Tea processing process fermentation monitoring method based on intelligent sensor

Abstract

The application discloses a tea processing fermentation monitoring method based on intelligent sensors and relates to the technical field of intelligent monitoring of tea processing. The method continuously collects temperature, humidity, volatile gas concentration and leaf reflectance spectrum data in the fermentation process through an intelligent sensor array; performs adaptive window segmentation on multi-source sensing data to obtain fermentation monitoring segments arranged in time; extracts the cross-covariance matrix between multi-source sensing data in each segment and performs characteristic decomposition to obtain a characteristic value sequence; calculates the spectral entropy of the characteristic value sequence to represent the information complexity of the fermentation state; performs dynamic time warping on the spectral entropy sequence of each segment and a pre-constructed fermentation reference spectral entropy curve to obtain a warping distance; and determines whether the tea is in an abnormal fermentation state according to the comparison result of the warping distance and an abnormal threshold. The method can utilize multi-source sensing interaction information and time sequence elastic alignment means to improve the accuracy of fermentation anomaly recognition.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for tea processing, specifically a method for monitoring the fermentation process of tea based on intelligent sensors. Background Technology

[0002] The quality control of tea fermentation directly affects the color, aroma, and flavor of the finished tea. Existing fermentation monitoring methods largely rely on human sensory evaluation or single-type sensors. Human experience alone is insufficient to quantify the fermentation state and suffers from strong subjectivity and poor consistency. Using a single temperature or humidity sensor only reflects an isolated aspect of the environment and cannot capture the complex biochemical changes within the fermentation system. Some solutions introduce gas sensor arrays to monitor volatile components, but these typically only perform threshold judgments or simple fusion of signals from each channel, neglecting the dynamic interactions between signals from different sensing dimensions. The fermentation process itself is a strongly nonlinear time-varying process; the coupling patterns of various sensing variables change significantly with the fermentation stage. Changes in the amplitude of a single sensor signal often lag behind the actual state shift, leading to delayed anomaly detection. Current technologies also commonly employ a comparison method based on a fixed time scale, comparing the currently collected data with a preset standard fermentation curve at the same moment. Because tea fermentation is affected by variety, environmental temperature and humidity fluctuations, and differences in enzyme activity, the fermentation time varies between different batches, resulting in a nonlinear time distortion between the time axis of the standard curve and the actual progress rate of the current batch. This temporal mismatch makes distance measurement methods for corresponding sample values ​​at fixed time points difficult to reflect the true degree of state deviation, easily leading to false alarms or missed alarms. Extracting features that characterize the overall information structure changes of fermentation from multi-source sensor signals and compensating for alignment errors caused by nonlinear distortions in temporal comparisons is key to improving the accuracy of abnormal fermentation monitoring. The challenge lies in effectively representing the covariant information from multi-source sensors and achieving flexible temporal matching. Summary of the Invention

[0003] This paper presents a method for monitoring the fermentation process of tea based on intelligent sensors. It solves the problems that existing methods fail to fully extract the dynamic correlation features between multi-source sensor data and cannot adapt to the nonlinear scaling of the fermentation time axis. It realizes the quantitative expression of the complexity of tea fermentation state information and achieves accurate judgment of abnormal fermentation through time-series distortion alignment.

[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a method for monitoring the fermentation process of tea processing based on intelligent sensors, comprising: acquiring multi-source sensor data continuously collected by an intelligent sensor array during the tea fermentation process, wherein the multi-source sensor data includes temperature data, humidity data, volatile gas concentration data, and leaf reflectance spectrum data; performing adaptive window segmentation on the multi-source sensor data to obtain multiple fermentation monitoring segments arranged in chronological order; extracting the cross-covariance matrix between the multi-source sensor data within each fermentation monitoring segment, and performing eigenvalue decomposition on the cross-covariance matrix to obtain an eigenvalue sequence; calculating the spectral entropy of the eigenvalue sequence of each fermentation monitoring segment, wherein the spectral entropy is used to characterize the information complexity of the fermentation state; performing dynamic time warping on the spectral entropy of each fermentation monitoring segment and a pre-constructed fermentation baseline spectral entropy curve, and calculating the warped distance between each fermentation monitoring segment and the fermentation baseline spectral entropy curve; and determining whether the tea corresponding to the current fermentation monitoring segment is in an abnormal fermentation state based on the comparison result of the warped distance and a preset anomaly threshold. This method integrates multi-source sensor information and combines spectral entropy with dynamic time warping to sensitively capture complex changes in the fermentation process and accurately identify abnormal fermentation states while accommodating normal time shifts.

[0005] As a preferred technical solution of the present invention, when acquiring multi-source sensor data, an array of intelligent sensors is deployed at multiple preset locations in the fermentation container. Each intelligent sensor integrates a temperature-sensitive element, a humidity-sensitive element, a metal oxide semiconductor gas sensor, and a miniature spectrometer. All intelligent sensors are set to synchronously collect data at the same sampling frequency. The sampling frequency is dynamically adjusted according to the fermentation stage of the tea. A first sampling frequency is used in the initial fermentation stage, and a second sampling frequency higher than the first sampling frequency is used in the main fermentation stage. The multi-source sensor data collected at each sampling moment is timestamped and outliers exceeding the sensor range are filtered out to form a continuous multi-source sensor data stream. Further, before the start of each fermentation process, all intelligent sensors are zero-point calibrated and range calibrated using a standard reference material, and the calibration coefficient of each intelligent sensor is recorded. During continuous acquisition, zero-point drift correction is automatically performed at fixed time intervals, and the difference between the current sensor output value in a sample-free environment and the initial zero-point value is used to compensate for subsequent data acquisition. The compensated multi-source sensor data is normalized according to the sensing dimension, so that the data value range of each dimension is mapped to the zero-to-one interval. This effectively eliminates the adverse effects of sensor drift and dimensional differences, ensuring the consistency of basic data.

