GC-MS characteristic fingerprint spectrum-based shaddock peel sauce quality monitoring system and method

Abstract

The invention discloses a GC-MS (Gas Chromatography-Mass Spectrometer) characteristic fingerprint spectrum-based shaddock peel sauce quality monitoring system and method, belongs to the technical field of shaddock peel sauce processing, and particularly relates to production quality monitoring and control of (Lixizhou) shaddock peel sauce. The problems that in an existing GC-MS combined technology, feature marker screening is lack of intelligence, the map stability is limited by environment interference, the system integration degree is low, and the real-time performance is insufficient are solved, and intelligent quality control over the whole shaddock peel sauce production chain is achieved. The system comprises a multi-source data fusion control module; and the multi-source data fusion control module is used for generating a process regulation and control instruction according to the characteristic peaks in the dynamic fingerprint spectrum model, the sensor data after drift correction and the equipment state parameters. The GC-MS characteristic fingerprint spectrum-based shaddock peel sauce quality monitoring system and method are suitable for the production quality monitoring and control of shaddock peel sauce (especially Sizhou shaddock peel sauce).

Description

Technical Field

[0001] The present invention relates to the technical field of pomelo peel jam processing, and particularly to the production quality monitoring and control of (Sizhou) pomelo peel jam. Background Art

[0002] Pomelo peel jam is a traditional fermented food, and its quality is affected by many factors such as raw material origin, processing technology, and storage conditions. The traditional quality control of pomelo peel jam mainly includes sensory evaluation (color, smell, texture) and physical and chemical index detection (water activity, total acid, sugar content). The following problems exist:

[0003] ① Limitations of sensory evaluation: Early production relied on empirical manual sensory evaluation, which was highly subjective and the standards could not be quantified, resulting in poor product consistency. Although some enterprises introduced equipment such as color difference meters and texture analyzers to assist in evaluation, the dynamic changes of flavor substances and functional components could not be analyzed.

[0004] ② Bottlenecks in conventional instrument detection: HPLC, ultraviolet spectrophotometry, etc. are used to detect single indicators such as naringin and citric acid. Although the detection accuracy of some parameters has been improved, the volatile flavor components (such as terpenoids and esters) in complex matrices cannot be comprehensively characterized, and the detection cycle is long, making it difficult to meet the requirements of continuous production.

[0005] In recent years, gas chromatography-mass spectrometry (GC-MS) technology has been tried for the analysis of volatile components in fermented foods. For example, patent document CN105548388A discloses a GC-MS multi-ion parameter detection method for acrylamide in foods. By using the internal standard method for bromination derivation treatment and conversion treatment of acrylamide in the sample, the accuracy of the calculation results is improved. However, this method only targets specific production links, does not form a full-chain monitoring system covering raw materials - processing - finished products, and the data processing relies on manual integration, and the problem of chromatogram drift between batches has not been solved.

[0006] The current GC-MS coupling technology still has the following defects and deficiencies:

[0007] (1) Lack of intelligence in screening characteristic markers: Existing GC-MS quality control methods mostly use fixed compounds (such as naringin and limonene) as indicators. However, key flavor substances such as furans and aldehydes and ketones generated during the fermentation of Sizhou pomelo peel jam have dynamic change characteristics. The traditional threshold method or PCA dimensionality reduction cannot effectively distinguish the subtle differences caused by process deviations, resulting in insufficient accuracy in screening key quality-related peaks.

[0008] (2) The stability of chromatograms is restricted by environmental interference: The existing GC-MS fingerprint map modeling does not consider the seasonal differences in raw material origin and the detection drift of equipment. Directly using static similarity algorithms (such as Euclidean distance) will cause misjudgment, and the false positive alarm rate in actual production exceeds 20%.

[0009] (3) Low system integration and insufficient real-time performance: Traditional quality control systems separate GC-MS detection from production line control, and data analysis relies on offline software (such as ChromaTOF). It takes more than 2 hours from sampling to feedback, and it is impossible to adjust key parameters such as sterilization temperature and mash turning frequency in real time.

[0010] In addition, although patent document CN119164913A proposes an online visible-near infrared fermentation liquid food production process monitoring system, it is only suitable for the detection of simple indicators such as the moisture content of liquid food, and its detection limit for trace characteristic substances (<1μg / g) is insufficient. Summary of the invention

[0011] The present invention proposes a pomelo paste quality monitoring system and method based on GC-MS characteristic fingerprint spectrum, which solves the problems existing in the existing GC-MS combination technology, such as lack of intelligence in characteristic marker screening, spectrum stability being restricted by environmental interference, and low system integration and insufficient real-time performance, thereby realizing intelligent quality control of the entire chain of pomelo paste production.

[0012] The grapefruit peel sauce quality monitoring system based on GC-MS characteristic fingerprint spectrum of the present invention comprises the following modules: GC-MS detection module, dynamic fingerprint spectrum modeling module, drift correction module, real-time analysis module, multi-source data fusion control module;

[0013] The GC-MS detection module is used to perform chemical analysis on the pomelo peel jam sample, separate the volatile components by gas chromatography, and perform qualitative and quantitative detection by mass spectrometry to obtain original GC-MS data;

[0014] The dynamic fingerprint modeling module is used to extract characteristic peaks from the original GC-MS data and construct a dynamic fingerprint model reflecting the product quality of the pomelo sauce sample;

[0015] The drift correction module is used to collect sensor data and equipment status parameters, and perform drift correction on the collected data to ensure data stability;

[0016] The multi-source data fusion control module is used to generate process control instructions based on characteristic peaks in the dynamic fingerprint spectrum model, drift-corrected sensor data and equipment status parameters.

[0017] Further, a preferred embodiment is provided, wherein the GC-MS detection module includes a sample pre-treatment unit, a chromatographic separation unit, a mass spectrometry detection unit and a data acquisition unit; wherein:

[0018] The sample pre-treatment unit is used for grinding, extracting (normal hexane ultrasound), filtering and loading into a sample injection bottle of the grapefruit peel jam sample;

[0019] The chromatographic separation unit is used to separate the volatile components in the pomelo peel jam sample through a DB-WAX chromatographic column;

[0020] The mass spectrometry detection unit is used to ionize and scan the separated volatile components;

[0021] The data acquisition unit is used to collect and store the original GC-MS data output by the mass spectrometry detection unit in real time.

[0022] Furthermore, a preferred embodiment is provided. The dynamic fingerprint map modeling module includes a data preprocessing unit, a characteristic peak screening unit, and a model training unit; wherein:

[0023] The data preprocessing unit is used to preprocess the original GC-MS data;

[0024] The characteristic peak screening unit is used to screen key peaks from the preprocessed data by using an improved random forest algorithm;

[0025] The model training unit is used to construct a multiple regression model as a dynamic fingerprint map model reflecting the product quality of the pomelo peel jam sample based on the screened key peaks and process parameters.

[0026] Furthermore, a preferred embodiment is provided. The drift correction module includes an environmental drift compensation unit;

[0027] The environmental drift compensation unit realizes double compensation for instrument fluctuations and raw material differences through the coupling mechanism of double internal standard correction and moving window similarity model, and improves the anti-interference ability of environmental drift.

[0028] Furthermore, a preferred embodiment is provided. The multi-source data fusion control module includes a heterogeneous data integration unit, an intelligent decision-making unit, and a production line regulation unit; wherein:

[0029] The heterogeneous data integration unit is used to obtain the characteristic peaks in the dynamic fingerprint map model, the sensor data after drift correction, and the device status parameters through the OPC-UA protocol;

[0030] The intelligent decision-making unit is used to predict the acid value change rate and flavor stability index based on the characteristic peaks in the dynamic fingerprint map model, the sensor data after drift correction, and the device status parameters by using a dual-channel network coupling LSTM and GRU;

[0031] The production line regulation unit is used to adjust the sterilization temperature and turning frequency in real time according to the prediction results of the acid value change rate and flavor stability index.

