high-speed visible light camera at ≥1000 fps; Thermodynamic parameters: The plasma temperature gradient is monitored by an infrared thermal imager with a sampling frequency of ≥1 kHz, with a detection range of 400-2000°C; Environmental parameters: collected through temperature and humidity sensors and dust sensors, with sampling frequencies of 1-10Hz and ≥50Hz respectively.

4. The method for rapid quantitative and qualitative material detection based on multimodal AI according to claim 1, characterized in that: In step S1, the spatiotemporal alignment is achieved by: The laser pulse signal is the main trigger source, and the FPGA controller sends synchronization signals to all sensors with a time jitter of <10ns. Using the timing compensation formula t all =t tri +Δt dev , where t all is the unified timestamp after alignment, t tri is the laser pulse triggering moment, Δt dev is the inherent delay compensation value of each sensor, which is determined by pre-selection in experiments; Calculate the actual time deviation δt=argmax∫S based on the cross-correlation function spe (t)·I cam (t+τ)dt, where S spe is the spectral intensity curve, I cam is the camera brightness curve.

5. The method for rapid quantitative and qualitative material detection based on multimodal AI according to claim 1, characterized in that: The spectrum decoupling in step S2 includes: separating the femtosecond spectrum f(λ)=BPF(S(λ), 200nm, 500nm) and the nanosecond spectrum g(λ)=S(λ)-f(λ)-WD(S(λ)), where BPF is bandpass filtering and WD is background noise elimination; when the ash characteristic peak is detected, the ash interference index is calculated. Among them, I is the spectral intensity of the characteristic peak of ash content, I back is the current spectral line intensity, A pla is the plasma area.

6. The method for rapid quantitative and qualitative material detection based on multimodal AI according to claim 1, characterized in that: The dynamic compensation in step S2 includes: Morphology Compensation S corr =g(λ)×(1-αΔρ), where Δρ is the deviation between the current bulk density and the calibration reference, and α is the attenuation coefficient; Thermodynamic correction R adj =R means *e -β▽T , R means is the measured ion and atomic line intensity ratio, β is the temperature sensitivity coefficient, and ▽T comes from the plasma temperature field monitored by the infrared thermal imager.

7. The method for rapid quantitative and qualitative material detection based on multimodal AI according to claim 1, characterized in that: The step S3 comprises: Light element channel: 1D-CNN with 5 convolution layers and 2 fully connected layers is used to process femtosecond spectra and output C, H, O, N content and volatile matter; Heavy element channel: 1D-CNN with residual connection is used to process nanosecond spectra and output Fe, Ca, and Al feature vectors; The fusion model integrates light element features, heavy element features, morphological parameters, and thermodynamic parameters through Transformer. The attention mechanism formula is: Among them, Q is the characteristic of light elements, K, V=[F h ; Mor; ▽T], Mor is the morphological feature, αΔρ is the physical rule embedding term, and the packing density deviation compensation.

8. The method for rapid quantitative and qualitative material detection based on multimodal AI according to claim 1, characterized in that: In step S4, the dual-source mutual verification mechanism includes: comparing the carbon content predicted by the femtosecond channel with the carbon content inferred by the nanosecond channel. If the relative error is greater than 5%, three re-inspections are triggered at the same point, and the average value is taken as the final result.

9. The method for rapid quantitative and qualitative material detection based on multimodal AI according to claim 1, characterized in that: When the model confidence is <90%, the femtosecond energy is automatically adjusted by ±0.5mJ or the nanosecond delay is ±0.5μs, and abnormal data re-sampling is triggered; standard coal samples are inserted into each batch of detection, and the model bias term is dynamically updated through moving window weighted regression.

10. The method for rapid quantitative and qualitative material detection based on multimodal AI according to claim 1, characterized in that: The output results of step S5 include: organic components C, H, O, N content, volatile matter, ash elements Fe, Ca, Al percentage, coal type classification and calorific value range, and are transmitted to the control system via the industrial bus.

Images