Large model feature classification algorithm and system for abnormal detection in traditional chinese medicine processing

Abstract

This invention relates to the field of traditional Chinese medicine (TCM) manufacturing technology, and discloses a large-scale feature classification algorithm and system for detecting anomalies in the TCM processing process. The algorithm includes: acquiring a multimodal real-time dataset, parsing it in parallel to obtain information emergence points, and reconstructing them into a feature respiration spectrum; using the feature respiration spectrum to construct a process semantic graph and process manifold, performing vector decomposition to obtain the semantic orthogonal decomposition distance between representative traits and progress deviations; constructing a virtual correction force field to obtain active intervention instructions, driving risk inference to generate a structured anomaly diagnosis report. This invention achieves accurate capture and digital reconstruction of the processing sensitivity window, transforming qualitative anomaly judgment into quantitative diagnosis, solving the problems of difficult identification of latent anomalies and lack of correction suggestions, and improving the inherent safety, batch quality uniformity, and reliability of quality traceability in the TCM processing process.

Description

Technical Field

[0001] This invention relates to the field of traditional Chinese medicine manufacturing technology, and more specifically, to a large-scale feature classification algorithm and system for detecting abnormalities in the processing of traditional Chinese medicine. Background Technology

[0002] With the expansion of modern manufacturing of traditional Chinese medicine (TCM), consistent quality control throughout the entire processing process has become crucial for ensuring medication safety. However, existing processing monitoring technologies primarily rely on threshold alarms for macroscopic physicochemical parameters such as temperature, humidity, and stirring frequency. These technologies generally face the bottleneck of inconsistent quality despite conforming to standard operating procedures; that is, even when process parameters meet standard operating procedures, significant differences still exist in the content of effective components or the rate of toxic degradation in the finished product. Existing technologies often focus on static monitoring of single-modal signals, neglecting the semantic aliasing effect caused by the random emergence of microscopic states during processing. This leads to a high degree of convergence between abnormal evolution trajectories and standard trajectories at specific sampling moments, making it difficult to uniquely determine the true state and evolution trend of the medicinal material through instantaneous features. Furthermore, the processing process involves nonlinear and strongly coupled dynamic changes. Traditional methods lack in-depth modeling of the temporal coupling relationships between multidimensional features such as sound, light, electricity, and heat, failing to effectively identify latent qualitative changes caused by moisture obstruction or physical noise interference. Moreover, the output alarm information lacks interpretability regarding pharmacological mechanisms. Therefore, how to shift from static single-parameter monitoring to dynamic multimodal semantic alignment, and transform macroscopic qualitative judgment into precise quantitative diagnosis of the processing evolution trajectory in order to eliminate semantic aliasing of the process trajectory, is a technical problem to be solved in this field.

[0003] In the prior art, patent application CN111487202A discloses a method for online control of the processing of Typha pollen. This method includes two steps: measurement and identification. Color parameter values ​​during the processing of Typha pollen are collected using a spectrophotometer, and the total color value and color difference value compared to the standard raw product are calculated. Alternatively, a discriminant function formula is used to determine whether the processed product is a qualified charcoal product. The method also specifies the corresponding processing conditions for qualified charcoal products, such as a processing temperature of 170-200℃ and a processing time of 10-30 minutes. It is objective, rapid, and reliable, solving the problem of lack of objectivity in traditional human judgment. Patent application CN119338329A discloses an online management method and system for quality data during the processing of traditional Chinese medicine decoction pieces. This system includes modules for data acquisition, preprocessing, model training, and quality prediction. By collecting key quality data during the processing, it uses an ant colony optimization algorithm to optimize the parameters of the neural network model, achieving real-time monitoring and anomaly warning of quality indicators, thus improving the accuracy and real-time performance of quality prediction.

[0004] However, while the two existing technologies mentioned above have some value in online monitoring and intelligent data management of traditional Chinese medicine processing, they fail to address the core pain points of "different qualities for the same specifications," semantic aliasing, and the identification of latent anomalies in the current processing process. Specifically, patent CN111487202A focuses only on static monitoring of a single color modal signal, without involving the fusion analysis of multi-dimensional features such as sound, light, electricity, and heat. It cannot capture the nonlinear dynamic changes in the processing process and is prone to misjudging the true processing state due to instantaneous features. While patent CN119338329A optimizes the prediction model parameters, it lacks in-depth modeling of the temporal coupling relationship between multi-dimensional feature terms, making it difficult to eliminate semantic aliasing effects caused by moisture obstruction and physical noise. Furthermore, it does not form an interpretable analysis system based on pharmacological mechanisms. Neither technology achieves the transformation from static single-parameter monitoring to dynamic multi-modal semantic alignment, making it impossible to accurately quantify the processing evolution trajectory and meet the refined requirements for quality consistency control in modern traditional Chinese medicine production. Summary of the Invention

[0005] This invention is applicable to Chinese medicine processing production lines of various scales, meeting the needs of multi-dimensional physical signal acquisition and automatic quality monitoring. By calibrating information emergence points through perceptual entropy sequences and reconstructing characteristic respiration spectra, it achieves precise capture and structured spatial representation of the sensitivity window of the processing process. The process semantic map and process manifold transform discrete process stage markers into continuous mathematical evolution trajectories and abstract pharmacological experience into digital mapping benchmarks, solving the technical problem that traditional algorithms cannot understand the logical continuity of process evolution. Semantic orthogonal decomposition distance deconstructs high-dimensional geometric deviations into quality degradation degree and process deviation degree, achieving accurate quantitative diagnosis of processing status. The virtual correction force field combined with risk inference automatically generates active intervention instructions for physical control parameters and outputs structured abnormality diagnosis reports, realizing a leap from passive detection to active early warning and interpretable intervention, ensuring the inherent safety and batch quality uniformity of modern Chinese medicine production.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] Large-scale feature classification algorithms for detecting anomalies in the processing of traditional Chinese medicine include:

[0008] A multimodal real-time dataset representing the co-evolution of the physicochemical properties of medicinal materials during the processing of traditional Chinese medicine was obtained. Parallel analysis was performed on the multimodal real-time dataset to obtain information emergence points that predict qualitative changes in the properties of medicinal materials. The multimodal real-time dataset was reconstructed into a feature respiration spectrum with the information emergence points as the central origin.

[0009] The process semantic map is constructed by calling the feature breathing spectrum, and the process manifold representing the continuous evolution trajectory of the ideal processing state is generated. Vector decomposition is performed using the process manifold to obtain the semantic orthogonal decomposition distance representing the essential characteristics and progress deviation of the processing process.

[0010] A virtual correction force field is constructed based on the semantic orthogonal decomposition distance, and the virtual correction force field is decomposed to obtain active intervention commands for physical control parameters. Based on the active intervention commands, risk inference is driven to generate an anomaly diagnosis report characterizing the causes of abnormalities in the processing process.

[0011] Furthermore, the method for acquiring the multimodal real-time dataset includes:

[0012] Raw visible light images were acquired using an industrial camera, from which the surface chromaticity values ​​of the medicinal materials were extracted and defined as visual data. Raw infrared thermal images were acquired using a thermal imaging camera, from which the material uniform temperature values ​​were extracted and defined as thermal field data. Instantaneous power spectral density, waveform kurtosis index, and characteristic absorption peak intensity were extracted using a microphone, Hall current sensor, and pushbroom camera and defined as acoustic spectrum data, load data, and chemical fingerprint data, respectively.

[0013] The global reference clock is invoked to perform timestamp association and synchronization on the feature items, which include visual data, thermal field data, acoustic spectrum data, load data, and chemical fingerprint data.

[0014] A baseline time axis is established using chemical fingerprint data. Visual data, thermal field data, acoustic spectrum data, and load data are mapped onto the baseline time axis using the nearest neighbor interpolation method to obtain the multimodal real-time dataset with instantaneous physical semantic alignment.

[0015] Furthermore, the information emergence points include:

[0016] The first derivative time series of visual data, thermal field data, acoustic spectrum data, and load data are acquired. The absolute sum of the first derivative time series is calculated and weighted to obtain the modal internal variability.

[0017] Extract the time-series feature values ​​of five feature items from the multimodal real-time dataset, calculate the real-time covariance values ​​between any two feature items and combine them to generate a real-time covariance matrix, and calculate the intermodal covariance divergence by comparing the real-time covariance matrix with the pre-stored benchmark covariance matrix.

[0018] Set energy weights and collaborative weights, and perform multiplication operations with intramodal variability and intermodal covariance divergence respectively, and sum them to obtain the single-dimensional perceptual entropy value;

[0019] The single-dimensional perceptual entropy values ​​are arranged in chronological order to form a perceptual entropy sequence. If the single-dimensional perceptual entropy value is greater than the dynamic feature threshold at multiple consecutive sampling times, it is marked as an information emergence point.