[0032] Further, a preferred embodiment is provided. When using the improved random forest algorithm to screen key peaks from the preprocessed data, the steps include:

[0033] Extract the peak area data of all samples as the data matrix from the preprocessed data, extract the classification labels of the samples as the label vector, and perform a standardization operation to obtain the standardized training data;

[0034] Construct an initial random forest model with a given number of decision trees, limit the maximum depth of a single tree, and use the Gini coefficient as the splitting criterion;

[0035] Train the initial random forest model with the standardized training data to obtain a trained random forest model;

[0036] Extract the feature importance scores from the trained random forest model;

[0037] Screen the chromatographic peaks with feature importance scores exceeding the dynamic threshold as candidate feature peaks; wherein, the dynamic threshold is: dynamic threshold = maximum value of the feature importance scores in the trained random forest model × 0.03;

[0038] Perform a fragment ion abundance ratio verification on the candidate feature peaks, eliminate interference peaks, and screen out key peaks.

[0039] Further, a preferred embodiment is provided for the coupling mechanism of the dual internal standard correction and the moving window similarity model:

[0040] Use the dual internal standard correction to correct the retention time and handle the time drift at the instrument hardware level;

[0041] Use the moving window similarity model to perform similarity matching on the peak areas to compensate for the peak area fluctuations caused by raw material batch differences.

[0042] Further, a preferred embodiment is provided. The dual-channel network coupled by LSTM and GRU includes an input layer, a hidden layer, and an output layer:

[0043] The input layer is used to input the feature peaks in the dynamic fingerprint model, the sensor data after drift correction, and the device status parameters;

[0044] The hidden layer includes LSTM units and GRU units; the LSTM units are used to process the time-series data in the data input by the input layer; the GRU units are used to extract the device status features in the data input by the input layer;

[0045] The output layer includes a fully connected network; the fully connected network is used to predict the acid value change rate and the flavor stability index according to the output of the hidden layer.

[0046] Further, a preferred embodiment is provided. According to the prediction results of the acid value change rate and the flavor stability index, the sterilization temperature and the turning frequency are adjusted in real time as follows:

[0047] The sterilization temperature is adjusted in real time using PID control:

[0048] When ΔAV > 0.5 mg / g·d-1, a temperature control instruction for the sterilization tank is triggered:

[0049]

[0050] where ΔAV is the acid value change rate; T new is the sterilization temperature after real-time adjustment; t set is the sterilization temperature set according to the sterilization process requirements; K p is the proportional gain coefficient; e(t) is the deviation of the acid value change rate; K i is the integral gain coefficient; K d is the derivative gain coefficient;

[0051] The turning frequency is adjusted in real time using variable frequency speed regulation:

[0052] The formula for adjusting the rotation speed of the turning machine:

[0053] Unit: Hz;

[0054] where FSI is the flavor stability index.

[0055] The present invention also proposes a method for monitoring the quality of pomelo peel jam based on the GC-MS characteristic fingerprint. The method is implemented using the quality monitoring system of pomelo peel jam based on the GC-MS characteristic fingerprint described in any one of the above; the method includes the following steps:

[0056] The GC-MS detection step: Chemically analyze the pomelo peel jam sample, separate the volatile components by gas chromatography, and perform qualitative and quantitative detection by mass spectrometry to obtain the original GC-MS data;

[0057] The dynamic fingerprint modeling step: Extract characteristic peaks from the original GC-MS data to construct a dynamic fingerprint model reflecting the product quality of the pomelo peel jam sample;

[0058] The drift correction step: Collect sensor data and equipment status parameters, and perform drift correction on the collected data to ensure data stability;

[0059] The multi-source data fusion control step: Generate a process control instruction according to the characteristic peaks in the dynamic fingerprint model, the sensor data after drift correction, and the equipment status parameters.

[0060] The present invention has the following beneficial effects:

[0061] 1. The quality monitoring system of pomelo peel jam based on GC-MS characteristic fingerprint spectrum according to the present invention solves the problem of lack of intelligence in screening characteristic markers through an improved random forest algorithm, and improves the accuracy of screening key quality correlation peaks when key flavor substances have dynamic change characteristics.

[0062] 2. The quality monitoring system of pomelo peel jam based on GC-MS characteristic fingerprint spectrum according to the present invention is realized through the coupling of double internal standard correction and moving window similarity model, which solves the problem that the stability of the spectrum is restricted by environmental interference. The coupling mechanism of double internal standard correction and moving window similarity model is that double internal standard correction is used to handle the time drift at the instrument hardware level (instrument fluctuation compensation), while the moving window similarity model is used to compensate for the peak area fluctuation caused by raw material batch differences (raw material difference compensation). The coupling method is to first correct the retention time through RRT, and then perform similarity matching on the peak area to double guarantee the consistency of the data.

[0063] Instrument fluctuation compensation: For example, the retention time drift caused by chromatographic column aging, calculate the relative retention time (RRT) through the double internal standard method to eliminate the influence of hardware fluctuation.

[0064] Raw material difference compensation: For example, when the content of a certain component in the new raw material batch decreases, perform peak area similarity evaluation through the moving window similarity model and call historical data for interpolation compensation.

[0065] 3. The quality monitoring system of pomelo peel jam based on GC-MS characteristic fingerprint spectrum according to the present invention integrates GC-MS detection and production line control through a multi-source data fusion control module. Without relying on offline software, key parameters such as sterilization temperature and turning frequency can be adjusted in real time, greatly saving the time from data collection to control feedback.

[0066] 4. The quality monitoring system of pomelo peel jam based on GC-MS characteristic fingerprint spectrum according to the present invention improves the sensitivity by optimizing the mass spectrometry parameters (such as solvent delay, scanning range) of the mass spectrometry detection unit and enhancing the algorithm of the data preprocessing unit (such as peak alignment of MS-DIAL). Especially the low-noise design and full-scan mode of the high-precision mass spectrometry detection unit (Agilent 5977B) are suitable for the detection of complex indicators in fermented foods such as pomelo peel jam, such as improving the detection limit of trace characteristic substances (<1 μg / g).

[0067] 5. The quality monitoring system of pomelo peel jam based on GC-MS characteristic fingerprint spectrum according to the present invention constructs a dual-channel network coupling LSTM and GRU, fuses GC-MS characteristic peaks and equipment operation parameters (such as the current harmonic of the turning motor), realizes the accurate traceability and early intervention of abnormal working conditions, and achieves real-time collaborative control of multi-modal data.

[0068] 6. The pomelo peel jam quality monitoring system based on the GC-MS characteristic fingerprint spectrum according to the present invention breaks through the limitations of the traditional threshold method or the linear dimensionality reduction of PCA through the improved random forest algorithm, significantly improving the accuracy of dynamic fingerprint modeling.

[0069] The pomelo peel jam quality monitoring system and method based on the GC-MS characteristic fingerprint spectrum according to the present invention are applicable to the production quality monitoring and control of pomelo peel jam (especially Sizhou pomelo peel jam). BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0071] Figure 1 It is a schematic structural diagram of the pomelo peel jam quality monitoring system based on the GC-MS characteristic fingerprint spectrum in an embodiment of the present invention;

[0072] Figure 2 It is a control flow chart of the pomelo peel jam quality monitoring system based on the GC-MS characteristic fingerprint spectrum in an embodiment of the present invention. SPECIFIC EMBODIMENTS

[0073] In order to more clearly describe the technical solutions and advantages of the present invention, the following will further describe the specific embodiments of the present invention in detail and completely in combination with the drawings. The following described embodiments are only some preferred embodiments of the present invention, rather than all the embodiments; the following described embodiments are intended to explain the present invention and should not be construed as a limitation of the present invention; the reasonable combination of the technical features defined in the various embodiments of the present invention, and all other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments of the present invention, fall within the scope of protection of the present invention.

[0074] Embodiment 1: A pomelo peel jam quality monitoring system based on the GC-MS characteristic fingerprint spectrum, the system includes the following modules: a GC-MS detection module, a dynamic fingerprint spectrum modeling module, a drift correction module, a real-time analysis module, and a multi-source data fusion control module;

[0075] The GC-MS detection module is used to chemically analyze the pomelo peel jam sample, separate volatile components by gas chromatography, and perform qualitative and quantitative detection by mass spectrometry to obtain the original GC-MS data;

[0076] The dynamic fingerprint map modeling module is used to extract characteristic peaks from the original GC-MS data and construct a dynamic fingerprint map model reflecting the product quality of the pomelo peel jam samples;

[0077] The drift correction module is used to collect sensor data and device status parameters, and perform drift correction on the collected data to ensure data stability;

[0078] The multi-source data fusion control module is used to generate process regulation instructions according to the characteristic peaks in the dynamic fingerprint map model, the sensor data after drift correction, and the device status parameters.