[0006] During adaptive window segmentation, the short-time energy change rate of each sensing dimension in the multi-source sensor data is calculated, which is the absolute value of the difference between the sensing values ​​at the current time and the previous time. When the short-time energy change rate of any sensing dimension exceeds a first threshold, that time is marked as a candidate segmentation point. The short-time energy change rates of all sensing dimensions within a set time range before and after the candidate segmentation point are accumulated. If the accumulated value exceeds a second threshold, the candidate segmentation point is determined as the formal segmentation boundary. The multi-source sensor data is segmented according to all formal segmentation boundaries, and the data segment between two adjacent formal segmentation boundaries is taken as a fermentation monitoring segment. Based on the preset minimum and maximum window durations for the tea fermentation stage, the values ​​of the first and second thresholds are dynamically adjusted so that the segmentation can capture key state transitions without generating too many sporadic fragments. As a further improvement, when the time interval between three or more consecutive formal segmentation boundaries is less than the minimum window duration, the segmentation boundary with the smallest interval is deleted, and two adjacent fermentation monitoring segments are merged. When the duration of a single fermentation monitoring segment exceeds the maximum window duration, additional segmentation boundaries are inserted within that segment at equal time intervals, dividing the segment into multiple sub-segments. For each fermentation monitoring segment generated after merging or segmentation, its start and end times are recalculated to ensure that all fermentation monitoring segments are temporally continuous and non-overlapping. Thus, adaptive segmentation can obtain monitoring segments with reasonable structure, providing a high-quality time window for subsequent analysis.

[0007] For a fermentation monitoring segment, the multi-source sensor data from all sampling times are arranged into a data matrix, with rows corresponding to sensor dimensions and columns corresponding to sampling times. Each row of the data matrix is ​​centered by subtracting its mean, resulting in a centered data matrix. The centered data matrix is ​​multiplied by its transpose and then divided by the number of sampling times minus one, yielding a cross-covariance matrix. Jacobian iterative eigenvalue decomposition is performed on the cross-covariance matrix to obtain all eigenvalues, which are then sorted in descending order to form an eigenvalue sequence. When calculating the spectral entropy of this fermentation monitoring segment, each eigenvalue in the eigenvalue sequence is divided by the sum of all eigenvalues ​​to obtain a normalized weight for each eigenvalue. The logarithm of each normalized weight is taken to the base of the natural constant and then multiplied by the normalized weight itself to obtain the entropy contribution value corresponding to each eigenvalue. The negative values ​​of the entropy contributions for all eigenvalues ​​are then summed to obtain the spectral entropy of the fermentation monitoring segment. The higher the spectral entropy value, the greater the information complexity of the multi-source sensor data at the current moment, and the more drastic the microscopic changes in the fermentation state. The spectral entropy of normal fermentation usually evolves smoothly along the baseline path, while abnormal fermentation shows significant entropy perturbation.

[0008] In the dynamic time warping stage, the pre-constructed fermentation baseline spectral entropy curve is a time series that records the spectral entropy reference values ​​sampled chronologically under standard fermentation processes. For the spectral entropy time series of the current fermentation monitoring segment and the fermentation baseline spectral entropy curve, a distance matrix is ​​constructed, where each element represents the absolute difference in spectral entropy values ​​at corresponding positions in the two sequences. A curved path with the minimum cumulative distance is searched from the upper left to the lower right corner of the distance matrix. The path search allows movement along the horizontal, vertical, and diagonal directions, with the cumulative distance at each step being the sum of the absolute difference at the current step and the minimum cumulative distance at the previous step. The cumulative distance at the endpoint of the curved path is taken as the warped distance between the current fermentation monitoring segment and the fermentation baseline spectral entropy curve. Preferably, during dynamic time warping, a window constraint is applied to the search for the curved path, limiting the path offset on either side of the diagonal in the distance matrix to no more than a preset curved window width. This reduces computational cost and avoids distorted excessive curvature, allowing the warped distance to more accurately reflect the substantial shift in the fermentation state.

[0009] When determining an abnormal fermentation state, the mean and standard deviation of the regularized distances of all completed fermentation monitoring segments are calculated. The anomaly threshold is set to the mean plus twice the standard deviation. If the regularized distance of the current fermentation monitoring segment is greater than the anomaly threshold, the tea is determined to be in an abnormal fermentation state, and the time interval of the fermentation monitoring segment is recorded. If the regularized distance of the current fermentation monitoring segment is less than or equal to the anomaly threshold, the tea is determined to be in a normal fermentation state. When multiple consecutive fermentation monitoring segments are determined to be in an abnormal fermentation state, the anomaly threshold is dynamically adjusted to the current mean plus three times the standard deviation to improve the specificity of anomaly determination and reduce false alarms caused by short-term fluctuations.

[0010] In a preferred embodiment of the present invention, after determining an abnormal fermentation state, the original multi-source sensor data, corresponding normalization distance, and anomaly determination results of the fermentation monitoring segment determined to be in an abnormal fermentation state are saved to an anomaly record database. Historical anomaly segments with similar multi-source sensor data change patterns to the current abnormal fermentation monitoring segment are retrieved from the anomaly record database, and subsequent fermentation treatment measures for these historical anomaly segments are extracted. Adjustment instructions are generated based on the extracted subsequent fermentation treatment measures, and these adjustment instructions are used to control the operation of the temperature regulator, humidity regulator, and ventilation device of the fermentation apparatus. Through the above closed-loop adjustment, rapid response and precise intervention to fermentation anomalies are achieved, significantly reducing the quality loss of tea caused by uncontrolled fermentation.