[0020] Furthermore, the method for obtaining the characteristic respiratory spectrum includes:

[0021] Extract the numerical sequence before and after the information emergence point for a preset time period and normalize it to obtain the temporal components of each feature item;

[0022] A two-dimensional feature space containing an evolution time axis and a causal feature axis is constructed. Based on the order of the physical response of the five feature items in the processing technology, decreasing values ​​are assigned sequentially from the source end to the end end, which are defined as weight scalar values. The causal energy level weights are composed of each weight scalar value.

[0023] Based on the order of the weight scalar values ​​in the causal energy level weight from large to small, the causal feature axis is divided into five feature mapping rows in the vertical direction, and the time-series components corresponding to the thermal field data, visual data, load data, acoustic spectrum data and chemical fingerprint data are sequentially locked into the corresponding feature mapping rows.

[0024] The amplitude of each time-series component is mapped to the pixel brightness value of the corresponding feature mapping row coordinate point, and the combination is used to generate the feature breathing spectrum.

[0025] Furthermore, the method for obtaining the process manifold includes:

[0026] Construct a process semantic graph consisting of process nodes representing key stages of processing and logical edges representing time evolution constraints, and use process text to parameterize the ideal physical state of each feature item to associate process nodes.

[0027] A large number of ideal processing batches that are judged to be qualified are obtained and corresponding characteristic respiration spectra are generated, which are defined as ideal characteristic respiration spectra. Ideal feature vectors are extracted using a contrastive language image pre-trained model to form an ideal feature vector set, and the process text is converted into semantic vectors.

[0028] Extract the local neighborhood topological relationships of the ideal feature vector set, fit and generate a smooth and continuous low-dimensional subspace image to obtain the process manifold.

[0029] Furthermore, the method for obtaining the semantic orthogonal decomposition distance includes:

[0030] Map the currently generated real-time feature respiration spectrum to a real-time feature vector, and locate the ideal projection point on the process manifold that has the smallest Euclidean distance from the real-time feature vector;

[0031] Calculate the Euclidean distance between the real-time feature vector and the ideal projection point to obtain the quality degradation degree that represents the deviation of the essential attribute;

[0032] Obtain the semantic vector corresponding to the current process node, calculate the geodesic distance from the ideal projection point to the position of the semantic vector on the process manifold, and obtain the process deviation that characterizes the achievement rate of the processing temperature.

[0033] The semantic orthogonal decomposition distance is composed of the degree of quality degradation and the degree of process deviation.

[0034] Furthermore, the active intervention command includes:

[0035] A virtual correction force field is constructed using the process manifold as the zero potential energy reference and the quality degradation degree as the potential energy gradient.

[0036] Calculate the virtual correction force pointing from the real-time feature vector to the ideal projection point, and decompose the virtual correction force into a set of intervention components in the dimensions of heating power, stirring speed and amount of additives;

[0037] Dynamic gain correction is performed on the intervention component set based on the positive or negative attribute of the process deviation: if the process deviation is positive, the power compensation value is increased; if the process deviation is negative, the power reduction value is increased.

[0038] The set of intervention components after dynamic gain correction is encapsulated into an active intervention command consisting of the target heating power value, the target stirring speed value, and the target amount of auxiliary material added.

[0039] Furthermore, the risk simulation includes:

[0040] Obtain the characteristic respiratory spectrum sequence of historical failed batches, and map the characteristic respiratory spectrum sequence into a high-dimensional feature vector sequence;

[0041] The Long Short-Term Memory (LSTM) neural network is invoked, and real-time feature vectors are input into the LSM to simulate the predicted evolution trajectory under conditions where no active intervention commands are executed.

[0042] Extract the feature vector generated by mapping historical failed batches, define it as an abnormal feature vector, perform cluster analysis on the abnormal feature vector, delineate the risk endpoint region and bind anomaly labels;

[0043] By using the abnormal feature vectors of historical failed batches, risk endpoint regions are delineated in a multidimensional feature space through clustering, and abnormal labels including scorched labels, half-cooked labels, and labels indicating excessive loss of effective ingredients are bound.

[0044] Collision detection is performed between the predicted evolution trajectory and the risk endpoint region. If the trajectory enters the risk endpoint region, the corresponding anomaly label is automatically extracted.

[0045] Furthermore, the method for obtaining the abnormal diagnostic report includes:

[0046] Based on the values ​​of quality degradation degree and process deviation degree, a description of trait deviation is generated;

[0047] Based on the depth and anomaly labels of the predicted evolution trajectory falling into the risk endpoint area, the expected risk consequences are generated.

[0048] Active intervention commands are mapped into natural language command sets to generate process adjustment schemes;

[0049] By combining descriptions of trait deviations, expected risk consequences, and process adjustment schemes, a structured anomaly diagnosis report is output.

[0050] A large-scale feature classification system for detecting anomalies in the processing of traditional Chinese medicine, which is used to implement the aforementioned large-scale feature classification algorithm for detecting anomalies in the processing of traditional Chinese medicine, the system comprising:

[0051] Feature Reconstruction Module: This module is used to acquire a multimodal real-time dataset that characterizes the co-evolution of the physicochemical properties of medicinal materials during the processing of traditional Chinese medicine. Parallel analysis is performed on the multimodal real-time dataset to obtain information emergence points that predict qualitative changes in the properties of medicinal materials. Using the information emergence points as the central origin, the multimodal real-time dataset is reconstructed into a feature respiration spectrum.

[0052] The trajectory quantization module is used to construct a process semantic map by calling the feature breathing spectrum and generate a process manifold that represents the continuous evolution trajectory of the ideal processing state. The process manifold is used to perform vector decomposition to obtain the semantic orthogonal decomposition distance that represents the essential characteristics and progress deviation of the processing process.

[0053] Correction and Diagnosis Module: This module is used to construct a virtual correction force field based on the semantic orthogonal decomposition distance, and decompose the virtual correction force field to obtain active intervention commands for physical control parameters. Based on the active intervention commands, risk inference is driven to generate an anomaly diagnosis report characterizing the causes of abnormalities in the processing process.

[0054] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0055] This invention uses perceptual entropy sequences to identify information emergence points and reconstructs multimodal real-time datasets into characteristic respiratory spectra, achieving precise capture and structured representation of sensitive state windows in the processing process. This solves the technical problem of key signals being smoothly masked due to rigid data sampling strategies in traditional methods. Process semantic graphs and process manifolds transform discrete process nodes into continuous high-dimensional evolutionary trajectories, reconstructing abstract pharmacological experience into digital mapping benchmarks, overcoming the limitation of traditional anomaly detection algorithms in understanding the logical continuity of process evolution. Semantic orthogonal decomposition distance deconstructs high-dimensional geometric deviations into quality degradation degree and process deviation degree, achieving accurate quantitative diagnosis of trait quality changes and progress deviations. Virtual correction force fields combined with risk inference realize an intelligent leap from passive detection to proactive early warning and interpretable intervention, significantly improving the inherent safety and reliability of quality traceability in traditional Chinese medicine manufacturing. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 The flowchart of the large-scale feature classification algorithm for detecting abnormalities in the processing of traditional Chinese medicine provided in this embodiment of the invention;

[0058] Figure 2 This is a schematic diagram of the spatial deployment of a multimodal sensor and a processing pot provided in an embodiment of the present invention;

[0059] Figure 3 A schematic diagram showing the generation logic and morphological comparison of a characteristic respiratory spectrum provided in an embodiment of the present invention;

[0060] Figure 4 A mathematical topological diagram of a process manifold provided in an embodiment of the present invention;

[0061] Figure 5 The functional module diagram of the large-scale feature classification system for detecting abnormalities in the processing of traditional Chinese medicine provided in this embodiment of the invention is shown.

[0062] Reference numerals: 1. Industrial camera; 2. Thermal imaging camera; 3. Pushbroom camera; 4. Cooking pot; 5. Stirring paddle; 6. Microphone 1; 7. Microphone 2; 8. Drive shaft; 9. Stirring motor; 10. Hall current sensor. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] Example 1

[0065] Please see Figure 1 As shown, this embodiment provides a large-scale feature classification algorithm for detecting anomalies in the processing of traditional Chinese medicine, including:

[0066] Step S10: Obtain a multimodal real-time dataset that characterizes the co-evolution of the physical and chemical properties of medicinal materials during the processing of traditional Chinese medicine. Perform parallel analysis on the multimodal real-time dataset to obtain information emergence points that predict qualitative changes in the properties of medicinal materials. Using the information emergence points as the central origin, reconstruct the multimodal real-time dataset into a feature respiration spectrum.

[0067] Further, step S10 includes:

[0068] Step S11: Obtain a multimodal real-time dataset that characterizes the co-evolution of the physicochemical properties of medicinal materials during the processing of traditional Chinese medicine.