[0079] In this embodiment, the drift correction includes correcting instrument drift and raw material batch differences.

[0080] In this embodiment, the sensor data (or production line sensor data) comes from real-time monitoring of the production line, including temperature, humidity, pH, etc. Among them:

[0081] Temperature: It refers to the ambient temperature inside the fermentation tank, which directly affects the microbial activity and fermentation rate.

[0082] Humidity: It refers to the air humidity in the fermentation workshop, which affects the moisture balance of the raw materials.

[0083] pH: It refers to the acidity and alkalinity of the fermentation mash, which is crucial for enzyme activity and the formation of flavor substances.

[0084] Real-time monitoring object: These sensors monitor production equipment (such as fermentation tanks and sterilization tanks) and the workshop environment, rather than directly monitoring the pomelo peel jam samples themselves.

[0085] In this embodiment, the device status parameters come from the sensors of the production line execution devices (including the turning machine and the sterilization tank). For example, the drive motor current of the turning machine in the production line equipment (typical value: 5 - 20 A), and the opening percentage of the steam regulating valve of the sterilization tank (0 - 100%).

[0086] Embodiment 2: The GC-MS detection module includes a sample pretreatment unit, a chromatographic separation unit, a mass spectrometry detection unit, and a data acquisition unit; among them:

[0087] The sample pretreatment unit is used to grind, extract (n-hexane ultrasonic), filter, and load the pomelo peel jam samples into the injection vial;

[0088] The chromatographic separation unit is used to separate the volatile components in the pomelo peel jam samples through a DB-WAX chromatographic column;

[0089] The mass spectrometry detection unit is used to ionize and scan the separated volatile components;

[0090] The data acquisition unit is used to collect and store the original GC-MS data output by the mass spectrometry detection unit in real time.

[0091] In this embodiment, the sample pretreatment unit performs extraction by ultrasonic treatment with n-hexane.

[0092] In this embodiment, the volatile components for separation include terpene compounds and ester compounds.

[0093] In this embodiment, the original GC-MS data is mass spectrometry raw data, represented in.D format, including the mass-to-charge ratio and intensity information at each time point. More specifically:

[0094] The original GC-MS data in.D format is sourced from the mass spectrometry detection unit (such as generated by an Agilent 5977B mass spectrometer), stored as a.D file, and contains complete mass spectrometry data (m / z-intensity matrix).

[0095] In this embodiment, the mass spectrometry detection unit is implemented using an Agilent 5977B mass spectrometer with a low-noise design and full-scan mode.

[0096] In this embodiment, the mass spectrometry parameters of the mass spectrometry detection unit are optimized: the mass spectrometry scanning range is 35 - 450 m / z; the solvent delay is 3 min. Among them, the solvent delay is used to avoid solvent peak interference in mass spectrometry detection.

[0097] Embodiment 3: The dynamic fingerprint model building module includes a data preprocessing unit, a characteristic peak screening unit, and a model training unit; where:

[0098] The data preprocessing unit is used to preprocess the original GC-MS data;

[0099] The characteristic peak screening unit is used to screen key peaks from the preprocessed data using an improved random forest algorithm;

[0100] The model training unit is used to construct a multiple regression model based on the screened key peaks and process parameters as a dynamic fingerprint model reflecting the product quality of the pomelo jam sample.

[0101] In this embodiment, the preprocessing includes normalization, baseline correction, peak alignment (retention time tolerance ±0.2 min), and noise filtering.

[0102] In this embodiment, the sensitivity is improved by optimizing the mass spectrometry parameters (such as solvent delay, scanning range) of the mass spectrometry detection unit and enhancing the algorithm of the data preprocessing unit (such as peak alignment of MS-DIAL). In particular, the low-noise design and full-scan mode of the high-precision mass spectrometry detection unit (Agilent 5977B) are applicable to the detection of complex indicators in fermented foods such as pomelo jam, and the detection limit of trace characteristic substances (<1 μg / g) is improved.

[0103] In this embodiment, the key peak is the key mass correlation peak.

[0104] In this embodiment, the multivariate regression model is PLS-DA, which is constructed based on the characteristic peak area and the (processing) process parameters of the pomelo jam sample. Among them, the process parameters include three parts:

[0105] (1) Sterilization parameters: temperature (121 °C), time (15 min), steam pressure (0.3 MPa).

[0106] (2) Fermentation parameters: turning frequency (45 - 55 Hz), oxygen concentration (3 - 5%), stirring speed.

[0107] (3) Drying parameters: temperature gradient (50 - 70 °C), wind speed (2 - 5 m / s), humidity control.

[0108] Embodiment 4: The drift correction module includes an environmental drift compensation unit;

[0109] The environmental drift compensation unit realizes the dual compensation of instrument fluctuations and raw material differences through the coupling mechanism of double internal standard correction and moving window similarity model, and improves the anti-interference ability of environmental drift.

[0110] In this embodiment, the consistency of data is double-guaranteed through the coupling mechanism of double internal standard correction and moving window similarity model.

[0111] Embodiment 5: The multi-source data fusion control module includes a heterogeneous data integration unit, an intelligent decision-making unit, and a production line regulation unit; among them:

[0112] The heterogeneous data integration unit is used to obtain the characteristic peaks in the dynamic fingerprint map model, the sensor data after drift correction, and the equipment status parameters through the OPC-UA protocol;

[0113] The intelligent decision-making unit is used to predict the acid value change rate and flavor stability index based on the characteristic peaks in the dynamic fingerprint map model, the sensor data after drift correction, and the equipment status parameters, based on a dual-channel network that couples LSTM and GRU;

[0114] The production line control unit is used to adjust the sterilization temperature and the turning frequency in real time according to the prediction results of the acid value change rate and the flavor stability index.

[0115] In this embodiment, the dual-channel network coupling LSTM and GRU is an improvement of the existing network. On the basis of the existing network, a dual-channel structure (LSTM + GRU) is realized, and the ability to extract time-series features is improved.

[0116] In this embodiment, a dual-channel network coupling LSTM and GRU is constructed to fuse the GC-MS characteristic peaks and the equipment operation parameters (such as the harmonic current of the turning motor), so as to realize the accurate traceability and early intervention of abnormal working conditions and achieve real-time collaborative control of multi-modal data.

[0117] In this embodiment, the acid value change rate (ΔAV): refers to the change rate of the acid value of pomelo peel jam during the fermentation process.

[0118] In this embodiment, the flavor stability index (FSI): is an index used to comprehensively evaluate the stability of the flavor components of the product.

[0119] In this embodiment, GC-MS data is obtained from the characteristic peaks (specifically 15) in the dynamic fingerprint map model. The heterogeneous data integration unit integrates the GC-MS data (specifically 15-dimensional) obtained from the characteristic peaks through the OPC-UA protocol.

[0120] In this embodiment, the equipment state parameters are 6-dimensional equipment parameter data, including temperature, humidity, harmonic distortion rate of turning current, motor speed, steam valve opening, and oxygen concentration.

[0121] In this embodiment, the sterilization temperature is adjusted in real time by using PID adjustment control.

[0122] In this embodiment, the turning frequency is adjusted in real time by using variable frequency speed regulation.

[0123] In addition, in one embodiment, the multi-source data fusion control module further includes a multi-dimensional quality correlation analysis module;

[0124] The multi-dimensional quality correlation analysis module is used to establish a multi-variable model by using PLS-DA (partial least squares discriminant analysis) based on the characteristic peaks in the dynamic fingerprint map model, the sensor data after drift correction, the equipment state parameters, and the prediction results of the acid value change rate and the flavor stability index, quantify the influence of each component in the pomelo peel jam sample on the quality of the pomelo peel jam sample, and generate the ranking of the contribution degrees of key factors (such as the contribution degree of β-caryophyllene to ΔAV is 0.32).

[0125] In this embodiment, the multi-dimensional quality correlation analysis module is used for real-time process adjustment (setting the parameters that the PID controller or frequency converter prioritizes to adjust according to the contribution ranking of key factors: sterilization temperature or turning over frequency) and long-term optimization (generating monthly quality reports to guide the screening of raw material suppliers and iteration of process parameters).