[0011] The technical effects and advantages provided by the present invention in the above technical solution are as follows: By extracting the cross-covariance matrix among multi-source sensor data within each fermentation monitoring segment and performing eigenvalue decomposition to obtain the eigenvalue sequence, the spectral entropy is calculated. Multi-source sensor signals are not isolated changes; temperature, humidity, gas concentration, and leaf spectra exhibit covariant relationships driven by the coupling of fermentation biochemical reactions. The cross-covariance matrix fully captures the linear correlation strength and temporal co-structure between any two sensing dimensions. After eigenvalue decomposition of this matrix, the eigenvalue sequence reflects the energy distribution of multi-source signals under different interaction modes. Based on the eigenvalue sequence, the normalized spectral entropy is calculated, transforming this distribution into a single information complexity index. When fermentation proceeds smoothly, the covariant structure of multi-source sensing is relatively stable, and the spectral entropy remains within a certain range. Once fermentation deviates from the normal path, the correlation between dimensions becomes disordered, the eigenvalue distribution tends to homogenize, or a dominant mode abruptly changes, resulting in a significant change in the spectral entropy. This measurement method directly captures changes in the internal organizational state of the fermentation system. Compared to monitoring only the absolute value or rate of change of a single sensor quantity, it is more sensitive to anomalies and is not significantly affected by independent noise such as sensor zero-point drift, thus improving the robustness of monitoring. Dynamic time warping is used to align the spectral entropy sequence of each segment with the pre-constructed fermentation baseline spectral entropy curve by bending the path, and the warping distance is calculated as the basis for anomaly judgment. The fermentation rates of different batches of tea naturally differ, and even if the final quality is qualified, the trajectory of its spectral entropy change over time may be stretched or compressed. The traditional method of calculating the Euclidean distance point by point according to a fixed time index will amplify the distance measurement due to slight deviations in the time axis, thus misreporting normal fluctuations as anomalies. Dynamic time warping searches for the bending path with the minimum cumulative distance in the distance matrix, allowing for local scaling and displacement compensation of the time dimension of the sequence, so that spectral entropy change patterns with similar shapes can be matched after deformation. The warping distance obtained in this way reflects the essential morphological difference between the two curves after extension and contraction when ignored. By comparing the regularized distance with the adaptive anomaly threshold, it is possible to effectively distinguish between tolerable time-series shifts caused by changes in fermentation rate and pattern deviations caused by abnormal fermentation degree, significantly reducing the false alarm rate and enabling the capture of abnormal signals in the early stages of fermentation and timely triggering of regulation. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0013] Figure 1 This is a flowchart of a method for monitoring the fermentation process of tea based on intelligent sensors; Figure 2 This is a flowchart of multi-source sensor data acquisition; Figure 3 This is a flowchart of multi-source sensor data calibration and preprocessing; Figure 4 This is a flowchart of a multi-source sensor data adaptive window segmentation method; Figure 5 This is a flowchart of the calculation of the cross-covariance matrix and spectral entropy of multi-source sensor data for fermentation monitoring segments; Figure 6 It is a flowchart for judging and handling abnormal fermentation states based on dynamic time regularization. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0015] See Figure 1 This invention provides a method for monitoring the fermentation process of tea based on intelligent sensors, enabling fully automated monitoring and anomaly detection of the tea fermentation state. The method utilizes an array of intelligent sensors deployed at multiple preset locations within the fermentation container to continuously collect temperature, humidity, volatile gas concentration, and leaf reflectance spectrum data during tea fermentation, forming a multi-source sensor data stream. The collected multi-source sensor data is adaptively windowed to obtain multiple fermentation monitoring segments arranged in chronological order. For each fermentation monitoring segment, the cross-covariance matrix between its internal multi-source sensor data is extracted, and eigenvalue decomposition is performed on the cross-covariance matrix to obtain a sequence of eigenvalues ​​sorted from largest to smallest. The spectral entropy of the eigenvalue sequence for each fermentation monitoring segment is calculated, representing the information complexity of the fermentation state within the corresponding segment. The time series formed by the spectral entropy of each fermentation monitoring segment is dynamically time-warped with a pre-constructed fermentation baseline spectral entropy curve, and the warped distance between each fermentation monitoring segment and the fermentation baseline spectral entropy curve is calculated. The fermentation baseline spectral entropy curve is a pre-obtained sequence of spectral entropy reference values ​​under standard fermentation process conditions. The calculated regularity distance is compared with the preset anomaly threshold. If the regularity distance of the current fermentation monitoring segment is greater than the anomaly threshold, the tea corresponding to the segment is determined to be in an abnormal fermentation state; otherwise, it is determined to be in a normal fermentation state.

[0016] Example 1: In specific implementation, please refer to Figure 2The process of acquiring multi-source sensor data continuously collected by an intelligent sensor array during tea fermentation includes deploying the intelligent sensor array, synchronous acquisition, dynamically adjusting the sampling frequency, timestamp alignment and outlier filtering, and forming a multi-source sensor data stream.

[0017] Multiple pre-selected locations are chosen inside the fermentation container, distributed at different depths and horizontal orientations, allowing the collected sensor data to reflect the spatial characteristics of the tea fermentation pile. Each pre-selected location houses a smart sensor, forming a smart sensor array. The smart sensor array integrates temperature-sensitive elements, humidity-sensitive elements, metal oxide semiconductor gas sensors, and a miniature spectrometer. The temperature-sensitive elements are contact-type sensors used to detect the fermentation environment temperature and localized temperature changes within the leaf pile. The humidity-sensitive elements are capacitive or resistive humidity sensors used to acquire air humidity and humidity information near the leaf surface. The metal oxide semiconductor gas sensors respond to changes in resistance in response to volatile organic compounds such as ethanol and acetaldehyde released during fermentation, detecting volatile gas concentrations. The miniature spectrometer uses an internal grating or filter structure to collect leaf reflectance spectra in the visible to near-infrared range, acquiring leaf reflectance spectral data.

[0018] All smart sensors synchronize via a unified clock mechanism, achieving millisecond-level accuracy. At each sampling moment, each smart sensor simultaneously triggers a data acquisition action, reading temperature data, humidity data, volatile gas concentration data, and leaf reflectance spectrum data from the temperature-sensitive element, humidity-sensitive element, metal oxide semiconductor gas sensor, and miniature spectrometer, respectively, thus achieving synchronous acquisition at the same sampling frequency. The sampling frequency is not fixed but dynamically adjusted according to the tea fermentation stage. The tea fermentation stage is divided into the initial fermentation stage and the main fermentation stage. The stage switching can be automatically determined by a pre-agreed fermentation time or manually switched by the operator based on experience such as changes in tea color and aroma. In the initial fermentation stage, the oxidation degree of polyphenols in the leaves is low, and the biochemical reaction rate is relatively slow. Therefore, the first sampling frequency is used, set to once per minute. In the main fermentation stage, as the oxidation reaction intensifies, the changes in temperature, humidity, and gas concentration accelerate significantly, requiring more frequent data sampling. Therefore, the second sampling frequency is used, set to four times per minute.