[0069] In the nonlinear, strongly coupled thermodynamic and chemical reaction process of traditional Chinese medicine processing, the evolution of the processing state is not uniform. Instead, within a mostly stable period, there are several extremely short critical windows that foreshadow quality turning points. Traditional monitoring methods, due to their fixed data sampling and processing strategies, often smooth out the weak signal fluctuations carrying high information content within these critical windows as noise during averaging or filtering, leading to a lag in the identification of potential quality risks. Therefore, establishing a dynamic mechanism that can adaptively identify and focus on key windows transforms passive, undifferentiated data recording into an active, intelligent perception mode capable of distinguishing between mundane and critical transition states. This provides a multimodal real-time dataset with the highest information density and no contamination. The multimodal real-time dataset possesses the characteristics of the synergistic evolution of the physicochemical properties of medicinal materials. Specifically, it refers to the dynamic evolution of the external physical properties of medicinal materials, such as surface color, overall thermal field, physical brittleness caused by moisture loss, and mechanical viscous resistance, and the internal chemical properties, such as deep reactions at the molecular level and the transformation of effective chemical components, over time during the nonlinear and strongly coupled thermodynamic and kinetic reaction process of traditional Chinese medicine processing.

[0070] Specifically, a multimodal real-time dataset capable of comprehensively and multidimensionally describing the state of the processing process is constructed. This multimodal real-time dataset is a time-synchronized collection of signals acquired in parallel at specific locations on the processing pot by various physical sensors at preset frequencies. (See [link to relevant documentation]). Figure 2 This is a schematic diagram of the spatial deployment of a multimodal sensor and a processing pot according to an embodiment of the present invention. The processing pot 4 in the figure is a gray, drum-shaped pot with a distinctly thick-walled structure. The thick-walled structure is composed of an outer dark gray protective shell and an inner light gray thermally conductive lining. The function of the thick-walled structure is to maintain heat and conduct heat evenly. An industrial camera 1 and a thermal imaging camera 2 are fixed horizontally and parallelly above the opening of the processing pot 4. The purpose is to ensure that the industrial camera 1 and the thermal imaging camera 2 have the same field of view, defined as the core monitoring field of view. Specifically, the core monitoring field of view refers to the spatial range covering the bottom and sidewall material tumbling area of ​​the light gray thermally conductive lining inside the processing pot 4, defined by the installation height and lens focal length of the industrial camera 1 and the thermal imaging camera 2, as shown by the light blue transparent area in the figure. This is used to lock onto the area for collecting all visible light and infrared radiation information during the heating and stir-frying process of the medicinal materials. The two gray lines extending from the positions of the industrial camera 1 and the thermal imaging camera 2 in the figure schematically represent the field of view range of the industrial camera 1 and the thermal imaging camera 2. Two circular sensor components, represented by blue circles in the figure, namely microphone 6 and microphone 7, are deployed on the outer edge of the thick-walled structure of the processing pot 4. Microphone 6 and microphone 7 maintain strict geometric symmetry as the processing pot 4 rotates. A Hall current sensor 10 is represented by a striking orange ring component; the Hall current sensor 10 is surrounded by the power supply cable of the driving stirring motor 9 in bright blue in the figure. To improve physical identification and assist in the positioning of the Hall current sensor 10, a pushbroom camera 3 is deployed on the upper side of the processing pot 4. The dashed line emanating from the center of the lens of the pushbroom camera 3 in the figure represents the direction of the detection optical axis of the pushbroom camera 3, which is intended to point to the medicinal material area at the bottom of the processing pot 4. The stirring motor 9 is located on the equipment base, and the power output end of the stirring motor 9 is connected to the stirring paddle 5 inside the processing pot 4 through the drive shaft 8 to drive the stirring paddle 5 to rotate. The blade edge of the stirring paddle 5 is designed with a curved structure that matches the curvature of the inner wall of the light gray thermally conductive lining, ensuring that the end curve of the blade can completely fit the inner wall contour of the light gray thermally conductive lining during the rotation of the stirring paddle 5, so as to achieve the turning of the medicinal materials without dead angles.

[0071] Specifically, the industrial camera and the thermal imaging camera collect visual data and thermal field data at the same frequency, respectively. The reason for setting the same frequency is that the color change on the surface of the medicinal material and the macroscopic thermal field distribution are slow-evolving physical processes. It is necessary to capture the dynamic trajectory of the rolling of the medicinal material while satisfying the persistence of vision effect, and at the same time avoid generating too much image data redundancy. For example, it is set to 30 Hz. The visual data refers to the surface chromaticity values ​​of medicinal materials extracted from raw visible light images captured by an industrial camera. These chromaticity values ​​are color space vectors representing the degree of Maillard reaction after the medicinal materials are heated. The purpose of collecting these values ​​is to quantify the processing temperature of the medicinal materials by analyzing the evolution of their color intensity. The thermal field data refers to the average temperature of the medicinal materials extracted from raw infrared thermal images of the core monitoring field of view within the processing pot, captured by a thermal imaging camera. The average temperature of the medicinal materials is defined as a statistical measure representing the overall thermal energy level of the medicinal material group within the core monitoring field of view. The purpose of collecting this data is to penetrate the water vapor interference generated during the processing of traditional Chinese medicine and obtain the true heat absorption efficiency of the medicinal materials in real time. Microphone 6 and microphone 7 are of the same model and collect voiceprint spectrum data at a preset frequency. The frequency setting is based on the fact that the voiceprint signal contains a large number of high-frequency mechanical impact sounds and moisture vaporization and bursting sounds, ensuring complete coverage of the subtle audio features above the range of human hearing and ensuring the fidelity of the voiceprint spectrum data. For example, the frequency is set to 44.The 1 kHz acoustic spectrum data refers to the instantaneous power spectral density extracted from the original audio signals picked up by two microphones. This instantaneous power spectral density characterizes the energy distribution of the original audio signal at different frequency components. The purpose of this acquisition is to determine the progress of the drying and crispening of the medicinal materials due to internal moisture loss by monitoring changes in the dullness or crispness of the sound frequencies of the collisions between the medicinal materials. The Hall current sensor collects load data at a preset frequency by sensing the alternating magnetic field in the bright blue power cable. The frequency setting is based on the fact that the phase current fluctuation of the stirring motor is mainly affected by the viscous resistance of the material, and its evolution rate is between visual and acoustic characteristics. Therefore, it is set to a frequency sufficient to capture the current distortion and kurtosis changes generated by the motor rotor when overcoming instantaneous agglomeration resistance. For example, it is set to 1 kHz. The load data refers to the waveform kurtosis index extracted from the original current waveform signal sensed by the Hall current sensor. This waveform kurtosis index characterizes the degree of sharpness of the flat distribution of the stirring motor current waveform. The purpose of collecting dimensional parameters is to quantify the instantaneous viscous resistance mutations in medicinal materials after the addition of excipients, thereby identifying whether abnormal agglomeration or clustering occurs. The pushbroom camera samples the optical axis of the probe and collects chemical fingerprint data at a preset frequency. The frequency setting is based on the transformation of chemical components within the medicinal material. For example, alkaloid hydrolysis is a deep reaction at the molecular level, and the shift of its characteristic absorption peak has a significant time lag. A frequency sufficient to monitor the transformation gradient of effective chemical components and reduce the processing pressure on the hyperspectral data cube is set; for example, it is set to 0.1 Hz. The chemical fingerprint data refers to the intensity of characteristic absorption peaks extracted from the original spectral data cube located within the spectral scanning field of view obtained from the pushbroom camera. The intensity of the characteristic absorption peak is the light energy absorption density of a specific molecular bond in the near-infrared band. The purpose of collecting this data is to perceive the transformation depth of effective components within the medicinal material in real time through molecular-level energy level transition information, thereby achieving the monitoring of the essential properties of medicinal material quality. By employing multi-source data time alignment technology, the five feature items—visual data, thermal field data, acoustic signature spectrum data, load data, and chemical fingerprint data—are time-stamped and synchronized based on a global reference clock, resulting in a structured vector sequence containing multiple independent physical semantic dimensions, i.e., a multimodal real-time dataset.

[0072] Step S12: Perform parallel analysis on the multimodal real-time dataset to obtain information emergence points that predict qualitative changes in the properties of medicinal materials.

[0073] After acquiring the multimodal real-time dataset, to transform it from a simple collection of physical signals into a comprehensive index that quantifies the evolutionary characteristics of the processing process, a composite index perceptual entropy calculation is performed to generate a composite index perceptual entropy that reflects the overall evolution of the processing process. This composite index perceptual entropy is a dynamic entropy value that measures the overall uncertainty and information density of the processing process, including intramodal variability and intermodal covariance divergence. Parallel parsing refers to synchronously stripping features from data streams such as visual, acoustic, thermal, and current data with different sampling frequencies and physical dimensions, ensuring that features from different modalities are processed under the same logical benchmark through global timestamp alignment.