[0126] The multi-dimensional quality correlation analysis module is used for real-time process adjustment, that is, according to the ranking of key factor contributions, the parameters that the PID controller or the frequency converter should adjust first are set: sterilization temperature or mash turning frequency, specifically:

[0127] During process adjustment, when an abnormality in the acid value change rate or flavor stability index (such as ΔAV exceeding the standard) is detected, it will not be directly used as a feedback signal to control the equipment, but will be:

[0128] For abnormal indicators (i.e., acid value change rate or flavor stability index), PLS-DA (partial least squares discriminant analysis) was used to establish a multivariate model to quantify the impact of each component in the pomelo peel sauce sample on the quality of the pomelo peel sauce sample and generate a ranking of key factor contributions (for example, it was found that the β-caryophyllene component in the pomelo peel sauce sample had the highest contribution to the change in acid value, which was 0.32);

[0129] The key factor contribution ranking is to rank the components in the grapefruit paste sample that have an impact on the change of the abnormal index according to the degree of influence;

[0130] Determine whether the main reason for the change of the top key factors in the ranking of key factor contribution is the sterilization temperature or the frequency of mash turning:

[0131] If it is judged that the main reason is the sterilization temperature, the PID controller is preferably used to adjust the sterilization temperature;

[0132] If it is determined that the main cause is the frequency of turning the mash, it is preferred to use a frequency converter to adjust the frequency of turning the mash.

[0133] Specifically, the key factor contribution ranking is used to set the priority and target value of the adjustment:

[0134] For example, when the multi-dimensional quality correlation analysis module determines that the sterilization temperature is the main cause of the current abnormality, a temperature adjustment instruction (such as "increase the temperature by 3°C") will be generated, and the PID controller will automatically adjust the steam valve opening according to this target value; if the frequency of turning the mash is determined to be the main cause, a frequency adjustment signal (such as "increase by 15Hz") will be sent to the frequency converter, and the frequency converter will accurately control the motor speed.

[0135] In this embodiment, an improved random forest algorithm and a multi-dimensional quality correlation analysis module are adopted to break through the limitations of the traditional threshold method or PCA's linear dimensionality reduction, significantly improving the accuracy of feature fingerprint dynamic modeling.

[0136] Embodiment 6: The improved random forest algorithm is used to screen key peaks from the preprocessed data, including:

[0137] From the preprocessed data, extract the peak area data of all samples as the data matrix, extract the classification labels of the samples as the label vector, and perform a standardization operation to obtain the standardized training data;

[0138] Construct an initial random forest model with a given number of decision trees, limit the maximum depth of a single tree, and use the Gini coefficient as the splitting criterion;

[0139] Use the standardized training data to train the initial random forest model to obtain a trained random forest model;

[0140] Extract the feature importance scores from the trained random forest model;

[0141] Screen the chromatographic peaks whose feature importance scores exceed the dynamic threshold as candidate feature peaks; among them, the dynamic threshold is: dynamic threshold = the maximum value of the feature importance scores in the trained random forest model × 0.03;

[0142] Perform a fragment ion abundance ratio verification on the candidate feature peaks, eliminate the interference peaks, and screen out the key peaks.

[0143] In this embodiment, during the process of using the improved random forest algorithm to screen key peaks from the preprocessed data, based on the cosine similarity threshold, it is judged whether it is necessary to re-execute the random forest algorithm to update the feature peaks:

[0144] Every time 10 batches of qualified (pomelo peel jam) samples are added, calculate the cosine similarity threshold between the new samples and the peak area vectors of the historical qualified samples through the moving window similarity model (MW-SIM):

[0145]

[0146] Among them, A i represents the peak area vector of the new pomelo peel jam sample, and B i represents the peak area vector of the historical pomelo peel jam sample;

[0147] When the standard deviation of the similarity within the window > 0.05, re-execute the random forest algorithm to update the feature peaks.

[0148] In this embodiment, the fragment ion abundance ratio verification is performed on the candidate feature peaks, the interference peaks are eliminated, and the key peaks are screened out as follows:

[0149] For the selected candidate characteristic peaks (such as m / z 136.2), extract the intensity values of their fragment ions (such as m / z 93.1, 77.0), calculate the abundance ratio (such as 93.1 / 77.0 = 2.1), and compare it with the historical data of the reference substance in the NIST library (such as the standard range 1.8 - 2.2):

[0150] If it exceeds the standard range of the historical data, it is determined as an interference peak and excluded.

[0151] The verification of the fragment ion abundance ratio is embedded in the branch of the improved random forest algorithm and is executed after dynamic threshold screening, ensuring the chemical credibility of the characteristic peaks through rule base matching.

[0152] In this embodiment, the data for fragment ion verification is the fragment ion intensity data exported from MS-DIAL (.csv file). During the characteristic peak screening process, the fragment ion intensity data is called, and the compliance of the abundance ratio is verified through the rule base (NIST 17 standard) (the standard abundance ratio range of the NIST library is built into the algorithm for dynamic comparison with the measured values).

[0153] In this embodiment, the improvements of the improved random forest algorithm are as follows:

[0154] (1) Dynamic threshold screening: The traditional method uses a fixed threshold (such as Gini > 0.01). In this embodiment, the screening threshold is automatically adjusted according to the feature importance distribution, that is, dynamically adjusted according to 3% of the maximum Gini coefficient: threshold = 0.03 × max(feature importance), avoiding manual setting deviation.

[0155] (2) Verification of fragment ion ratio: Perform (mass) fragment ion abundance ratio verification on candidate characteristic peaks to ensure the chemical credibility of the characteristic peaks.

[0156] Embodiment 7: The coupling mechanism of the dual internal standard correction and the moving window similarity model:

[0157] Use dual internal standard correction to correct the retention time and handle the time drift at the instrument hardware level;

[0158] Use the moving window similarity model to perform similarity matching on the peak areas to compensate for the peak area fluctuations caused by raw material batch differences.

[0159] In this embodiment, it is achieved through the coupling of dual internal standard calibration and moving window similarity model, which solves the problem that the stability of the chromatogram is restricted by environmental interference. The coupling mechanism of dual internal standard calibration and moving window similarity model is that dual internal standard calibration is used to handle the time drift at the instrument hardware level (instrument fluctuation compensation), while the moving window similarity model is used to compensate for the peak area fluctuation caused by raw material batch differences (raw material difference compensation). The coupling method is to first correct the retention time through RRT calibration and then perform similarity matching on the peak area to double guarantee the consistency of the data.

[0160] Instrument fluctuation compensation: For example, the retention time drift caused by chromatographic column aging is eliminated by calculating the relative retention time (RRT) through the dual internal standard method to eliminate the influence of hardware fluctuations.

[0161] Raw material difference compensation: For example, when the content of a certain component in the new raw material batch decreases, the peak area similarity evaluation is carried out through the moving window similarity model, and historical data is called for interpolation compensation.

[0162] In this embodiment, double standardization is adopted to overcome drift and anomalies:

[0163] RRT calculation (dual internal standard calibration): Eliminate the absolute time drift through the internal standard substance;

[0164] Similarity comparison (moving window similarity model): Detect relative distribution anomalies through cosine similarity.

[0165] In addition, in one embodiment, the retention time is corrected by dual internal standard calibration to handle the time drift at the instrument hardware level as follows:

[0166] Internal standard peak identification: Locate the characteristic peaks of 2-octanol and n-tetracosane in the original GC-MS data;

[0167] RRT conversion: Convert the absolute retention time of the target compound to the RRT value.

[0168] In this embodiment, 2-octanol (m / z 84.1) and n-tetracosane (m / z 338.5) are added as internal standards in each batch of pomelo peel jam samples.

[0169] In this embodiment, the target compound (or simply referred to as the target substance) refers to the key volatile compounds in pomelo peel jam, including: D-limonene (characteristic aroma component), β-caryophyllene (flavor stability marker), terpinyl acetate (fermentation process product).

[0170] The absolute retention time of D-limonene is 9.8 min.

[0171] In this embodiment, RRT, that is, the standardized relative retention time. The RRT value, for example, RRT = 372.

[0172] In this embodiment, the object of double internal standard method calibration is the instrument drift (such as chromatographic column aging, detector sensitivity change) of the GC-MS detection system (GC-MS detection module). Its function is to convert the absolute retention time into a relative value (RRT) by anchoring the retention times of 2-octanol and n-tetracosane, and eliminate the influence of equipment fluctuations.