[0019] In practice, after each data acquisition, the smart sensors package the collected temperature, humidity, volatile gas concentration, and leaf reflectance spectrum data along with a high-precision timestamp of the current sampling moment, and send them to the data processing unit via a wired communication bus or wireless communication module. Upon receiving the data packets from each smart sensor, the data processing unit timestamps the multi-source sensor data uploaded by different smart sensors at the same timestamp. If a smart sensor fails to upload data at a specific sampling moment, it uses data from the same dimension from the previous valid sampling moment or completes the upload using linear interpolation to maintain the synchronous correspondence of the multi-source data in the time dimension. Simultaneously, the data processing unit performs range verification on the data for each sensor dimension, eliminating outliers exceeding the sensor's range. Temperature-sensitive elements, humidity-sensitive elements, metal-oxide-semiconductor gas sensors, and miniature spectrometers are all preset with their own normal operating ranges. The upper and lower limits of the temperature range are 80℃ and 0℃, respectively; the upper and lower limits of the humidity range are 100%RH and 0%RH, respectively; the upper and lower limits of the volatile gas concentration range are determined according to the specifications of the selected metal-oxide-semiconductor gas sensor; and the upper and lower limits of the spectral intensity of the leaf reflectance spectral data are determined by the dynamic range of the miniature spectrometer. When a sensor value exceeds the corresponding range, the value is marked as an outlier and filtered out. The result of time-series interpolation of adjacent valid values ​​is used to replace the filtered outlier. After timestamp alignment and outlier filtering, the multi-source sensor data is arranged in ascending order of timestamps, forming a continuous multi-source sensor data stream for subsequent adaptive window segmentation.

[0020] Example 2: In specific implementation, please refer to Figure 3 The process of acquiring multi-source sensor data continuously collected by the intelligent sensor array during the tea fermentation process also includes sensor calibration and data preprocessing, specifically zero-point calibration and range calibration, calibration coefficient recording, zero-point drift correction, data compensation, and normalization processing.

[0021] Before each fermentation process, all smart sensors underwent zero-point and range calibration using standard reference materials. The standard reference materials were selected based on the type of sensitive element being calibrated. For temperature-sensitive elements, the temperature-sensitive element of the smart sensor was placed in a constant temperature bath along with a calibrated standard platinum resistance thermometer. When the constant temperature bath was set to 0°C, the output value of the temperature-sensitive element was recorded as the zero-point temperature value. When the constant temperature bath was set to a known standard temperature value, the output value of the temperature-sensitive element was recorded and compared with the reading of the standard platinum resistance thermometer to obtain the temperature range calibration relationship. For humidity-sensitive elements, the humidity-sensitive element of the smart sensor was placed in a standard humidity environment provided by a saturated salt solution humidity generator. The output value of the humidity-sensitive element was recorded in a standard environment with a relative humidity of 0%RH as the humidity zero-point value. The output value of the humidity-sensitive element was recorded in a standard environment with a relative humidity of 100%RH and compared with the standard humidity value to obtain the humidity range calibration relationship. For metal oxide semiconductor gas sensors, zero-level air is introduced into the metal oxide semiconductor gas sensor of the smart sensor, and the sensor's output resistance value is recorded as the gas zero-point value. Then, a standard ethanol gas of known concentration is introduced, and the sensor's output resistance value is recorded again. Based on this, a gas range calibration relationship is established. For miniature spectrometers, the incident end of the miniature spectrometer is aligned with a standard diffuse reflectance white plate, and the light source is turned off to collect the dark spectrum as the spectral zero-point value. Then, the standard light source is turned on to illuminate the standard diffuse reflectance white plate, and the reflectance spectrum is collected. The collected spectral response is compared with the known reflectance spectrum of the standard diffuse reflectance white plate to establish a spectral range calibration relationship. After completing the zero-point calibration and range calibration, the calibration coefficients of each smart sensor are recorded. The calibration coefficients include temperature calibration coefficients, humidity calibration coefficients, gas calibration coefficients, and spectral calibration coefficients. The calibration coefficients are stored in the sensor parameter configuration table of the data processing unit in the form of the slope and intercept of a linear function or a piecewise linear function, respectively.

[0022] During continuous data acquisition, zero-point drift correction is automatically performed at fixed time intervals. These fixed time intervals are set to twenty minutes. Each time zero-point drift correction is performed, the intelligent sensor array exposes its sensitive elements to a sample-free environment via an internal gas path switching mechanism or with the assistance of an operator. In a sample-free environment, temperature and humidity sensors are placed in a clean air chamber isolated from the fermentation vessel. For metal oxide semiconductor gas sensors, zero-level air is introduced. For the miniature spectrometer, the external light source is turned off and the incident end is blocked to obtain dark spectral signals. In this sample-free environment, each sensitive element acquires its current output value, referred to as the current zero-point output value. The difference between the current zero-point output value and the initial zero-point value is calculated. The initial zero-point value is the zero-point output value recorded during the zero-point calibration step before the fermentation process begins. If the difference between the current zero-point output value and the initial zero-point value is not zero, the calculated difference is used as the zero-point drift compensation value. Zero-point drift compensation values ​​are used to compensate for subsequent data acquisition. Specifically, the zero-point drift compensation value for the corresponding sensing dimension is added to or subtracted from each subsequent acquisition of temperature, humidity, volatile gas concentration, and leaf reflectance spectrum data. For temperature data, the temperature zero-point drift compensation value is subtracted from subsequent temperature acquisitions; for humidity data, the humidity zero-point drift compensation value is subtracted from subsequent humidity acquisitions; for volatile gas concentration data, based on the response characteristics of the metal oxide semiconductor gas sensor, the subsequent volatile gas concentration data is compensated using a logarithmic scale combined with the gas zero-point drift compensation value; for leaf reflectance spectrum data, the intensity value of each wavelength channel in the subsequent acquired spectrum is subtracted from the dark spectrum zero-point drift compensation value.

[0023] After zero-point drift correction and compensation, the multi-source sensor data is further normalized according to each sensing dimension, so that the data value range of each dimension is mapped to the zero-to-one interval. For temperature data, the global maximum and global minimum values ​​are determined. The global maximum and global minimum values ​​are derived from the statistical values ​​of the temperature range recorded by the intelligent sensor array in the actual fermentation environment, which are the upper and lower temperature limits, respectively. and lower limit of temperature Then normalize the temperature data The calculation method is as follows:

[0024] in This represents the original temperature data after compensation. This represents the global minimum value of the temperature data. This represents the global maximum value of the temperature data.

[0025] For volatile gas concentration data, determine the global maximum value of the volatile gas concentration data. and global minimum Normalized volatile gas concentration data The calculation method is as follows:

[0026] in, This represents the original volatile gas concentration data after compensation. This represents the global minimum value of the volatile gas concentration data. This represents the global maximum value of the volatile gas concentration data.

[0027] For leaf reflectance spectral data, normalization is performed independently for each wavelength channel. A single wavelength channel is selected. Determine the global maximum value of the spectral intensity in this wavelength channel. and global minimum Normalized spectral intensity The calculation method is as follows:

[0028] in, Indicates the wavelength channel after compensation The original spectral intensity, Indicates wavelength channel The global minimum of spectral intensity, Indicates wavelength channel The global maximum value of the spectral intensity.