[0074] The modal variability is a feature quantity characterizing the degree of amplitude fluctuation of each feature item (excluding chemical fingerprint data) on a microscopic time scale in a multimodal real-time dataset. Specifically, the calculation process of the modal variability is as follows: a short time window covering a single collision cycle of medicinal materials is preset, for example, the short time window is set to 100 milliseconds. Within the short time window, the surface chromaticity value of the medicinal materials corresponding to the visual data, the average temperature of the medicinal materials corresponding to the thermal field data, the instantaneous power spectral density corresponding to the acoustic signature spectrum data, and the waveform kurtosis index corresponding to the load data are extracted respectively. The first derivative time series of the surface chromaticity value of the medicinal materials, the average temperature of the medicinal materials, the instantaneous power spectral density, and the waveform kurtosis index are calculated by differential calculation. The first derivative time series is a set of values ​​characterizing the instantaneous rate of change of each feature item with the sampling step size. The absolute values ​​of all values ​​in the first derivative time series are accumulated and summed. The weighted sum of the absolute values ​​of each feature item is then obtained to obtain the modal variability. Furthermore, the reason why chemical fingerprint data is not incorporated into the intramodal variability is that the sampling frequency of chemical fingerprint data is much lower than that of a short time window, and the transformation of the effective components of medicinal materials is a deep and slow reaction, lacking the fluctuation characteristics to characterize variability. The purpose of the intramodal variability is to capture the instantaneous jump signals caused by the explosion, sticking, or heating of medicinal materials inside the processing pot. The intermodal covariance divergence is a deviation measure characterizing the degree of deviation of the inherent process co-logic between the feature items, including chemical fingerprint data, in the multimodal real-time dataset. Specifically, the calculation process of the intermodal covariance divergence is as follows: a long time window covering the complete stirring cycle is preset, for example, the long time window is set to 1 second, and a pre-stored benchmark covariance matrix is ​​called; the benchmark covariance matrix records the ideal linear correlation pattern between the surface color value of the medicinal materials, the average temperature of the medicinal materials, the instantaneous power spectral density, the waveform kurtosis index, and the intensity of the characteristic absorption peak under standard process conditions. Specifically, a large amount of sample data of standard processing conditions is acquired, and the joint distribution characteristics of each feature item in the sample data on the time axis are statistically analyzed. The covariance value between any two feature items is calculated, and the covariance values ​​are arranged in row and column order to form a benchmark covariance matrix that records the ideal linear correlation mode between the surface color value of the medicinal material, the average temperature of the medicinal material, the instantaneous power spectral density, the waveform kurtosis index, and the intensity of the characteristic absorption peak. By extracting the time-series feature value set of each feature item in the multimodal real-time dataset within a long time window, the time-series feature value set refers to the numerical vector composed of the amplitude of the corresponding feature item at all sampling times within the long time window arranged in chronological order; the real-time covariance value between any two numerical vectors of different categories in the time-series feature value set is calculated in turn. The real-time covariance value is used to quantify the degree of synchronous evolution of the two feature items within the long time window.All real-time covariance values ​​are combined in the same order as the baseline covariance matrix to obtain the real-time covariance matrix. The Frobenius norm distance between the real-time covariance matrix and the baseline covariance matrix is ​​calculated to obtain the intermodal covariance divergence. The Frobenius norm distance is a spatial distance index that measures the overall numerical difference between two matrices. The purpose of collecting the intermodal covariance divergence is to quantify the degree of deconstruction of the inherent correlation patterns between feature items in the multimodal real-time dataset constrained by physicochemical laws when anomalies occur in the processing process. After obtaining the intramodal variability and intermodal covariance divergence, in order to perform dimensionality compression and feature aggregation on the energy activity of the processing process at the microscale and the co-logistic deviation at the macroscale, a weighted summation operation is performed to combine the intramodal variability and intermodal covariance divergence into a perceptual entropy sequence reflecting the overall uncertainty transition of the processing process. The specific process is as follows: A set of energy weights and collaborative weights reflecting the contribution of different physical properties are preset. The setting is based on the physical brittleness characteristics of the current processed medicinal materials, such as whether they are prone to cracking or breaking during the frying process. The weights are fitted by the amplitude contribution ratio of abnormal sudden signals in historical production data to enhance the sensitivity of the perception entropy to micro-energy jumps. The energy weights are used to quantify the influence of the internal variability of the modality on the instantaneous stability of the processing process. The collaborative weights are used to quantify the influence of the covariance divergence between modalities on the process matching degree of the processing process. The weights are set based on the causal coupling strength between each feature item. For example, when the processing is in the stage of drastic chemical component transformation, the weight ratio of the covariance item associated with the chemical fingerprint data is increased to enhance the monitoring depth of the perception entropy on the loss of macro-process logic.

[0075] The micro-perturbation component is obtained by multiplying the intramodal variability with the energy weight; the macro-instability component is obtained by multiplying the intermodal covariance divergence with the cooperative weight; and the micro-perturbation component and the macro-instability component are summed to obtain the single-dimensional perceptual entropy value at the corresponding time moment. The perceptual entropy sequence is defined as a scalar numerical sequence composed of the single-dimensional perceptual entropy values ​​corresponding to each sampling time moment arranged in chronological order. The purpose is to provide a unique discrimination criterion for subsequent processing.

[0076] To pinpoint sensitive signal moments indicating physical or chemical changes during the processing from a continuous and redundant multimodal real-time dataset, the information emergence points are calibrated by monitoring the fluctuation patterns of the perception entropy sequence in real time. Specifically, the current single-dimensional perception entropy value is acquired, and a dynamic feature threshold is set. This dynamic feature threshold is based on the fact that the periodic mechanical tumbling of the agitator and the random rolling of materials inside the processing pot cause background random fluctuations accompanying the process rhythm, resulting in a slow baseline drift in the single-dimensional perception entropy value. The dynamic feature threshold is constructed by extracting the moving average of the perception entropy sequence within a historical sliding window and superimposing it with the fluctuation margin obtained by weighted summation of the standard deviations of the five feature items within the historical sliding window. A continuous threshold judgment is performed on the single-dimensional perceptual entropy value. Specifically, the single-dimensional perceptual entropy value obtained at each sampling moment is compared with the dynamic feature threshold. Only when the single-dimensional perceptual entropy value is greater than the dynamic feature threshold for N consecutive sampling moments is it determined that a non-stationary signal with process quality evolution significance has appeared in the multimodal real-time dataset. The moment when the single-dimensional perceptual entropy value first exceeds the dynamic feature threshold in N sampling moments is marked as the information emergence point. The purpose is to lock the critical trigger moment of process quality change in complex processing environment. The value of N is set according to the ratio of the maximum sampling frequency, i.e., the sampling frequency of the acoustic spectrum data, to the minimum physical evolution period of the medicinal material in the processing pot for quality change. The minimum physical evolution period is the minimum duration required for the medicinal material to undergo a single burst event or irreversible carbonization mutation due to heating in the processing pot. The purpose is to filter out isolated jump interference caused by non-trend noise and ensure that the marked information emergence point belongs to a steady-state response with physical continuity.

[0077] Step S13: Using the information emergence point as the central origin, reconstruct the multimodal real-time dataset into a feature breathing spectrum.

[0078] After identifying the information emergence points, in order to transform the heterogeneous and discrete numerical sequences contained in the multimodal real-time dataset into a structured spatial image that can characterize the qualitative changes in physical states, a feature breathing spectrum reflecting the state evolution texture of the processing process is generated by executing data truncation and feature arrangement logic. The aim is to transform the temporal fluctuations of each feature item into a two-dimensional image with intuitive physical causal logic.

[0079] Specifically, taking the sampling time corresponding to the information emergence point as the center origin of the time axis, a numerical sequence of a preset duration before and after the corresponding information emergence point is extracted from the multimodal real-time dataset to obtain the original feature sequence fragment corresponding to each feature item. The preset duration is set based on the complete causal transmission cycle of the medicinal materials in the processing pot, from the induction of physical property changes to the end of chemical component response. This aims to ensure that the original feature sequence fragment can completely cover all related feature evolution patterns before and after the information emergence point is triggered. For example, the preset duration is set to 2 seconds. The original feature sequence fragment refers to a subset extracted from the multimodal real-time dataset, consisting of the numerical points corresponding to each feature item within the preset duration, arranged in time sequence. Normalization processing is performed on the original feature sequence fragment to obtain the standardized temporal component corresponding to each feature item. The temporal component is a numerical sequence consisting of multiple discrete sampling time points and their corresponding normalized amplitudes after uniformly mapping the amplitude range of the five feature items within the preset duration to the [0,1] interval.