[0173] In addition, in one embodiment, the RRT value (normalized relative retention time) is calculated according to the following formula to eliminate the influence of column efficiency attenuation:

[0174]

[0175] In this embodiment, the double internal standard method minimizes the influence of the hardware fluctuations of the instrument (such as chromatographic column aging, detector sensitivity change) and external environmental interference (temperature and humidity fluctuations) on the detection results through the dual mechanisms of time anchoring and environmental compensation, so as to ensure: comparability of cross-batch data: the RRT values of the same component detected at different times are highly consistent. At the same time, the input stability of the model is improved: standardized input data is provided for the subsequent dual-channel network coupled with LSTM and GRU.

[0176] In addition, in one embodiment, a moving window similarity model is used to perform similarity matching on the peak areas to compensate for the peak area fluctuations caused by raw material batch differences, as follows:

[0177] Reference window maintenance: A sliding window is used to retain the peak area data of the most recent given number of qualified pomelo jam samples from the original GC-MS data as the reference window; the reference window is updated according to the first-in, first-out principle;

[0178] Similarity calculation: Compare the cosine similarity between the peak area data of the current pomelo jam sample and the mean value of the reference window. If it is lower than the similarity threshold, it is determined to be abnormal;

[0179] Interpolation compensation: When it is determined to be abnormal, the original GC-MS data of a given number of batches of pomelo jam samples with the most similar cosine similarity to the mean value of the reference window in the historical database is called, and the peak area data of the current pomelo jam sample with abnormalities is generated by weighted average for interpolation compensation.

[0180] In this embodiment, if the sliding window size is 10, a sliding window is used to retain the peak area data (such as 15 characteristic peaks) of the most recent 10 pomelo jam samples from the original GC-MS data as the reference window.

[0181] In this embodiment, a dynamic reference benchmark is adopted: the statistical characteristics of the most recent given number of qualified pomelo jam samples are maintained through a sliding window, and the calculation of the reference window mean value reflects the stable state of the current instrument.

[0182] In this embodiment, the similarity threshold is set to 0.92.

[0183] In this embodiment, a real-time feedback mechanism is adopted:

[0184] The drift detection result (i.e., judging abnormality according to similarity calculation) directly affects the update of the reference window; the similarity threshold (or alarm threshold, for example, set to 0.92) can balance sensitivity and fault tolerance.

[0185] In this embodiment, the original GC-MS data of a given number of batches of pomelo peel jam samples that are most similar to the mean value of the reference window in the historical database is called, and the peak area data of the current pomelo peel jam sample with abnormality is generated by weighted average interpolation compensation:

[0186] The number of batches of pomelo peel jam samples with a given number is 5.

[0187] In this embodiment, through interpolation compensation data replacement, that is, using the interpolation result to replace the current abnormal data, the validity of the subsequent model input is ensured.

[0188] In addition, in one embodiment, a moving window similarity model is used to perform similarity matching on the peak area to compensate for the peak area fluctuation caused by the difference in raw material batches, and the following steps are further included:

[0189] Read the ion source temperature and GC-MS vacuum degree through the OPC-UA protocol; if the data of the ion source temperature and GC-MS vacuum degree is abnormal, it is judged that the instrument hardware state is abnormal, and the hardware maintenance instruction is triggered preferentially, and the similarity calculation result is frozen to prevent making wrong decisions based on distorted data.

[0190] In this embodiment, the abnormal judgment based on similarity calculation and the abnormal judgment of the instrument hardware state based on the ion source temperature and GC-MS vacuum degree constitute a double guarantee relationship. Specifically:

[0191] The abnormal judgment based on similarity calculation is to directly identify the abnormality caused by the fluctuation of raw material components or process deviation by calculating the cosine similarity (threshold 0.92) between the GC-MS characteristic peak area of the current batch of samples and the historical qualified data.

[0192] The abnormal judgment of the instrument hardware state based on the ion source temperature and GC-MS vacuum degree is an abnormal warning for the instrument hardware state. For example, when the GC-MS vacuum degree is too low, it will cause mass axis drift, indirectly resulting in distorted similarity calculation; the fluctuation of the ion source temperature will cause errors in the fragment ion abundance ratio and affect the authenticity of the peak area.

[0193] The connection between the two is as follows: when the hardware status of the instrument is abnormal (such as a sudden increase in the ion source temperature or a GC-MS vacuum leak), the system will first trigger a hardware maintenance instruction and freeze the similarity calculation result to prevent making wrong decisions based on distorted data. That is, the judgment of the abnormal hardware status of the instrument based on the ion source temperature and GC-MS vacuum is the first layer of protection. When the hardware status of the instrument is abnormal, it is meaningless to calculate the similarity.

[0194] In this embodiment, the instrument in which the abnormal hardware status of the instrument is detected refers to the Agilent 5977B mass spectrometer, and an alarm is triggered when the vacuum or ion source temperature is abnormal.

[0195] In this embodiment, when the ion source temperature and GC-MS vacuum have data anomalies, it means that the data exceeds the given normal value range for the data to be abnormal.

[0196] In this embodiment, the ion source temperature is generally 230.3 °C.

[0197] In this embodiment, the GC-MS vacuum is generally 2.8×10 -4 Pa.

[0198] Embodiment 8: The dual-channel network coupled by LSTM and GRU includes an input layer, a hidden layer, and an output layer:

[0199] The input layer is used to input the characteristic peaks in the dynamic fingerprint model, the sensor data after drift correction, and the device status parameters;

[0200] The hidden layer includes LSTM units and GRU units; the LSTM units are used to process the time-series data in the data input by the input layer; the GRU units are used to extract the device status features in the data input by the input layer;

[0201] The output layer includes a fully connected network; the fully connected network is used to predict the acid value change rate and the flavor stability index according to the output of the hidden layer.

[0202] In this embodiment, the loss function of the dual-channel network coupled by LSTM and GRU adopts the Huber loss.

[0203] Embodiment 9: According to the prediction results of the acid value change rate and the flavor stability index, the sterilization temperature and the turning frequency are adjusted in real time as follows:

[0204] The sterilization temperature is adjusted in real time using PID control:

[0205] When ΔAV > 0.5 mg / g·d-1, trigger the temperature control instruction of the sterilization tank:

[0206]

[0207] Where ΔAV is the rate of change of acid value; T new is the sterilization temperature adjusted in real time; T set Sterilization temperature set for sterilization process requirements; K p is the proportional gain coefficient; e(t) is the acid value change rate deviation; K i is the integral gain coefficient; K d is the differential gain coefficient;

[0208] Use frequency conversion speed regulation to adjust the frequency of turning the mash in real time:

[0209] Formula for adjusting the speed of the mash maker:

[0210] Unit: Hz;

[0211] Among them, FSI is the flavor stability index.

[0212] In this embodiment, T new It is the sterilization temperature after real-time adjustment, in ℃.

[0213] In this embodiment, T set The sterilization temperature set for the sterilization process requirements, in °C.

[0214] In this embodiment, e(t) is the acid value change rate deviation, and the unit is mg / g·d-1.

[0215] e(t)=ΔAV 实际 -ΔAV 设定 ;in:

[0216] ΔAV 实际 is the actual acid value change rate, ΔAV 设定 The acid value change rate is set according to the sterilization process requirements.

[0217] In this embodiment, K p is the proportional gain coefficient, and its unit is ℃·d-1 / mg / g.

[0218] In this embodiment, K i is the integral gain coefficient, and its unit is ℃·d-2 / mg / g.

[0219] In this embodiment, K d is the differential gain coefficient, and its unit is ℃·d / mg / g.

[0220] Embodiment 10: A method for monitoring the quality of pomelo paste based on GC-MS characteristic fingerprint, wherein the method is implemented by using the pomelo paste quality monitoring system based on GC-MS characteristic fingerprint described in any of the above embodiments; the method comprises the following steps:

[0221] The GC-MS detection steps: Chemically analyze the pomelo peel jam sample, separate volatile components through gas chromatography, and perform qualitative and quantitative detection through mass spectrometry to obtain the original GC-MS data;

[0222] The dynamic fingerprint map modeling steps: Extract characteristic peaks from the original GC-MS data to construct a dynamic fingerprint map model reflecting the product quality of the pomelo peel jam sample;

[0223] The drift correction steps: Collect sensor data and equipment status parameters, and perform drift correction on the collected data to ensure data stability;

[0224] The multi-source data fusion control steps: Generate process control instructions based on the characteristic peaks in the dynamic fingerprint map model, the sensor data after drift correction, and the equipment status parameters.

[0225] Embodiment 11: Provide a specific embodiment of the system.