[0029] Optional, global maximum value of temperature data and global minimum The global maximum value of humidity data and global minimum Global maximum value of volatile gas concentration data and global minimum and the global maximum value of spectral intensity for each wavelength channel. and global minimum All data were obtained through statistical analysis of multi-source sensor data collected from the current fermentation batch after zero-point drift compensation, and the statistical scope covered all data from the start of the current fermentation process to the current sampling time.

[0030] In some implementations, the upper temperature limit The value is 45℃, which is the lower limit of the temperature range. The value is set at 15℃; upper limit of humidity. The value is taken as 100%RH, lower limit of humidity. The value is set at 50% RH; upper limit of volatile gas concentration. The value is 500 ppm, the lower limit of volatile gas concentration. The value is 0 ppm; the upper limit of spectral intensity for each wavelength channel of the leaf reflectance spectrum. The value is the saturation intensity value of the corresponding miniature spectrometer in that wavelength channel, and the lower limit of spectral intensity. The value is the minimum dark noise level for that wavelength channel of the corresponding miniature spectrometer.

[0031] The multi-source sensor data, after the above normalization process, are arranged according to the sensing dimension to form a normalized multi-source sensor data stream. In the normalized multi-source sensor data stream, the values ​​of temperature data, humidity data, volatile gas concentration data, and leaf reflectance spectrum data are all in the range of zero to one, which serves as the direct input for subsequent adaptive window segmentation and data analysis.

[0032] Example 3: In specific implementation, please refer to Figure 4 The process of adaptive window segmentation for multi-source sensor data includes calculating the short-time energy change rate, marking candidate segmentation points, filtering formal segmentation boundaries, dynamically adjusting thresholds, handling abnormal window sizes, and recalculating start and end times.

[0033] In practical implementation, the multi-source sensor data stream, after compensation and normalization processing, is used as input. This stream includes temperature, humidity, volatile gas concentration, and leaf reflectance spectrum data sequences, each arranged in ascending order of sampling time. The short-time energy change rate is calculated independently for each sensing dimension, defined as the absolute value of the difference between the sensing value at the current sampling time and the sensing value at the previous sampling time. Representing sensing dimensions At sampling time The sensor values, among which, The values ​​are indices corresponding to the temperature dimension, humidity dimension, volatile gas concentration dimension, and leaf reflectance spectrum dimension, respectively. Indicates the sequence number of the sampling time. , This represents the total number of sampling moments in the current multi-source sensor data stream. Sensing dimension. At sampling time short-time energy change rate The calculation method is as follows:

[0034] in, Representing sensing dimensions At sampling time Normalized sensor values, Representing sensing dimensions At sampling time Normalized sensor values, The value ranges from to .

[0035] In practice, when the short-time energy change rate of any sensing dimension exceeds a first threshold, the corresponding sampling time is marked as a candidate segmentation point. The initial value of the first threshold is preset according to the tea fermentation stage. In the initial fermentation stage, the first threshold is set to 0.08, and in the main fermentation stage, the first threshold is set to 0.15. The basis for setting the first threshold is that in the initial fermentation stage, the changes in multi-source sensor data are gradual, and a small short-time energy change rate can reflect the local changes in the fermentation state. Therefore, the first threshold is set to a lower value of 0.08. In the main fermentation stage, the fluctuation range of multi-source sensor data is larger. To avoid oversensitivity to noise, the first threshold is set to a higher value of 0.15.

[0036] In practice, for each candidate segmentation point, the short-time energy change rate of all sensing dimensions within a set time range before and after the candidate segmentation point is obtained. The set time range is a continuous time interval consisting of the three sampling times before the candidate segmentation point and the three sampling times after the candidate segmentation point. The short-time energy change rates of all sensing dimensions at all sampling times within the set time range are accumulated to obtain the accumulated value. If the accumulated value exceeds a second threshold, the candidate segmentation point is determined as the formal segmentation boundary. The initial value of the second threshold is preset according to the tea fermentation stage. In the initial fermentation stage, the second threshold is set to 0.60, and in the main fermentation stage, the second threshold is set to 1.20. The basis for setting the second threshold is: the accumulated value of the short-time energy change rate is generally low in the initial fermentation stage, and a second threshold of 0.60 can identify meaningful fermentation state transition points; the accumulated value of the short-time energy change rate is generally higher in the main fermentation stage, and a second threshold of 1.20 can filter out non-critical fluctuations while retaining the true state transition boundary.

[0037] In practice, the values ​​of the first and second thresholds are dynamically adjusted based on the preset minimum and maximum window durations for each tea fermentation stage. The minimum window duration is defined as the shortest allowed time between two adjacent formal segmentation boundaries, and the maximum window duration is defined as the longest allowed time between two adjacent formal segmentation boundaries. In the initial fermentation stage, the minimum window duration is set to three minutes, and the maximum window duration is set to fifteen minutes; in the main fermentation stage, the minimum window duration is set to one minute, and the maximum window duration is set to eight minutes. When the frequency of consecutive formal segmentation boundaries generated during adaptive window segmentation is too high, causing the time interval between adjacent formal segmentation boundaries to frequently fall below the minimum window duration, the values ​​of both the first and second thresholds are multiplied by an adjustment factor of 1.1 to reduce segmentation sensitivity and decrease the number of formal segmentation boundaries. When no formal segmentation boundaries are generated for an extended period during adaptive window segmentation, causing the duration of a fermentation monitoring segment to approach the maximum window duration, the values ​​of both the first and second thresholds are multiplied by an adjustment factor of 0.9 to increase segmentation sensitivity and increase the number of formal segmentation boundaries. The minimum allowable value for the first threshold is 0.02, and the maximum allowable value is 0.30; the minimum allowable value for the second threshold is 0.20, and the maximum allowable value is 2.00. When the first or second threshold exceeds its respective allowable range after adjustment, the excess value will be clamped to the corresponding minimum or maximum allowable value.