[0080] Specifically, in order to quantify the causal relationship between the five features, the temporal components are subjected to spatial topological rearrangement based on the causal chain, according to the thermodynamic and kinetic conduction logic in the processing. The specific process is as follows: A two-dimensional feature space is created, with the horizontal axis defined as the evolution time axis. This is obtained by extracting each discrete sampling time point contained in the time-series component and mapping it to the horizontal axis coordinate according to the chronological order of the discrete sampling time points. This is used to lock the temporal evolution trajectory of the processing process within the preset time period. The vertical axis of the two-dimensional feature space is defined as the causal feature axis reflecting the logical hierarchy of the five feature items. According to the response order of the five feature items in the processing process, decreasing weight scalar values ​​are assigned sequentially from the source end to the end end. The weight scalar values ​​constitute the causal energy level weight. The causal energy level weight is used to lock the vertical distribution pose of each feature item on the causal feature axis. That is, the thermal field data is assigned the highest weight scalar value, the visual data is assigned the second highest weight scalar value, the load data is assigned the medium weight scalar value, the acoustic spectrum data is assigned the second lowest weight scalar value, and the chemical fingerprint data is assigned the lowest weight scalar value. Based on the descending order of the weight scalar values, the causal feature axis is divided vertically into five parallel feature mapping rows at equal intervals. This ensures that the time-series component corresponding to the thermal field data with the largest weight scalar value is locked in the topmost mapping row of the causal feature axis, while the time-series component corresponding to the chemical fingerprint data with the smallest weight scalar value is locked in the bottommost mapping row. The time-series components corresponding to the remaining feature items are then sequentially filled into the middle feature mapping rows according to their weight scalar values. This spatial layout of the causal feature axis effectively solidifies the physical progression logic of energy input inducing phenotype responses during processing, leading to component transformations. The amplitude of each feature item after arrangement at each discrete sampling time point is mapped to the pixel brightness value of the corresponding coordinate point in the feature mapping row in the two-dimensional feature space, thereby combining to generate a feature breathing spectrum with a specific qualitative change texture. The feature breathing spectrum is a dynamic feature map generated by reconstructing multimodal feature vectors under spatial causal topological constraints using the identified information emergence points as spatiotemporal anchor points. The causal energy level weight is a weight scalar value used to lock the vertical distribution pose of each feature item on the causal feature axis. The setting is based on using the order of the numerical values ​​to quantify the physical response priority of the five feature items in the internal composition qualitative change from external heat energy exchange inside the cooking pot.For example, the thermal field data characterizing the initial energy input is assigned the highest weight scalar value of 0.9, the visual data characterizing the degree of Maillard reaction is assigned the second highest weight scalar value of 0.7, the load data characterizing the sudden change in material viscous resistance is assigned the medium weight scalar value of 0.5, the acoustic spectrum data characterizing the progress of moisture bursting and drying is assigned the second lowest weight scalar value of 0.3, and the chemical fingerprint data characterizing the final goal of effective component conversion is assigned the lowest weight scalar value of 0.1. By sorting the weight scalar values ​​by size, the causal feature axis is divided into five equidistant feature mapping rows in the vertical direction, thereby locking the topological pose of different physical semantics on the causal chain in the two-dimensional feature space, see below. Figure 3 This is a schematic diagram comparing the generation logic and morphology of a characteristic breathing spectrum provided by an embodiment of the present invention. The diagram illustrates the process of transforming the multimodal real-time dataset into a structured spatial image capable of characterizing the qualitative changes in physical states. The left side of the diagram shows the constructed two-dimensional feature space, which includes a causal feature axis and an evolution time axis. On the causal feature axis, according to the weight scalar values ​​of the causal energy level weights, the two-dimensional feature space is divided into five parallel regions, i.e., feature mapping rows, in the vertical direction from top to bottom; adjacent feature mapping rows are separated by horizontal dashed lines. This solidifies the physical progression relationship from external heat energy exchange to internal component qualitative changes. Figure 3 As shown on the right, by mapping the amplitude of the temporal component corresponding to each feature item at each discrete sampling time point to a pixel brightness value, a feature respiration spectrum with specific qualitative change texture is finally generated. An ideal respiration spectrum exhibits smooth transitions between color blocks at each level, continuous texture, and temporal rhythm, indicating that the evolution of various physical quantities of the medicinal material inside the processing pot conforms to a preset physical causality law. An abnormal respiration spectrum, however, exhibits a pathological texture, specifically manifested as follows: due to abnormal evolution of visual data, a visually abnormal area represented by light red blocks appears in the second feature mapping row; due to lag or distortion in the molecular energy level transformation of chemical fingerprint data, a chemically abnormal area represented by light red blocks appears in the fifth feature mapping row; the visually abnormal area and the chemically abnormal area lose coordination due to a loss of evolutionary logic, which is shown in the figure as red circles representing temporal misalignment spots. This intuitive presentation of multimodal feature logic misalignment can identify hidden quality changes in medicinal materials within the processing pot.

[0081] Step S10 addresses the technical challenges in traditional Chinese medicine processing monitoring, such as the smooth masking of key qualitative change signals due to rigid data sampling strategies and the difficulty in establishing intuitive causal relationships between heterogeneous multidimensional signals. This is achieved by acquiring a multimodal real-time dataset, calibrating information emergence points, and reconstructing the characteristic respiration spectrum. Specifically, the multimodal real-time dataset provides comprehensive underlying data support covering appearance, thermal energy, acoustic signature, load, and chemical fingerprint; information emergence points pinpoint the triggering moments indicating process qualitative changes from massive amounts of bland background signals, resolving the computational overload problem caused by data redundancy; and the characteristic respiration spectrum transforms discrete time-series numerical sequences into two-dimensional images with physical logical depth, providing a structured input carrier for subsequent classification and recognition of processing states.

[0082] Step S20: Construct a process semantic graph by calling the feature breathing spectrum, and generate a process manifold that represents the continuous evolution trajectory of the ideal processing state. Perform vector decomposition using the process manifold to obtain the semantic orthogonal decomposition distance that represents the essential characteristics and progress deviation of the processing process.

[0083] Further, step S20 includes:

[0084] Step S21: Invoke the feature breathing spectrum to construct the process semantic map and generate the process manifold representing the continuous evolution trajectory of the ideal processing state.

[0085] After obtaining the characteristic respiratory spectrum, in order to elevate the processing procedure from simple image feature recognition to a process diagnostic level with pharmacological logical understanding, a process semantic graph is constructed and a process manifold is generated by executing knowledge graph mapping and high-dimensional space modeling logic. The aim is to establish a digital reference system containing expert knowledge for heterogeneous multimodal real-time datasets, thereby addressing the technical deficiency of traditional detection technologies in lacking interpretability of pharmacological mechanisms.

[0086] The process semantic graph is a topological knowledge network composed of multiple process nodes and logical edges connecting them. Each process node is defined as a digital logical marker for a key stage in a specific traditional Chinese medicine processing process. For example, in the processing of Aconitum carmichaelii (Fuzi), the process nodes include the initial uniform heating stage, the intermediate rapid dehydration stage, and the later component transformation stage. Each process node is associated with a process text, which is a parameterized definition of the ideal physical state that visual data, thermal field data, acoustic spectrum data, load data, and chemical fingerprint data should achieve at the stage of the process node, using a digital description language. For example, for the process node corresponding to the intermediate rapid dehydration stage, the corresponding process text is: the surface chromaticity value of the medicinal material is within a preset light yellow feature vector range; the average temperature of the medicinal material remains within a preset moisture evaporation plateau value; the instantaneous power spectral density exhibits a preset dense collision high-frequency characteristic; the waveform kurtosis index is within a stable fluctuation range; and the intensity of the characteristic absorption peak shows that the aconitine content begins to decrease linearly. Logical edges are established between two adjacent process nodes. These logical edges represent legitimate path constraints that characterize the evolution of the processing procedure from one process stage to the next on the timeline. By combining the process node text and logical edges, the process semantic graph can transform the originally abstract process specifications into a semantic reference framework composed of multiple feature vectors with causal logical relationships.