[0226] As Figure 2 shown, the control flow chart of the intelligent monitoring system for the quality of Sizhou pomelo peel jam based on the GC-MS characteristic fingerprint map, the detailed control flow and implementation method of the system are as follows:

[0227] 1. The hardware configuration of the system is as follows:

[0228] (1) The C-MS detection unit corresponds to the function of the GC-MS detection module and is responsible for chemical component separation and detection:

[0229] ① Select an Agilent 7890B gas chromatograph and a 5977B mass spectrometry system, and configure a DB-WAX capillary column (60m × 0.25mm × 0.25μm);

[0230] The 5977B mass spectrometry system corresponds to the mass spectrometry detection unit, which is used to ionize compounds and perform mass-to-charge ratio scanning to collect mass spectrometry diagrams (m / z-intensity data).

[0231] The Agilent 7890B gas chromatograph and the DB-WAX capillary column correspond to the chromatographic separation unit, where:

[0232] The Agilent 7890B gas chromatograph is used to separate volatile components and collect time-chromatographic peak sequences;

[0233] The DB-WAX capillary column is used to optimize the separation of polar compounds and collect the retention times of the separated chromatographic peaks.

[0234] ② Select an Agilent G4513A automatic sampler, set the injection volume to 1 μL, in splitless mode, with helium (purity ≥ 99.999%) as the carrier gas, and a flow rate of 1.0 mL / min.

[0235] The Agilent G4513A auto sampler corresponds to the sample pretreatment unit, which is used for auto sampling and collecting the injection volume record.

[0236] (2) The edge computing node corresponds to the dynamic fingerprint model building module, which is responsible for real-time data processing:

[0237] The edge computing node is as follows:

[0238] Edge computing node

[0239] ① Deploy the Huawei Atlas 500 intelligent edge server (model: 500-3010), which is built-in with NVIDIA Tesla T4 GPU and equipped with the TensorRT 8.0 acceleration framework;

[0240] ② Connect directly to the GC-MS workstation through the PCIe 4.0 interface, and the data throughput rate ≥ 5 Gbps.

[0241] (3) The production line control unit corresponds to the multi-source data fusion control module, which is responsible for executing control instructions:

[0242] ① Siemens S7-1200 PLC controller, integrated with an analog input module (6ES7231-4HD32-0XB0) and a PROFINET communication interface;

[0243] The Siemens S7-1200 PLC controller corresponds to the production line regulation unit, which is used to receive and execute control instructions and collect device status parameters (current, valve opening).

[0244] The 6ES7231-4HD32-0XB0 module corresponds to the heterogeneous data integration unit, which is used to collect analog signals, such as temperature, humidity, and pH sensor data.

[0245] The PROFINET communication interface is used to communicate with the edge computing node and production line equipment (such as frequency converters and temperature control modules).

[0246] ② The temperature control module of the sterilization tank uses an SSR solid-state relay (output accuracy ±0.5 °C); the model of the turning machine frequency converter is ABB ACS880-01-025A-3.

[0247] 2. The construction process of the dynamic fingerprint model:

[0248] (1) Data collection and preprocessing:

[0249] ① Collect samples of the whole cycle of pomelo peel jam fermentation (sampling once every 15 days), freeze them quickly with liquid nitrogen, grind them to 80 mesh, take 2.0 g and add 10 mL of n-hexane for ultrasonic extraction for 30 min;

[0250] ② GC-MS parameters:

[0251] In the GC-MS detection stage, after ultrasonic extraction, a temperature programming is carried out to optimize the chromatographic separation effect; the parameters of the temperature programming are:

[0252] 50 °C (1 min) → 4 °C / min → 230 °C (5 min).

[0253] The volatile components in the pomelo peel jam sample are separated by a DB-WAX chromatographic column, and then ionization and scanning are carried out to cover the molecular weight range of the main volatile components of the pomelo peel jam; among them, the relevant parameters are:

[0254] Mass spectrometry scanning range 35 - 450 m / z; solvent delay 3 min;

[0255] Among them, the solvent delay is used to avoid solvent peak interference in mass spectrometry detection.

[0256] (2) Intelligent screening of characteristic peaks:

[0257] ① Use MS-DIAL 4.0 to perform peak alignment and match with the NIST 17 library for the preprocessed data (mass spectrometry data), and export a.csv file containing peak area, retention time, and characteristic fragment ions (such as m / z 93, 136, 161);

[0258] ② Run the improved random forest algorithm (RFRF-MS):

[0259]

[0260]

[0261] The above program implements a dynamic feature screening process based on an improved random forest, mainly for the analysis of GC-MS (gas chromatography - mass spectrometry) datasets. The following is a step-by-step description of its core processing process:

[0262] Step 1: Initialize parameters:

[0263] When creating a DynamicFingerprint object, preset the number of trees in the random forest (n_trees = 500) and the dynamic Gini threshold (gini_threshold = 0.03);

[0264] Initialize the standardization processor StandardScaler and the variable selected_peaks for storing the selected features;

[0265] Step 2: Data loading stage:

[0266] Read the input data in CSV format through the load_data method;

[0267] Extract the data matrix X (peak area data of all samples, such as 200 chromatographic peaks);

[0268] Extract the label vector y (classification labels of samples);

[0269] Step 3: Data preprocessing:

[0270] Use the preprocess method to perform Z-score normalization on the peak area data, converting the original data into a distribution with a mean of 0 and a variance of 1 to eliminate the dimensional difference;

[0271] Step 4: Model training stage:

[0272] Build a random forest classifier (train_rf method, random forest model) containing 500 decision trees;

[0273] Set the maximum depth of a single tree to 10 and use the Gini coefficient as the splitting criterion;

[0274] Train the classification model on the standardized data;

[0275] Step 5: Dynamic feature screening:

[0276] Extract the feature importance scores from the trained random forest (model) (select_features method);

[0277] Calculate the dynamic threshold: maximum importance score × 0.03;

[0278] Screen the chromatographic peaks with importance exceeding the dynamic threshold, and store the indices in selected_peaks;

[0279] Process integration execution:

[0280] The run_pipeline method concatenates the entire processing chain: data loading → standardization → model training → feature screening;

[0281] Finally, output the number and specific indices of the selected features (peaks), and return the screening results.

[0282] Technical features:

[0283] Adopt a dynamic threshold mechanism to automatically adjust the screening strictness according to the actual feature importance distribution;

[0284] Prevent overfitting by restricting the tree depth (max_depth = 10) and improve the reliability of feature importance evaluation;

[0285] The final output of this process is a set of chromatographic peaks with significant discriminatory ability for classification labels, which can be used in subsequent application scenarios such as mass spectrometry fingerprint analysis or biomarker discovery.

[0286]

[0287]

[0288]

[0289] The above program (module code 1.1) includes the process of verifying the fragment ion abundance ratio.

[0290] (3) Dynamic reference library update:

[0291] For every 10 newly added batches of qualified samples, calculate the cosine similarity threshold between the peak area vectors of the new samples and the historical qualified samples through the moving window similarity model (MW-SIM):

[0292]

[0293] Among them, A i represents the peak area vector of the new sample, and B i represents the peak area vector of the historical sample;

[0294] When the standard deviation of the similarity within the window > 0.05, automatically trigger the retraining of the reference library (i.e., re-execute the random forest algorithm to update the characteristic peaks).

[0295] 3. Implementation steps of drift correction

[0296] (1) Dual internal standard correction:

[0297] ① Add 2-octanol (10 μg / mL) and n-tetracosane (20 μg / mL) as internal standards to each batch of samples;

[0298] ② Calculate the relative retention time (RRT) according to the formula to eliminate the influence of column efficiency attenuation:

[0299]

[0300] (2) Equipment status compensation

[0301] Read the GC-MS vacuum degree (< 3×10 -4 Pa) and the ion source temperature (230 ± 5 °C) through the OPC-UA protocol;

[0302] When an abnormality is detected, automatically call the historical reference spectrum for interpolation compensation.