[0038] In practice, multi-source sensor data is segmented according to all formal segmentation boundaries, with the data segment between two adjacent formal segmentation boundaries forming a fermentation monitoring segment. When the time interval between three or more consecutive formal segmentation boundaries is less than the minimum window duration, the formal segmentation boundary with the smallest interval is deleted, and two adjacent fermentation monitoring segments are merged. The specific merging operation is as follows: for three consecutive formal segmentation boundaries... , , ,calculate and The time interval between as well as and The time interval between ,like and If both are less than the minimum window duration, then compare. and ,when Less than Delete formal partition boundaries ,when Less than Delete formal partition boundaries ,when equal Delete formal partition boundaries After deleting the formal dividing boundary, the original boundary will be... and The defined fermentation monitoring fragment is different from the original one. and The defined fermentation monitoring segments are merged into one fermentation monitoring segment.

[0039] In practice, when the duration of a single fermentation monitoring segment exceeds the maximum window duration, additional segmentation boundaries are inserted within that fermentation monitoring segment at equal time intervals to divide the segment into multiple sub-segments. The equal time intervals are selected as follows: the maximum integer number of minutes not exceeding the maximum window duration is used as the time step of the additional segmentation boundary. Starting from the beginning of the fermentation monitoring segment, an additional segmentation boundary is inserted every other time step until the entire fermentation monitoring segment is covered.

[0040] In practice, for each fermentation monitoring segment generated after merging or dividing, the start and end times of the fermentation monitoring segment are recalculated. The start time is the first sampling time included in the fermentation monitoring segment, and the end time is the last sampling time included in the fermentation monitoring segment. To ensure that all fermentation monitoring segments are temporally continuous and non-overlapping, all recalculated fermentation monitoring segments are arranged in ascending order of their start times. The end and start times of two adjacent fermentation monitoring segments are checked to see if they are equal. If they are not equal, the end time of the previous fermentation monitoring segment is used as the start time of the next fermentation monitoring segment. When the next fermentation monitoring segment experiences data loss as a result, linear interpolation is used to supplement the multi-source sensor data at the corresponding sampling time. Finally, all fermentation monitoring segments constitute a fermentation monitoring segment sequence arranged in chronological order. Each fermentation monitoring segment contains normalized temperature data, humidity data, volatile gas concentration data, and leaf reflectance spectrum data.

[0041] Example 4: In specific implementation, please refer to Figure 5 The process of extracting the cross-covariance matrix between multi-source sensor data and calculating the spectral entropy for a fermentation monitoring segment includes constructing a data matrix, centering, calculating the cross-covariance matrix, performing Jacobian iterative eigenvalue decomposition, obtaining the sorted eigenvalue sequence, and calculating the spectral entropy.

[0042] In practice, a fermentation monitoring segment obtained after adaptive window segmentation is used as the processing object. This segment contains temperature, humidity, volatile gas concentration, and leaf reflectance spectrum data arranged according to sampling time. The multi-source sensor data from all sampling times within the fermentation monitoring segment are arranged into a data matrix. Data matrix The rows correspond to the sensing dimensions, and the columns correspond to the sampling times. Let the total number of sensing dimensions be... , Values , representing the temperature dimension, humidity dimension, volatile gas concentration dimension, and leaf reflectance spectrum dimension, respectively, and let the number of sampling times included in the fermentation monitoring segment be . Data matrix The size is Data matrix The Line 1 Column elements Indicates the first The sensing dimension in the first... The normalized sensor values ​​at each sampling time, where The range of values ​​is arrive , The range of values ​​is arrive .

[0043] For data matrix Each row is centered, and the mean of that row is subtracted to obtain a centered data matrix. The specific method of centralized processing is as follows: calculating the data matrix. The row mean mean The calculation method is as follows ,in, Representing a data matrix The Line 1 Column elements, This indicates the number of sampling times within the fermentation monitoring segment. The range of values ​​is arrive Data matrix The Subtract the mean of the corresponding row from each element in the row. To obtain a centralized data matrix The Line 1 Column elements , The calculation method is as follows .

[0044] Centralized data matrix Multiply by its transpose matrix and then divide by the number of sampling times minus one to obtain the cross covariance matrix. Cross-covariance matrix The size is Cross covariance matrix The calculation method is as follows:

[0045] in, Represents a centralized data matrix of size . , Represents a centralized data matrix The transpose of the matrix has a size of , This indicates the number of sampling times within the fermentation monitoring segment. This represents the total number of sensing dimensions. Cross-covariance matrix. The Line 1 Column elements Indicates the first The sensing dimension and the first The covariance between each sensing dimension, where and The range of values ​​is arrive .

[0046] For the cross covariance matrix Perform Jacobi iterative eigenvalue decomposition to obtain all eigenvalues. The Jacobi iterative eigenvalue decomposition process is as follows: [The text abruptly shifts to a different topic] ...the cross covariance matrix... As the initial matrix, the maximum number of iterations is set to... The convergence accuracy is In each round of Jacobi iteration, the element with the largest absolute value among the current off-diagonal elements of the cross-covariance matrix is ​​searched. ,in, and satisfy Calculate the rotation angle Rotation angle The calculation method is as follows ,in, Indicates the current cross-covariance matrix. Line 1 Column elements, Indicates the current cross-covariance matrix. Line 1 diagonal elements of the column Indicates the current cross-covariance matrix. Line 1 Column diagonal elements. Construct a rotation matrix. Rotation matrix The size is Rotation matrix diagonal elements excluding the first Line 1 Column and number Line 1 The column values ​​are The values ​​of the remaining diagonal elements are: Non-diagonal elements except the first Line 1 The column values ​​are , No. Line 1 The column values ​​are The values ​​of the remaining non-diagonal elements are: Perform a similarity transformation on the cross-covariance matrix, and update the cross-covariance matrix as follows: ,in, Representing the rotation matrix The transpose of the matrix. Repeat the steps above to search for the element with the largest absolute value off-diagonal and the similarity transformation, until the absolute value of all off-diagonal elements is less than the convergence precision. Or reach the maximum number of iterations After the Jacobi iteration is complete, the diagonal elements of the cross covariance matrix are all the eigenvalues ​​obtained from the eigenvalue decomposition.

[0047] All that will be obtained The eigenvalues ​​are sorted in descending order of value to form an eigenvalue sequence. ,in This represents the largest eigenvalue. Represents the smallest eigenvalue. This represents the total number of sensing dimensions.

[0048] Calculate the spectral entropy of the eigenvalue sequence of the fermentation monitoring fragment. Divide each eigenvalue in the sequence by the sum of all eigenvalues ​​to obtain the normalized weight of each eigenvalue. Indicates the sorted order of the first... There are eigenvalues, among which... The range of values ​​is arrive The sum of eigenvalues The calculation method is as follows . No. Normalized weights of each eigenvalue The calculation method is as follows:

[0049] in, Indicates the sorted order of the first... 1 eigenvalue, Represents the sum of all eigenvalues. The range of values ​​is arrive .