[0087] To further establish continuous and high-precision ideal evolution criteria within the semantic reference framework, a process manifold representing the ideal state set is generated by performing data clustering and surface fitting logic. The aim is to connect discrete process nodes in the process semantic graph into a continuous mathematical trajectory, thereby achieving precise quantification of the trajectory deviation throughout the entire processing process. Specifically, a large number of ideal processing batches judged to be of acceptable quality are obtained, and corresponding feature respiration spectra are generated for these ideal processing batches, defined as ideal feature respiration spectra. A contrastive language image pre-training model is invoked. This model includes a visual encoding branch for image feature extraction and a text encoding branch for text feature extraction, with the text encoding branch using a Transformer network model. A multi-dimensional feature space is opened using this model; this multi-dimensional feature space is used to unify a high-dimensional logical mapping field carrying five feature items: visual data, thermal field data, acoustic spectrum data, load data, and chemical fingerprint data. The ideal feature breath spectrum is input into the visual encoding branch of the contrastive language image pre-training model. Texture features of the ideal feature breath spectrum are extracted through a residual convolutional neural network or visual network within the visual encoding branch, outputting an ideal feature vector. This ideal feature vector is a high-dimensional numerical vector representing the instantaneous features of the standard processing state, formed by nonlinearly compressing the multidimensional physical evolution attributes contained in the ideal feature breath spectrum. The ideal feature vectors corresponding to each ideal feature breath spectrum are combined into an ideal feature vector set. Simultaneously, the process text in the process semantic graph is input into the text encoding branch of the contrastive language image pre-training model, outputting a semantic vector. This semantic vector is a set of numerical scalars with unique pose coordinates in a multidimensional feature space, transforming the ideal characteristic descriptions of the five feature items contained in the process text. This achieves joint alignment between the image space and the text space, placing the ideal feature vector and the semantic vector in the same multidimensional feature space with physical causal logic.The ideal feature vector set is reconstructed using an equidistant mapping algorithm. Specifically, the local neighborhood topological relationships between ideal feature vectors in the ideal feature vector set are extracted. Specifically, the Euclidean distance between any two ideal feature vectors in the ideal feature vector set is calculated, and the Euclidean distance is compared with a preset similarity threshold. If the Euclidean distance is less than the similarity threshold, the two ideal feature vectors are determined to be local neighbors, and a topological connection relationship is established. The similarity threshold is set based on the average Euclidean distance between ideal feature vectors at adjacent sampling times in the ideal batch. For example, 1.2 times the average value of the Euclidean distance is taken. An M×M all-zero matrix is ​​initialized, where M is the total number of feature vectors in the ideal feature vector set, and each ideal feature vector is numbered and indexed according to the process time sequence. The determination results of any two ideal feature vectors are traversed. If the condition of being local neighbors is met, the element values ​​of the corresponding two ideal feature vectors in the all-zero matrix are changed from 0 to 1. The symmetric matrix generated after the traversal is completed is the adjacency matrix. Then, spatial interpolation is used to map the ideal feature vector set into a low-dimensional embedding space, generating a smooth and continuous low-dimensional subspace image, defined as the process manifold. The process manifold is the unique evolutionary trajectory formed by all ideal processing states conforming to the process text in the multidimensional feature space. See also. Figure 4 This is a mathematical topological diagram of a process manifold provided in an embodiment of the present invention. The diagram illustrates the multidimensional feature space generated by the contrastive language image pre-training model. The X, Y, and Z axes of the multidimensional feature space schematically represent the high-dimensional feature distribution within it. The white dashed line traversing the multidimensional feature space represents the ideal evolution trajectory, which is a feature line representing the continuous migration trajectory of a standard processing state, formed by connecting a set of ideal feature vectors in chronological order. The green solid dots on the ideal evolution trajectory represent semantic vectors, indicating the specific ideal process pose corresponding to the process text on the trajectory. The light blue area shows the geometric surface, which is the process manifold. This manifold is a smooth and continuous low-dimensional subspace image fitted in the multidimensional feature space after performing an isometric mapping algorithm on the set of ideal feature vectors, including the ideal evolution trajectory. This digital reference benchmark, which spatially aligns discrete process text with continuous physical signals, provides a precise mathematical carrier for subsequent processing.

[0088] Step S22: Perform vector decomposition using the process manifold to obtain the semantic orthogonal decomposition distance characterizing the essential properties of the processing process and the progress deviation.

[0089] After generating the process manifold, in order to deeply align the characteristic breathing spectrum with the ideal evolution trajectory and accurately isolate anomalous properties, spatial projection and vector decomposition logic are executed to calculate the semantic orthogonal decomposition distance corresponding to the real-time processing state. The aim is to deconstruct the abstract deviations in the high-dimensional feature space into orthogonal physical components that can independently characterize process progress deviations and intrinsic quality degradation, thereby solving the semantic aliasing defects of traditional detection methods when facing the nonlinear evolution of processing, and providing a quantitative basis for subsequent steps.

[0090] Specifically, using a visual encoding branch, the currently generated feature respiration spectrum is mapped to a multi-dimensional feature space, outputting a real-time feature vector representing the processing state at the current sampling moment. This real-time feature vector is a high-dimensional numerical coordinate located in the multi-dimensional feature space, formed by feature extraction of the real-time amplitudes of five feature items at the current sampling moment through the visual encoding branch. Nearest neighbor search is used to locate the coordinate point with the smallest Euclidean distance to the real-time feature vector on the continuous surface of the process manifold, and this coordinate point is defined as the ideal projection point. The ideal projection point represents the reference benchmark that is closest to the actual state at the current sampling moment among all ideal states conforming to the process specifications. Based on the real-time feature vector, the ideal projection point, and the semantic vector, the semantic orthogonal decomposition distance is calculated. This semantic orthogonal decomposition distance is composed of the quality degradation degree, reflecting deviations from essential attributes, and the process deviation degree, reflecting deviations from process progress. Among them, the quality deterioration degree characterizes the offset intensity of the real-time feature vector perpendicular to the process manifold surface, and is used to quantitatively determine whether the medicinal material has undergone irreversible qualitative changes that deviate from the process text. It is obtained by calculating the Euclidean distance between the real-time feature vector and the ideal projection point. The process deviation degree characterizes the degree of advancement or lag of the processing progress at the current sampling moment relative to the preset stage target on the ideal evolution trajectory, and is used to quantitatively determine the achievement rate of the processing temperature. The process deviation degree is obtained by calculating the geodesic distance from the ideal projection point on the process manifold surface to the pose to which the semantic vector belongs. The stage target is the ideal state pose corresponding to the current process node locked by the semantic vector in the multi-dimensional feature space. The setting basis is to retrieve the process text corresponding to the current process stage in the process semantic map, and use the text encoding branch to convert the process text into a semantic vector, thereby establishing a dynamic time benchmark in the multi-dimensional feature space in real time for evaluating the achievement rate of the processing temperature.

[0091] Step S20, by constructing a process semantic graph, generating a process manifold, and calculating the semantic orthogonal decomposition distance, solves the technical problem that traditional anomaly detection algorithms can only identify isolated features and cannot understand the logical continuity of process evolution, and that the diagnostic results lack quantitative guidance. This achieves the digital reconstruction of experience in traditional Chinese medicine processing techniques and the precise quantification of deviation. Specifically, the process semantic graph establishes a digital reference system containing expert knowledge for feature identification, overcoming the technical deficiency of traditional algorithms that lack interpretability of pharmacological mechanisms; the process manifold solidifies the ideal processing evolution trajectory in the feature space, solving the problem of blind spots in monitoring transition states between discrete nodes; and the semantic orthogonal decomposition distance deconstructs high-dimensional deviations into quality degradation degree and process deviation degree, providing a mathematical benchmark with physical causal interpretation for the qualitative and localization of abnormal states.

[0092] Step S30: Construct a virtual correction force field based on the semantic orthogonal decomposition distance, and decompose the virtual correction force field to obtain active intervention commands for physical control parameters. Drive risk inference based on active intervention commands to generate an abnormal diagnosis report characterizing the cause of abnormalities in the processing process.

[0093] Further, step S30 includes:

[0094] Step S31: Construct a virtual correction force field based on the semantic orthogonal decomposition distance, and decompose the virtual correction force field to obtain active intervention commands for physical control parameters.

[0095] After obtaining the semantic orthogonal decomposition distance, in order to transform the abstract high-dimensional geometric deviation into physical control kinetic energy that can drive the production equipment back to the ideal evolution trajectory and realize automated closed-loop correction of abnormal processing states, a virtual correction force field is constructed and active intervention commands are generated by executing potential energy mapping and vector decomposition logic. The aim is to establish a dynamic control framework in a multi-dimensional feature space that can guide real-time feature vectors to regress to the process manifold, thereby overcoming the technical deficiency of traditional monitoring systems that can only provide abnormal alarms but cannot provide quantitative correction operation suggestions.

[0096] The virtual correction force field refers to a vector field in a multidimensional feature space, with the process manifold as the zero potential energy reference and the quality degradation degree as the potential energy gradient. The zero potential energy reference refers to the surface of the process manifold in the multidimensional feature space, composed of all ideal eigenvectors, which does not require correction intervention. The potential energy gradient refers to the slope of energy evolution determined by the quality degradation degree of the real-time eigenvectors deviating from the zero potential energy reference. Specifically, the construction process of the virtual correction force field is as follows: based on the magnitude of the quality degradation degree, a virtual correction force pointing towards the process manifold is calculated; the direction of the virtual correction force is the unit vector direction from the real-time eigenvectors to the ideal projection point, and its magnitude is obtained by performing a weighted multiplication operation between the unit vector direction and the quality degradation degree. Using a preset parameter mapping weight matrix, the virtual correction force is decomposed into the physical control parameter dimension composed of component heating power, stirring speed, and auxiliary material addition amount, resulting in a preliminary set of intervention components. The heating power is a scalar value of electrical power used to adjust the heat input intensity of the cooking pot; the stirring speed is a frequency value used to adjust the number of rotations of the stirring paddle per unit time; the auxiliary material addition amount is a flow control value used to adjust the flow rate of auxiliary materials sprayed or added into the cooking pot per unit time. The parameter mapping weight matrix is ​​a coefficient matrix used to quantify the contribution ratio of the dimensional deviation components of the real-time feature vector in each coordinate dimension of the multi-dimensional feature space to different actuators; the setting is based on: statistically analyzing the changes of each feature item in historical data, mapping and setting the sensitivity influence factors of heating power, stirring speed, and auxiliary material addition amount, to ensure that the virtual correction force can be decomposed into physical control variables.