[0303] The relevant program is as follows:

[0304]

[0305]

[0306] Chromatographic data drift correction system, mainly targeting the instrument drift problem of retention time. The following is a detailed description of its core processing flow:

[0307] Initialize parameters:

[0308] Preset when creating a DriftCorrector object:

[0309] Sliding window size (window_size = 10): Store the most recent 10 qualified samples as a reference benchmark;

[0310] Similarity threshold (similarity_threshold = 0.92): The critical value for determining whether the current sample has drifted;

[0311] Reference window (reference_window): A dynamically maintained reference sample queue;

[0312] Relative retention time calculation:

[0313] calculate_rrt method:

[0314] (target_rt - std1_rt) / (std2_rt - std1_rt)*1000;

[0315] Use the retention times of two internal standards (2-octanol, n-tetracosane) as anchor points;

[0316] Convert the absolute retention time of the target substance to relative retention time (RRT) to eliminate the absolute time offset caused by instrument fluctuations;

[0317] Dynamic update of the reference window:

[0318] update_reference_window method:

[0319] Initialize the reference window as the current sample (in the form of a two-dimensional array) when called for the first time;

[0320] Subsequently, append new samples through np.vstack to maintain a first-in, first-out (FIFO) queue;

[0321] The window length is always ≤ 10, and when it exceeds, the earliest sample is automatically removed;

[0322] Drift detection mechanism:

[0323] compute_similarity method:

[0324] Calculate the cosine similarity between the current sample and the reference window mean using the formula: 1 - cosine(sample, ref_mean);

[0325] In the initial state (no reference sample), the highest similarity of 1.0 is returned by default.

[0326] Main calibration process:

[0327] drift_correction method:

[0328] Step 1: Calculate the similarity between the current sample and the reference window

[0329] Step 2: If the similarity is lower than the threshold (<0.92):

[0330] Trigger an alarm (Alert prompt);

[0331] Return False to indicate that drift is detected and the reference window is not updated;

[0332] Step 3: If the similarity meets the standard (≥0.92):

[0333] Add the current sample to the reference window;

[0334] Return True to indicate that the data is normal.

[0335] Technical features:

[0336] Dynamic reference benchmark: Maintain the statistical characteristics of the last 10 qualified samples through a sliding window, and the mean calculation reflects the stable state of the current instrument.

[0337] Dual standardization:

[0338] RRT calculation: Eliminate absolute time drift through internal standards;

[0339] Similarity comparison: Detect relative distribution anomalies through cosine similarity.

[0340] Real-time feedback mechanism:

[0341] The drift detection result directly affects the update of the reference window;

[0342] The alarm threshold (0.92) can balance sensitivity and fault tolerance.

[0343] 4. Multi-source data fusion control

[0344] (1) Dual-channel network coupling LSTM and GRU:

[0345] ① Input layer: 15-dimensional GC-MS characteristic peak area + 6-dimensional equipment parameters (temperature, humidity, harmonic distortion rate of turning pile current, motor speed, steam valve opening, oxygen concentration);

[0346] ② Hidden layer: LSTM units (128 nodes) process time series data, and GRU units (64 nodes) extract equipment status features;

[0347] ③ Output layer: A fully connected network is used to predict the acid value change rate (ΔAV) and flavor stability index (FSI), and the Huber loss is used as the loss function.

[0348] The relevant program is as follows:

[0349]

[0350] The above program implements a dual-channel time series prediction model based on a dual-channel network coupled with LSTM and GRU for variable regression tasks (such as predicting two indicators of ΔAV and FSI). The following is a detailed description of its core processing flow:

[0351] Step 1: Construction of the input layer:

[0352] Input dimension: (time_steps = 10, input_dim = 21);

[0353] Time step: 10 consecutive time points form a sample segment;

[0354] Feature dimension: Each time point contains 21-dimensional features (the code comment indicates that it consists of 15 + 6 features);

[0355] Step 2: Bidirectional LSTM channel:

[0356] Bidirectional structure: Capture forward and backward time series dependencies simultaneously;

[0357] Parameter configuration:

[0358] Number of LSTM units: 128 (the actual output dimension of the bidirectional structure is 128 × 2 = 256);

[0359] return_sequences = True: Retain the output of all time steps (for subsequent GRU processing);

[0360] Step 3: GRU channel:

[0361] Parameter configuration:

[0362] Number of GRU units: 64

[0363] Default return_sequences = False: Only output the state of the last time step

[0364] Data flow:

[0365] Receive the 256-dimensional time series data output by the LSTM channel;

[0366] Further extract the time series features and compress them into a 64-dimensional vector;

[0367] Step 4: Output layer:

[0368] Regression prediction: The fully connected layer outputs 2 continuous values: acid value change rate (ΔAV) and flavor stability index (FSI); Step 5: Model compilation:

[0369] Optimizer: Adam adaptive learning rate optimization;

[0370] Loss function: Mean squared error (MSE) is applicable to regression tasks.

[0371] Analysis of technical features:

[0372] Dual-channel collaboration:

[0373] Advantages of LSTM: Capture long-distance time series dependencies through the gating mechanism;

[0374] Advantages of GRU: Simplify the parameter structure (no cell state) and improve the training efficiency.

[0375] Hierarchical design: First, use the high-capacity LSTM to extract deep features, and then use GRU for feature compression.

[0376] Input-output adaptation:

[0377] Input structure: (10, 21) adapts to the multivariate time series data generated by sensors and other devices;

[0378] Output design: Dual-objective prediction can simultaneously monitor the acid value change rate (ΔAV) and the flavor stability index (FSI).

[0379] (2) Closed-loop control of the production line:

[0380] When ΔAV > 0.5 mg / g·d-1, trigger the temperature control instruction of the sterilization tank:

[0381]

[0382] Turning machine speed adjustment formula:

[0383]

[0384] The relevant program is as follows:

[0385]

[0386] The above program implements a discrete-time PID controller, which is mainly used for the closed-loop regulation of dynamic systems (such as temperature control).

[0387] Step 1: Initialize parameters:

[0388] Initialize when creating an object:

[0389] Kp = 2.5 # Proportional gain coefficient

[0390] Ki = 0.1 # Integral gain coefficient

[0391] Kd = 0.8 # Derivative gain coefficient

[0392] last_error = 0 # Store the previous error

[0393] integral = 0 # Cumulative value of the integral term

[0394] Step 2: Calculate the error:

[0395] Current error:

[0396] error = setpoint - measured_value # Deviation between the target value and the measured value; for example: temperature setpoint 25°C, measured 23°C → error = 2°C;

[0397] Step 3: Update the integral term:

[0398] Discrete integral calculation:

[0399] self.integral += error * dt # Integral term = Σ(error × time step); assume dt = 1 minute, continuous error of 2 for 3 minutes → integral term accumulates to 6;

[0400] Step 4: Calculate the derivative term:

[0401] Rate of change of error:

[0402] derivative = (error - self.last_error) / dt # Current rate of change of error, previous error 1°C, 2°C → rate of change = (2 - 1) / 1 = 1°C / minute.

[0403] Step 5: Combine the PID formula:

[0404] Calculate the control quantity:

[0405] control = Kp * error + Ki * integral + Kd * derivative;

[0406] Example calculation:

[0407] error = 2, integral = 6, derivative = 1;

[0408] control = 2.5 * 2 + 0.1 * 6 + 0.8 * 1 = 5 + 0.6 + 0.8 = 6.4;

[0409] Step 6: State update:

[0410] Save the current error:

[0411] self.last_error = error # Prepare for the next derivative calculation

[0412] Data flow:

[0413] Set value --> Error calculation

[0414] Measured value --> Error calculation

[0415] Error --> Integral term

[0416] Error --> Derivative term

[0417] Integral term --> Control synthesis

[0418] Derivative term --> Control synthesis

[0419] Control synthesis --> Output control quantity.

[0420] Technical features:

[0421] Discrete-time implementation, suitable for digital control systems (such as PLCs, microcontrollers); the time step dt is default 1 minute and can be adjusted according to the sampling frequency.