[0050] For each normalized weight, take the logarithm to the base of the natural constant, then multiply by the normalized weight itself to obtain the entropy contribution value corresponding to each feature value. Entropy contribution value corresponding to each feature value The calculation method is as follows ,in Indicates the first Normalized weights for each feature value Represented by natural constant Logarithmic operations with base 0, The range of values ​​is arrive If normalized weights If it is zero, then it is agreed that .

[0051] The spectral entropy of the fermentation monitoring segment is obtained by summing the negative entropy contributions of all eigenvalues. Spectral entropy The calculation method is as follows:

[0052] in, Indicates the first The entropy contribution value corresponding to each feature value Represents the total number of sensing dimensions. Spectral entropy. The value of is a non-negative real number, and the spectral entropy is The numerical value of spectral entropy represents the information complexity of multi-source sensor data within this fermentation monitoring segment. The larger the value, the more complex the correlation structure between multi-source sensor data.

[0053] Example 5: In specific implementation, please refer to Figure 6 The process involves dynamically warping the spectral entropy of each fermentation monitoring segment against a pre-constructed fermentation baseline spectral entropy curve, calculating the warping distance, determining abnormal fermentation states based on the warping distance, and implementing subsequent processing measures after determining anomalies. This process includes constructing the fermentation baseline spectral entropy curve, constructing a distance matrix, searching for curved paths, applying window constraints, obtaining the warping distance, statistically analyzing the mean and standard deviation of the warping distance, setting anomaly thresholds, comparison and determination, dynamically adjusting the anomaly thresholds, and recording and processing anomalies.

[0054] In practical implementation, a fermentation baseline spectral entropy curve is pre-constructed. This curve is a time series, recording spectral entropy reference values ​​sampled chronologically under standard fermentation conditions. The construction of the baseline spectral entropy curve involves selecting multiple batches of normally fermented tea samples. Under the same standard fermentation conditions, multi-source sensor data is collected using the aforementioned intelligent sensor array. Through adaptive window segmentation, cross-covariance matrix extraction, feature decomposition, and spectral entropy calculation, the spectral entropy values ​​for each normally fermented batch at different fermentation monitoring segments are obtained. The arithmetic mean of the spectral entropy values ​​of each normally fermented batch at the same relative time position is calculated to obtain a spectral entropy reference value sequence arranged chronologically, i.e., the fermentation baseline spectral entropy curve. This curve is stored in the reference database of the data processing unit for use in each fermentation monitoring process. For new processes or new tea varieties, initial multi-source sensor data can be collected through small-batch pre-fermentation experiments to calculate an initial spectral entropy sequence as a temporary baseline curve; alternatively, the baseline spectral entropy curve can be updated in real-time during fermentation, with thresholds dynamically adjusted based on expert experience to gradually improve the baseline library.

[0055] In practical implementation, a distance matrix is ​​constructed between the spectral entropy time series of a current fermentation monitoring segment and the fermentation baseline spectral entropy curve. Let the spectral entropy time series of the current fermentation monitoring segment be... ,in, Indicates the number of fragments in the current fermentation monitoring sequence. Each spectral entropy value Indicates the length of the current fermentation monitoring fragment sequence. The range of values ​​is arrive Let the fermentation baseline entropy curve be... ,in, The first spectral entropy curve in the fermentation reference spectrum represents the... Individual spectral entropy reference value, This indicates the length of the fermentation baseline entropy curve. The range of values ​​is arrive The size of the structure is Distance matrix Distance matrix The Middle Line 1 Column elements This represents the absolute difference between the spectral entropy values ​​at corresponding positions in the two sequences. The calculation method is as follows:

[0056] in, Representing the spectral entropy time series The Middle Each spectral entropy value Indicates the fermentation reference spectrum entropy curve The Middle Individual spectral entropy reference value, The range of values ​​is arrive , The range of values ​​is arrive .

[0057] In practical implementation, in the distance matrix Search for a curved path with the minimum cumulative distance. Curved path Defined as by A sequence of coordinate points The first component of each coordinate point Second component Corresponding distance matrices Row index and column index, This represents the length of the curved path. (Curved path) From the distance matrix top left corner Departure, to the distance matrix bottom right corner End, boundary conditions satisfied. as well as Curved path The search allows starting from the current coordinates. Move horizontally to Move vertically to Or move along the diagonal direction to Curved path The cumulative distance is defined as the distance value corresponding to all coordinate points on the path. The cumulative sum. Dynamic time warping calculates the cumulative distance matrix in the distance matrix using dynamic programming. Cumulative distance matrix The Line 1 Column elements Indicates starting from the origin Reaching coordinates The minimum cumulative distance. Cumulative distance matrix. The recursive calculation method is as follows: For the first row element, For the first Column elements, For the remaining elements, After the recursion is completed, This is the cumulative distance to the end of the curved path.

[0058] In practical implementation, when performing dynamic time warping, a window constraint is applied to the search for curved paths, limiting the curved paths within the distance matrix. The offset on both sides of the center diagonal does not exceed the preset curved window width. The curved window width is set to... The width of the curved window is set to a value. The basis for this is that the local scaling amplitude of the spectral entropy time series and the fermentation baseline spectral entropy curve on the time axis is expected to be within the range of adjacent time windows, and the width of the curved window is [missing information]. It can allow local time alignment while avoiding excessive path distortion. The window constraint is applied by recursively calculating the cumulative distance matrix. At that time, only for those that satisfy coordinates calculate ,for The coordinates of the points are not involved in the path search.

[0059] In practical implementation, the cumulative distance at the end of the curved path will be... As the normalized distance between the current fermentation monitoring segment and the fermentation baseline spectral entropy curve, it is used express.

[0060] In practice, the comparison between the regularized distance and a preset anomaly threshold determines whether the tea corresponding to the current fermentation monitoring segment is in an abnormal fermentation state. The mean and standard deviation of the regularized distances of all completed fermentation monitoring segments are calculated. Let the number of completed fermentation monitoring segments be... The normalization distances corresponding to each completed fermentation monitoring segment are as follows: Mean of normalized distance The calculation method is as follows Standard deviation of regular distance The calculation method is as follows Set the outlier threshold to the mean plus twice the standard deviation. The calculation method is as follows .