[0097] Dynamic gain correction is performed on the intervention component set based on the process deviation, and the final active intervention command is generated by combining them. The specific process is as follows: The positive or negative attribute of the process deviation is determined; if the process deviation is positive, the processing progress at the current sampling time is determined to be lagging, and a power compensation value for increasing thermal energy input is added to the intervention component set; if the process deviation is negative, the processing progress at the current sampling time is determined to be ahead, and a power reduction value for reducing thermal energy input is added to the intervention component set. The intervention component set after dynamic gain correction is encapsulated to generate the active intervention command. The active intervention command is defined as a set of structured control commands containing target heating power values, target stirring speed values, and target auxiliary material addition values. The aim is to pull the deviated real-time feature vector back to the edge of the process manifold by adjusting the output parameters of the physical actuators without interrupting production.

[0098] Step S32: Based on the risk simulation driven by the active intervention command, generate an abnormal diagnosis report that characterizes the causes of abnormalities in the processing process.

[0099] Upon receiving an active intervention instruction, in order to achieve proactive early warning of processing anomalies and improve the interpretability of automated monitoring for the causes of anomalies, a structured anomaly diagnosis report is generated by executing time-series simulation and path matching logic. The aim is to transform the geometric offset logic in the multi-dimensional feature space into a natural language description that conforms to the logic of traditional Chinese medicine processing, thereby reducing the reliance on manual supervision and ensuring the integrity of quality traceability.

[0100] Specifically, when the quality degradation exceeds a preset risk threshold, a risk simulation for the processing process is initiated. The risk threshold is the quality deviation threshold value for forward evolution prediction. Determined using the local radius of curvature of the process manifold at the corresponding current process node, the risk simulation refers to simulating the future evolution path of real-time feature vectors in a multi-dimensional feature space without active intervention. The specific implementation process is as follows: A long short-term memory (LSTM) neural network is invoked. A large amount of multimodal real-time dataset of historical failed batches is acquired, and a feature breathing spectrum sequence corresponding to the historical failed batches is generated. The feature breathing spectrum sequence is mapped to a corresponding high-dimensional feature vector sequence using a visual encoding branch, and this sequence is used as training samples input into the LTM neural network. The LTM neural network extracts the nonlinear dynamic features of the high-dimensional feature vector sequence's evolution over time, establishing a mapping relationship between the feature pose at the current sampling moment and the feature pose at future moments. The real-time feature vector is input into the LTM neural network for iterative prediction to obtain the corresponding predicted evolution trajectory.

[0101] The predicted evolution trajectory is compared with a pre-defined risk endpoint region in a multi-dimensional feature space for collision detection. The construction process of the risk endpoint region involves: extracting ideal feature vectors generated from mapping historical failed batches, defining them as anomalous feature vectors; using clustering analysis algorithms to determine the distribution center and diffusion radius of these anomalous feature vectors in the multi-dimensional feature space, thereby defining the logical warning range characterizing irreversible failures in the processing. Simultaneously, the failure cause field associated with each historical failed batch is obtained, and this field is used as an anomaly label and logically bound to the corresponding risk endpoint region. The anomaly labels include scorched labels, undercooked labels, and labels indicating excessive loss of effective components, used to describe the root causes of medicinal material failure. If the predicted evolution trajectory enters a risk endpoint region, the corresponding anomaly label is automatically extracted and, combined with proactive intervention instructions, a structured anomaly diagnosis report is generated. The structured anomaly diagnostic report specifically includes a description of trait deviations, expected risk consequences, and process adjustment plans. The trait deviation description is a textual representation of the physical manifestations of the medicinal material deviating from its ideal evolutionary trajectory by mapping the degree of quality deterioration to the process deviation value. The expected risk consequences are a forward-looking judgment of the future quality deterioration trend of the medicinal material based on the depth to which the predicted evolutionary trajectory falls into the risk endpoint region. The process adjustment plan maps proactive intervention instructions into a set of natural language instructions that can be executed by operators. The aim is to provide a quality traceability basis with causal logic support for the production process of each batch of medicinal materials.

[0102] Step S30, by constructing a virtual correction force field, performing risk simulation, and generating a structured anomaly diagnosis report, solves the technical problems of traditional monitoring systems that can only provide anomaly alarms but cannot provide quantitative corrective action suggestions, and whose feedback results lag behind the occurrence of quality deterioration and lack forward-looking prediction. This achieves an intelligent leap from anomaly diagnosis in traditional Chinese medicine processing to proactive intervention and forward-looking early warning. Specifically, the virtual correction force field transforms abstract geometric pose offsets into dynamic components that can drive physical actuators, solving the problem of correction decisions relying on human experience; risk simulation overcomes the technical bottleneck of monitoring after the fact and loss of control beforehand; and the structured anomaly diagnosis report provides quality traceability evidence with causal logic support for the production process of each batch of medicinal materials, significantly improving the inherent safety of modern traditional Chinese medicine manufacturing.

[0103] Example 2

[0104] This embodiment, based on Embodiment 1, provides a large-scale feature classification system for detecting anomalies in the processing of traditional Chinese medicine, such as... Figure 5 As shown, it includes:

[0105] Feature Reconstruction Module: This module is used to acquire a multimodal real-time dataset that characterizes the co-evolution of the physicochemical properties of medicinal materials during the processing of traditional Chinese medicine. Parallel analysis is performed on the multimodal real-time dataset to obtain information emergence points that predict qualitative changes in the properties of medicinal materials. Using the information emergence points as the central origin, the multimodal real-time dataset is reconstructed into a feature respiration spectrum.

[0106] The trajectory quantization module is used to construct a process semantic map by calling the feature breathing spectrum and generate a process manifold that represents the continuous evolution trajectory of the ideal processing state. The process manifold is used to perform vector decomposition to obtain the semantic orthogonal decomposition distance that represents the essential characteristics and progress deviation of the processing process.

[0107] Correction and Diagnosis Module: This module is used to construct a virtual correction force field based on the semantic orthogonal decomposition distance, and decompose the virtual correction force field to obtain active intervention commands for physical control parameters. Based on the active intervention commands, risk inference is driven to generate an anomaly diagnosis report characterizing the causes of abnormalities in the processing process.

[0108] In the feature reconstruction module, the process of acquiring a multimodal real-time dataset representing the co-evolution of the physicochemical properties of medicinal materials during the processing of traditional Chinese medicine, performing parallel analysis on the multimodal real-time dataset to obtain information emergence points that predict qualitative changes in the properties of medicinal materials, and reconstructing the multimodal real-time dataset into a feature respiration spectrum with the information emergence points as the central origin includes:

[0109] Step S11: Obtain a multimodal real-time dataset that characterizes the co-evolution of the physicochemical properties of medicinal materials during the processing of traditional Chinese medicine.

[0110] Step S12: Perform parallel analysis on the multimodal real-time dataset to obtain information emergence points that predict qualitative changes in the properties of medicinal materials.

[0111] Step S13: Using the information emergence point as the central origin, reconstruct the multimodal real-time dataset into a feature breathing spectrum.

[0112] In the trajectory quantization module, the process semantic map is constructed by calling the feature breathing spectrum, and a process manifold representing the continuous evolution trajectory of the ideal processing state is generated. Vector decomposition is performed using the process manifold to obtain the semantic orthogonal decomposition distance representing the essential characteristics and progress deviation of the processing process, including:

[0113] Step S21: Invoke the feature breathing spectrum to construct the process semantic map and generate the process manifold representing the continuous evolution trajectory of the ideal processing state.

[0114] Step S22: Perform vector decomposition using the process manifold to obtain the semantic orthogonal decomposition distance characterizing the essential properties of the processing process and the progress deviation.

[0115] In the deviation correction and diagnosis module, the virtual deviation correction force field is constructed based on the semantic orthogonal decomposition distance, and the virtual deviation correction force field is decomposed to obtain active intervention commands for physical control parameters. Based on the active intervention commands, risk inference is driven to generate an anomaly diagnosis report characterizing the causes of abnormalities in the processing process, including:

[0116] Step S31: Construct a virtual correction force field based on the semantic orthogonal decomposition distance, and decompose the virtual correction force field to obtain active intervention commands for physical control parameters.

[0117] Step S32: Based on the risk simulation driven by the active intervention command, generate an abnormal diagnosis report that characterizes the causes of abnormalities in the processing process.

[0118] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0119] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A large-scale feature classification algorithm for detecting anomalies in the processing of traditional Chinese medicine, characterized in that, The algorithm includes: A multimodal real-time dataset representing the co-evolution of the physicochemical properties of medicinal materials during the processing of traditional Chinese medicine was obtained. Parallel analysis was performed on the multimodal real-time dataset to obtain information emergence points that predict qualitative changes in the properties of medicinal materials. The multimodal real-time dataset was reconstructed into a feature respiration spectrum with the information emergence points as the central origin. The process semantic map is constructed by calling the feature breathing spectrum, and the process manifold representing the continuous evolution trajectory of the ideal processing state is generated. Vector decomposition is performed using the process manifold to obtain the semantic orthogonal decomposition distance representing the essential characteristics and progress deviation of the processing process. A virtual correction force field is constructed based on the semantic orthogonal decomposition distance, and the virtual correction force field is decomposed to obtain active intervention commands for physical control parameters. Based on the active intervention commands, risk inference is driven to generate an anomaly diagnosis report characterizing the causes of abnormalities in the processing process.