[0422] 5. In the above four module codes, the key algorithms are described as follows:

[0423] (1) Improved random forest feature screening

[0424] Dynamic threshold calculation: dynamic_threshold = gini_threshold * max(importance)

[0425] Parameter description:

[0426] n_trees = 500: The number of decision trees in the forest

[0427] gini_threshold = 0.03: Dynamic threshold coefficient (optimized according to experimental data)

[0428] (2) LSTM-GRU dual-channel network

[0429] Input dimension: (None, 10, 21) (10 time steps, 21-dimensional features)

[0430] Network structure: Bidirectional LSTM(128) → GRU(64) → Dense(2)

[0431] (3) PID control algorithm

[0432] Control equation:

[0433]

[0434] Parameters:

[0435] K p = 2.5: Proportional term coefficient (rapid response to main deviation)

[0436] K i = 0.1: Integral term coefficient (eliminating steady-state error)

[0437] K d = 0.8: Derivative term coefficient (suppressing overshoot)

[0438] 6. In the above four module codes, the data interfaces are described as follows:

[0439] (1) Input data format:

[0440] GC-MS data: CSV file, the first column is the sample ID, the second column is the label (0 / 1), and the subsequent columns are the peak areas

[0441] SampleID,Label,Peak1,Peak2,...,Peak200

[0442] 1,1,0.532,1.234,...,0.987

[0443] 2,0,0.456,0.876,...,1.023

[0444] (2) Output results:

[0445] selected_peaks: List of indices of the selected key feature peaks

[0446] PID control instruction: Temperature adjustment amount (unit: °C) or turning frequency adjustment amount (unit: Hz).

[0447] The above further describes the technical solutions provided by the present invention through several specific embodiments to highlight the advantages and beneficial effects of the technical solutions provided by the present invention. However, the above-mentioned several specific embodiments are not used as a limitation to the present invention. Any reasonable modification and improvement of the present invention, reasonable combination of implementation manners, equivalent substitution, etc. within the spirit and principle of the present invention should be included within the protection scope of the present invention.

Claims

1. A quality monitoring system for pomelo peel jam based on GC-MS characteristic fingerprint, characterized in that The system includes the following modules: a GC-MS detection module, a dynamic fingerprint spectrum modeling module, a drift correction module, a real-time analysis module, and a multi-source data fusion control module; The GC-MS detection module is used to perform chemical analysis on the pomelo peel jam sample, separate volatile components through gas chromatography, and perform qualitative and quantitative detection through mass spectrometry to obtain the original GC-MS data; The dynamic fingerprint spectrum modeling module is used to extract characteristic peaks from the original GC-MS data and construct a dynamic fingerprint spectrum model reflecting the product quality of the pomelo peel jam sample; The drift correction module is used to collect sensor data and equipment status parameters, and perform drift correction on the collected data to ensure data stability; The multi-source data fusion control module is used to generate process control instructions based on the characteristic peaks in the dynamic fingerprint spectrum model, the sensor data and equipment status parameters after drift correction.

2. The quality monitoring system of pomelo peel jam based on GC-MS characteristic fingerprint spectrum according to claim 1, wherein, The GC-MS detection module includes a sample pretreatment unit, a chromatographic separation unit, a mass spectrometry detection unit, and a data acquisition unit; among them: The sample pretreatment unit is used to grind, extract, filter, and load the pomelo peel jam sample into the sample vial; The chromatographic separation unit is used to separate the volatile components in the pomelo peel jam sample through a DB-WAX chromatographic column; The mass spectrometry detection unit is used to ionize and scan the separated volatile components; The data acquisition unit is used to collect and store in real time the original GC-MS data output by the mass spectrometry detection unit.

3. The pomelo peel jam quality monitoring system based on the GC-MS characteristic fingerprint spectrum according to claim 1, characterized in that The dynamic fingerprint spectrum modeling module includes a data preprocessing unit, a characteristic peak screening unit, and a model training unit; among them: The data preprocessing unit is used to preprocess the original GC-MS data; The characteristic peak screening unit is used to screen key peaks from the preprocessed data by using an improved random forest algorithm; The model training unit is used to construct a multiple regression model based on the screened key peaks and process parameters as a dynamic fingerprint spectrum model reflecting the product quality of the pomelo peel jam sample.

4. The quality monitoring system of pomelo peel jam based on GC-MS characteristic fingerprint spectrum according to claim 1, characterized in that, The drift correction module includes an environmental drift compensation unit; The environmental drift compensation unit realizes double compensation for instrument fluctuations and raw material differences through the coupling mechanism of double internal standard correction and moving window similarity model, and improves the anti-interference ability of environmental drift.

5. The quality monitoring system of pomelo peel jam based on GC-MS characteristic fingerprint spectrum according to claim 1, characterized in that, The multi-source data fusion control module includes a heterogeneous data integration unit, an intelligent decision-making unit, and a production line control unit; among them: The heterogeneous data integration unit is used to obtain the characteristic peaks in the dynamic fingerprint spectrum model, the sensor data and equipment status parameters after drift correction through the OPC-UA protocol; The intelligent decision-making unit is used to predict the acid value change rate and flavor stability index based on the characteristic peaks in the dynamic fingerprint spectrum model, the sensor data and equipment status parameters after drift correction by using a dual-channel network coupling LSTM and GRU; The production line control unit is used to adjust the sterilization temperature and turning frequency in real time according to the prediction results of the acid value change rate and flavor stability index.

6. The quality monitoring system of pomelo peel jam based on the GC-MS characteristic fingerprint spectrum according to claim 1, wherein, The improved random forest algorithm is adopted to screen key peaks from the preprocessed data, including: From the preprocessed data, extract the peak area data of all samples as the data matrix, extract the classification labels of the samples as the label vector, and perform standardization operations to obtain the standardized training data; Construct an initial random forest model with a given number of decision trees, limit the maximum depth of a single tree, and use the Gini coefficient as the splitting criterion; Use the standardized training data to train the initial random forest model to obtain a trained random forest model; Extract the feature importance scores from the trained random forest model; Screen the chromatographic peaks with feature importance scores exceeding the dynamic threshold as candidate feature peaks; among them, the dynamic threshold is: dynamic threshold = maximum value of the feature importance scores in the trained random forest model × 0.03; Perform fragment ion abundance ratio verification on the candidate feature peaks, eliminate interference peaks, and screen out key peaks.

7. The quality monitoring system of pomelo peel jam based on GC-MS characteristic fingerprint spectrum according to claim 4, wherein The coupling mechanism of the double internal standard correction and the moving window similarity model: Use double internal standard correction to correct the retention time and handle the time drift at the instrument hardware level; Use the moving window similarity model to perform similarity matching on the peak areas to compensate for the peak area fluctuations caused by raw material batch differences.

8. The quality monitoring system of pomelo peel jam based on GC-MS characteristic fingerprint spectrum according to claim 1, characterized in that The dual-channel network coupling LSTM and GRU includes an input layer, a hidden layer, and an output layer: The input layer is used to input the feature peaks in the dynamic fingerprint map model, the sensor data after drift correction, and the device status parameters; The hidden layer includes LSTM units and GRU units; the LSTM units are used to process the time-series data in the data input by the input layer; the GRU units are used to extract the device status features in the data input by the input layer; The output layer includes a fully connected network; the fully connected network is used to predict the acid value change rate and the flavor stability index according to the output of the hidden layer.

9. The quality monitoring system of pomelo peel jam based on GC-MS characteristic fingerprint spectrum according to claim 1, characterized in that According to the prediction results of the acid value change rate and the flavor stability index, the sterilization temperature and the turning frequency are adjusted in real time as follows: Use PID control to adjust the sterilization temperature in real time: When ΔAV > 0.5 mg / g·d-1, trigger the temperature control instruction of the sterilization tank: where, ΔAV is the acid value change rate; T new is the sterilization temperature after real-time adjustment; T set is the sterilization temperature set according to the sterilization process requirements; K p is the proportional gain coefficient; e(t) is the deviation of the acid value change rate; K i is the integral gain coefficient; K d is the derivative gain coefficient; Use variable frequency speed regulation to adjust the turning frequency in real time: Turning machine speed adjustment formula: Unit: Hz; Where FSI is the flavor stability index.

10. A method for monitoring the quality of pomelo peel jam based on GC-MS characteristic fingerprint, characterized in that, The method is implemented by the quality monitoring system of pomelo peel jam based on the GC-MS characteristic fingerprint map described in any one of claims 1 to 9; the method includes the following steps: The GC-MS detection step: chemically analyze the pomelo peel jam sample, separate the volatile components by gas chromatography, and perform qualitative and quantitative detection by mass spectrometry to obtain the original GC-MS data; The dynamic fingerprint map modeling step: extract feature peaks from the original GC-MS data and construct a dynamic fingerprint map model reflecting the product quality of the pomelo peel jam sample; The drift correction step: collect sensor data and device status parameters, and perform drift correction on the collected data to ensure data stability; The multi-source data fusion control step: generate process control instructions according to the feature peaks in the dynamic fingerprint map model, the sensor data after drift correction, and the device status parameters.

Images