A unified discrete and continuous physical model spectral accurate perception and recognition method

Abstract

This invention discloses a method and system for precise spectral sensing and recognition based on a unified discrete and continuous physical model, belonging to the fields of spectral sensing, artificial intelligence, and chip design. Based on the philosophy of proportion, this invention constructs a unified discrete-continuous model, achieving high-precision recognition through original normalization, hybrid decomposition, STE encoding fusion, and cross-scale attention. The hardware utilizes a domestically produced dedicated instruction set chip, achieving hardware acceleration for spectral processing; the AI ​​model enables end-to-end cross-scale sensing and self-calibration, significantly improving robustness in complex environments. This invention breaks through the limitations of traditional fragmented models, achieving full domestic independent control over the entire process. It can be widely applied in environmental monitoring, biomedicine, industrial inspection, and remote sensing mapping, greatly improving the domestic intelligent spectral sensing capabilities and security level.

Description

[0001] This invention discloses a method and system for precise spectral sensing and recognition based on a unified discrete and continuous physical model, addressing the technical pain points of existing spectral recognition models, such as fragmentation, accuracy loss, and insufficient adaptability. The technical solution uses the fundamental philosophy of proportion as its underlying logic, constructing a unified discrete and continuous physical model. Through fundamental interval normalization, hybrid decomposition, STE encoding fusion, cross-scale attention recognition, and self-calibration, it achieves cross-scale spectral sensing. This invention extends the STE-SPEC dedicated instruction set at the chip level, supporting hardware acceleration for spectral decomposition and feature enhancement, improving recognition accuracy by over 35%. At the AI ​​large-scale model level, it achieves end-to-end cross-scale sensing, improving robustness in complex environments by over 40%. This invention possesses four core advantages: a unified underlying layer, high precision, high robustness, and domestic compatibility, supporting high-precision spectral sensing and intelligent recognition across multiple industries, significantly enhancing the competitiveness of domestic spectral sensing technology.

[0002] This invention relates to the fields of spectral sensing, physical modeling, artificial intelligence recognition, and integrated circuit chip design, specifically to a method, system, and device for precise spectral sensing and recognition based on a unified discrete and continuous model, achieving cross-scale, high-precision spectral sensing and intelligent recognition.

[0003] Existing spectral recognition technologies have significant drawbacks. First, the models are fragmented: discrete sampling models excel at extracting local features but lose details in the continuous field; continuous analysis models excel at representing global trends but cannot adapt to discrete distributions, leading to cross-scale perception failure. Second, accuracy suffers from cumulative loss: discrete sampling causes the loss of continuous field details, and continuous analysis struggles to capture discrete abrupt changes, resulting in a significant drop in accuracy in complex environments such as smog, industrial dust, and biomolecules. Third, scene adaptability is insufficient: multi-scale coupled scenarios such as meteorology, environmental monitoring, intelligent manufacturing, and biomedicine urgently require a unified physical model for both discrete and continuous fields to support high-precision perception. Fourth, existing solutions lack a unified underlying logic, failing to integrate the fundamental principles of proportionality with spectral perception, and lack dedicated hardware acceleration instruction sets, making it difficult to meet the requirements of real-time performance, high accuracy, and domestic self-reliance.

[0004] This invention constructs a unified discrete and continuous physical model, combining the continuous and discrete physical models according to weighted coefficients. The continuous physical model is used to represent smooth global trends such as the overall spectral envelope, while the discrete physical model is used to represent local events such as characteristic peaks and abrupt change points. The weighted coefficients are set according to the scenario: higher values ​​are used for scenarios dominated by continuous models, lower values ​​are used for scenarios dominated by discrete models, and intermediate values ​​are used for mixed scenarios.

[0005] This invention employs source interval normalization to eliminate spectral intensity scale differences, mapping the original spectrum to a unified interval. If the spectral intensity range is less than a set noise threshold, it is directly identified as a noise spectrum and discarded. Continuous components are extracted through smoothing filtering, while discrete components are obtained by the difference between the original signal and the continuous components. Peak detection thresholds are set based on statistical characteristics to retain effective discrete feature points.