2. The large-scale feature classification algorithm for detecting anomalies in the traditional Chinese medicine processing process as described in claim 1, characterized in that, The method for obtaining the multimodal real-time dataset includes: Raw visible light images were acquired using an industrial camera, from which the surface chromaticity values ​​of the medicinal materials were extracted and defined as visual data. Raw infrared thermal images were acquired using a thermal imaging camera, from which the material uniform temperature values ​​were extracted and defined as thermal field data. Instantaneous power spectral density, waveform kurtosis index, and characteristic absorption peak intensity were extracted using a microphone, Hall current sensor, and pushbroom camera and defined as acoustic spectrum data, load data, and chemical fingerprint data, respectively. The global reference clock is invoked to perform timestamp association and synchronization on the feature items, which include visual data, thermal field data, acoustic spectrum data, load data, and chemical fingerprint data. A baseline time axis is established using chemical fingerprint data. Visual data, thermal field data, acoustic spectrum data, and load data are mapped onto the baseline time axis using the nearest neighbor interpolation method to obtain the multimodal real-time dataset with instantaneous physical semantic alignment.

3. The large-scale feature classification algorithm for detecting anomalies in the traditional Chinese medicine processing process as described in claim 2, characterized in that, The information emergence points include: The first derivative time series of visual data, thermal field data, acoustic spectrum data, and load data are acquired. The absolute sum of the first derivative time series is calculated and weighted to obtain the modal internal variability. Extract the time-series feature values ​​of five feature items from the multimodal real-time dataset, calculate the real-time covariance values ​​between any two feature items and combine them to generate a real-time covariance matrix, and calculate the intermodal covariance divergence by comparing the real-time covariance matrix with the pre-stored benchmark covariance matrix. Set energy weights and collaborative weights, and perform multiplication operations with intramodal variability and intermodal covariance divergence respectively, and sum them to obtain the single-dimensional perceptual entropy value; The single-dimensional perceptual entropy values ​​are arranged in chronological order to form a perceptual entropy sequence. If the single-dimensional perceptual entropy value is greater than the dynamic feature threshold at multiple consecutive sampling times, it is marked as an information emergence point.

4. The large-scale feature classification algorithm for detecting anomalies in the traditional Chinese medicine processing process as described in claim 3, characterized in that, The method for obtaining the characteristic respiratory spectrum includes: Extract the numerical sequence before and after the information emergence point for a preset time period and normalize it to obtain the temporal components of each feature item; A two-dimensional feature space containing an evolution time axis and a causal feature axis is constructed. Based on the order of the physical response of the five feature items in the processing technology, decreasing values ​​are assigned sequentially from the source end to the end end, which are defined as weight scalar values. The causal energy level weights are composed of each weight scalar value. Based on the order of the weight scalar values ​​in the causal energy level weight from large to small, the causal feature axis is divided into five feature mapping rows in the vertical direction, and the time-series components corresponding to the thermal field data, visual data, load data, acoustic spectrum data and chemical fingerprint data are sequentially locked into the corresponding feature mapping rows. The amplitude of each time-series component is mapped to the pixel brightness value of the corresponding feature mapping row coordinate point, and the combination is used to generate the feature breathing spectrum.

5. The large-scale feature classification algorithm for detecting anomalies in the traditional Chinese medicine processing process as described in claim 4, characterized in that, The method for obtaining the process manifold includes: Construct a process semantic graph consisting of process nodes representing key stages of processing and logical edges representing time evolution constraints, and use process text to parameterize the ideal physical state of each feature item to associate process nodes. A large number of ideal processing batches that are judged to be qualified are obtained and corresponding characteristic respiration spectra are generated, which are defined as ideal characteristic respiration spectra. Ideal feature vectors are extracted using a contrastive language image pre-trained model to form an ideal feature vector set, and the process text is converted into semantic vectors. Extract the local neighborhood topological relationships of the ideal feature vector set, fit and generate a smooth and continuous low-dimensional subspace image to obtain the process manifold.

6. The large-scale feature classification algorithm for detecting abnormalities in the traditional Chinese medicine processing process as described in claim 5, characterized in that, The method for obtaining the semantic orthogonal decomposition distance includes: Map the currently generated real-time feature respiration spectrum to a real-time feature vector, and locate the ideal projection point on the process manifold that has the smallest Euclidean distance from the real-time feature vector; Calculate the Euclidean distance between the real-time feature vector and the ideal projection point to obtain the quality degradation degree that represents the deviation of the essential attribute; Obtain the semantic vector corresponding to the current process node, calculate the geodesic distance from the ideal projection point to the position of the semantic vector on the process manifold, and obtain the process deviation that characterizes the achievement rate of the processing temperature. The semantic orthogonal decomposition distance is composed of the degree of quality degradation and the degree of process deviation.

7. The large-scale feature classification algorithm for detecting anomalies in the traditional Chinese medicine processing process as described in claim 6, characterized in that, The active intervention instructions include: A virtual correction force field is constructed using the process manifold as the zero potential energy reference and the quality degradation degree as the potential energy gradient. Calculate the virtual correction force pointing from the real-time feature vector to the ideal projection point, and decompose the virtual correction force into a set of intervention components in the dimensions of heating power, stirring speed and amount of additives; Dynamic gain correction is performed on the intervention component set based on the positive or negative attribute of the process deviation: if the process deviation is positive, the power compensation value is increased; if the process deviation is negative, the power reduction value is increased. The set of intervention components after dynamic gain correction is encapsulated into an active intervention command consisting of the target heating power value, the target stirring speed value, and the target amount of auxiliary material added.

8. The large-scale feature classification algorithm for detecting anomalies in the traditional Chinese medicine processing process as described in claim 7, characterized in that, The risk simulation includes: Obtain the characteristic respiratory spectrum sequence of historical failed batches, and map the characteristic respiratory spectrum sequence into a high-dimensional feature vector sequence; The Long Short-Term Memory (LSTM) neural network is invoked, and real-time feature vectors are input into the LSM to simulate the predicted evolution trajectory under conditions where no active intervention commands are executed. Extract the feature vector generated by mapping historical failed batches, define it as an abnormal feature vector, perform cluster analysis on the abnormal feature vector, delineate the risk endpoint region and bind anomaly labels; By using the abnormal feature vectors of historical failed batches, risk endpoint regions are delineated in a multidimensional feature space through clustering, and abnormal labels including scorched labels, half-cooked labels, and labels indicating excessive loss of effective ingredients are bound. Collision detection is performed between the predicted evolution trajectory and the risk endpoint region. If the trajectory enters the risk endpoint region, the corresponding anomaly label is automatically extracted.

9. The large-scale feature classification algorithm for detecting anomalies in the traditional Chinese medicine processing process as described in claim 8, characterized in that, The methods for obtaining the abnormal diagnostic report include: Based on the values ​​of quality degradation degree and process deviation degree, a description of trait deviation is generated; Based on the depth and anomaly labels of the predicted evolution trajectory falling into the risk endpoint area, the expected risk consequences are generated. Active intervention commands are mapped into natural language command sets to generate process adjustment schemes; By combining descriptions of trait deviations, expected risk consequences, and process adjustment schemes, a structured anomaly diagnosis report is output.

10. A large-scale feature classification system for detecting anomalies in the processing of traditional Chinese medicine, used to implement the large-scale feature classification algorithm for detecting anomalies in the processing of traditional Chinese medicine as described in any one of claims 1-9, characterized in that, The system includes: Feature Reconstruction Module: This module is used to acquire a multimodal real-time dataset that characterizes the co-evolution of the physicochemical properties of medicinal materials during the processing of traditional Chinese medicine. Parallel analysis is performed on the multimodal real-time dataset to obtain information emergence points that predict qualitative changes in the properties of medicinal materials. Using the information emergence points as the central origin, the multimodal real-time dataset is reconstructed into a feature respiration spectrum. The trajectory quantization module is used to construct a process semantic map by calling the feature breathing spectrum and generate a process manifold that represents the continuous evolution trajectory of the ideal processing state. Vector decomposition is performed using the process manifold to obtain the semantic orthogonal decomposition distance that represents the essential characteristics and progress deviation of the processing process. Correction and Diagnosis Module: This module is used to construct a virtual correction force field based on the semantic orthogonal decomposition distance, and decompose the virtual correction force field to obtain active intervention commands for physical control parameters. Based on the active intervention commands, risk inference is driven to generate an anomaly diagnosis report that characterizes the causes of abnormalities in the processing process.

Images