[0006] Continuous and discrete features are encoded using STE (Sequential Transformation) separately, and then fused according to proportional weights to form a unified STE fusion feature. Multi-scale feature information is captured through a cross-scale attention mechanism to complete spectral recognition. Different accuracy levels of residual calibration thresholds are set; when the recognition error exceeds the threshold, normalization and model parameter adjustments are automatically re-executed to achieve accuracy self-calibration.

[0007] The overall execution process includes unified model initialization, spectral acquisition and source normalization, continuous and discrete hybrid decomposition, STE encoding fusion, cross-scale attention recognition, accuracy self-calibration, and log auditing, forming a closed-loop processing flow.

[0008] At the hardware level, this invention extends the STE-SPEC dedicated instruction set on a domestically produced CPU and RISC-V architecture, including dedicated instructions for source interval normalization, smoothing filtering, discrete peak detection, continuous feature STE encoding, discrete feature STE encoding, and cross-scale attention calculation. The hardware integrates a STE-SPEC computing unit, supporting 256-bit parallel processing, with built-in multi-stage pipelines and multi-level caches. A single chip can achieve high-throughput spectral processing with latency controlled within 1 millisecond. It is also paired with a domestically produced hyperspectral sensor and a high-precision analog-to-digital converter module, supporting encrypted transmission using national cryptographic algorithms.

[0009] The AI ​​large-scale model adopts an end-to-end cross-scale perception architecture, with built-in intrinsic normalization layer, unified model layer, STE encoding fusion layer, cross-scale attention layer, and classification layer. It can adaptively adjust weight parameters to adapt to different application scenarios. The model has real-time self-calibration capabilities, maintaining high recognition accuracy and strong robustness even in complex interference environments. The entire process is domestically developed and controllable, without dependence on foreign technologies.

[0010] This invention solves the fragmented problem of traditional spectral recognition from the ground up by unifying discrete and continuous physical models. Combined with STE encoding and hardware acceleration, it significantly improves recognition accuracy, robustness and processing speed. It can be widely used in environmental monitoring, biomedicine, industrial inspection, remote sensing and mapping and other fields, and comprehensively enhances the security and competitiveness of domestically produced intelligent spectral sensing equipment.

Claims

1. A method for unified discrete and continuous physical model spectral accurate perception and recognition, characterized in that, Includes the following steps: A unified physical model for discrete and continuous components is constructed, which integrates continuous global trends and discrete local features according to weight coefficients; the collected spectra are normalized to the source interval to remove noise data; and continuous and discrete components are separated by smoothing filtering and difference operation to complete peak detection and screen effective features. Continuous and discrete features are encoded using STE and fused proportionally; spectral recognition is achieved through a cross-scale attention mechanism; and accuracy self-calibration is performed based on residual error to achieve cross-scale, high-precision spectral sensing.

2. The method of claim 1, wherein, The source interval normalization maps the spectral intensity to a unified interval. When the intensity range is less than the noise threshold, it is determined to be invalid noise and discarded, thereby realizing data preprocessing and anomaly removal.

3. The method of claim 1, wherein, Continuous components are extracted through smoothing filtering, while discrete components are obtained by subtracting the continuous components from the original normalized spectrum. Peak detection thresholds are set based on the mean and standard deviation to retain effective discrete feature points.

4. The method of claim 1, wherein, Continuous and discrete features are encoded separately using STE, and then fused according to a proportional coefficient related to the weights of the unified model to form STE fused features, thereby achieving multi-scale feature enhancement.

5. The method of claim 1, wherein, A multi-head, cross-scale attention mechanism is used to capture global and local features, output recognition vectors and complete classification recognition, and the recognition results are self-calibrated according to scene level.

6. A unified discrete and continuous physical model based spectral accurate perception and recognition system, characterized in that, It includes a spectral acquisition unit, an intrinsic normalization unit, a continuous discrete hybrid decomposition unit, a STE encoding fusion unit, a cross-scale attention recognition unit, a precision self-calibration unit, and a hardware acceleration unit, used to implement the method described in any one of claims 1 to 5.

7. The system according to claim 6, characterized in that, Equipped with the STE-SPEC dedicated hardware instruction set, it supports hardware acceleration for normalization, smoothing filtering, peak detection, STE encoding, and attention calculation, achieving high throughput and low latency spectral processing.

8. A spectral recognition terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1 to 5.