[0061] In practice, the normalization distance of the current fermentation monitoring segment is compared with the anomaly threshold. If the normalization distance of the current fermentation monitoring segment... Greater than the abnormal threshold If the current tea leaves are in an abnormal fermentation state, the time interval of the fermentation monitoring segment is recorded, including the start and end sampling times of the fermentation monitoring segment. If the normalization distance of the current fermentation monitoring segment... Less than or equal to the abnormal threshold If so, it is determined that the tea leaves are currently in a normal fermentation state.

[0062] In practice, when multiple consecutive fermentation monitoring segments are identified as abnormal fermentation states, the abnormality threshold is dynamically adjusted to the current mean plus three times the standard deviation. The condition for multiple consecutive segments is set as three consecutive fermentation monitoring segments. When the normalized distance of three consecutive fermentation monitoring segments is greater than the current abnormality threshold, the mean of the normalized distances of all fermentation monitoring segments completed from the start of fermentation to the current time is recalculated. and standard deviation The abnormal threshold will be dynamically updated to The abnormal threshold is dynamically updated and then applied to the anomaly detection of subsequent fermentation monitoring segments.

[0063] In practice, after determining whether the tea corresponding to the current fermentation monitoring segment is in an abnormal fermentation state, anomaly recording and processing are performed on the fermentation monitoring segments determined to be in an abnormal fermentation state. The original multi-source sensor data, the corresponding normalization distance, and the anomaly determination result of the fermentation monitoring segment determined to be in an abnormal fermentation state are saved to the anomaly record database. The original multi-source sensor data refers to the temperature data sequence, humidity data sequence, volatile gas concentration data sequence, and leaf reflectance spectrum data sequence within the fermentation monitoring segment before compensation and normalization processing. The anomaly determination result includes the anomaly determination time, the normalization distance value, and the anomaly threshold used.

[0064] In practice, the system retrieves historical anomalous segments from the anomaly record database that exhibit similar multi-source sensor data change patterns to the current anomalous fermentation monitoring segment. The retrieval method for similar multi-source sensor data change patterns involves extracting the trend curves for temperature, humidity, volatile gas concentration, and reflectance of specific bands in the leaf reflectance spectrum within the current anomalous fermentation monitoring segment. Pearson correlation coefficients are calculated between these four trend curves and their corresponding curves in the anomaly record database for each historical anomalous segment. The average of the four Pearson correlation coefficients is then calculated. If the average value is greater than 0.85, the historical anomalous segment is considered to have a similar multi-source sensor data change pattern to the current anomalous fermentation monitoring segment. Subsequent fermentation treatment measures for the retrieved historical anomalous segments are then extracted. These measures refer to the temperature, humidity, and ventilation adjustments taken to correct the fermentation anomaly when the historical anomalous segment occurred. Adjustment instructions are generated based on the extracted subsequent fermentation treatment measures. The content of the adjustment instructions includes the target temperature setpoint of the temperature regulator, the target humidity setpoint of the humidity regulator, and the ventilation rate setpoint of the ventilation device. The adjustment instructions are sent to the temperature regulator, humidity regulator, and ventilation device of the fermentation device through the control bus, and the temperature regulator, humidity regulator, and ventilation device execute the corresponding actions.

[0065] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for monitoring fermentation in tea processing based on intelligent sensors, characterized in that, include: Acquire multi-source sensor data continuously collected by an intelligent sensor array during the tea fermentation process. The multi-source sensor data includes temperature data, humidity data, volatile gas concentration data, and leaf reflectance spectrum data. The multi-source sensor data is subjected to adaptive window segmentation to obtain multiple fermentation monitoring segments arranged in chronological order; Extract the cross-covariance matrix between multi-source sensor data within each fermentation monitoring segment, and perform eigenvalue decomposition on the cross-covariance matrix to obtain an eigenvalue sequence; Calculate the spectral entropy of the feature value sequence for each fermentation monitoring segment, whereby the spectral entropy is used to characterize the information complexity of the fermentation state; Dynamic time warping is performed on the spectral entropy of each fermentation monitoring segment and the pre-constructed fermentation baseline spectral entropy curve, and the warped distance between each fermentation monitoring segment and the fermentation baseline spectral entropy curve is calculated. Based on the comparison between the regular distance and the preset abnormal threshold, it is determined whether the tea corresponding to the current fermentation monitoring segment is in an abnormal fermentation state.

2. The method for monitoring fermentation in tea processing based on intelligent sensors according to claim 1, characterized in that, The acquisition of multi-source sensor data continuously collected by the intelligent sensor array during the tea fermentation process specifically includes: A smart sensor array is deployed at multiple preset locations in the fermentation vessel. Each smart sensor integrates a temperature-sensitive element, a humidity-sensitive element, a metal oxide semiconductor gas sensor, and a miniature spectrometer. All smart sensors are set to collect data synchronously at the same sampling frequency. The sampling frequency is dynamically adjusted according to the tea fermentation stage. The initial fermentation stage uses a first sampling frequency, and the main fermentation stage uses a second sampling frequency that is higher than the first sampling frequency. The multi-source sensor data collected at each sampling time is timestamped and outliers that exceed the sensor's range are filtered out to form a continuous multi-source sensor data stream.

3. The method for monitoring fermentation in tea processing based on intelligent sensors according to claim 1, characterized in that, The multi-source sensor data is adaptively windowed to obtain multiple fermentation monitoring segments arranged in chronological order, specifically including: Calculate the short-time energy change rate for each sensing dimension in the multi-source sensing data, where the short-time energy change rate is the absolute value of the difference between the sensing values ​​at the current time and the previous time. When the short-time energy change rate of any sensing dimension exceeds the first threshold, that moment is marked as a candidate segmentation point; The short-time energy change rates of all sensing dimensions within a set time range before and after the candidate segmentation point are accumulated. If the accumulated value exceeds the second threshold, the candidate segmentation point is determined as the formal segmentation boundary. The multi-source sensor data is segmented according to all formal segmentation boundaries, and the data segment between two adjacent formal segmentation boundaries is used as a fermentation monitoring segment.

4. The method for monitoring fermentation in tea processing based on intelligent sensors according to claim 3, characterized in that, During the adaptive window segmentation process, the values ​​of the first threshold and the second threshold are dynamically adjusted according to the preset minimum and maximum window durations for the tea fermentation stage.

5. The method for monitoring fermentation in tea processing based on intelligent sensors according to claim 1, characterized in that, The step of extracting the cross-covariance matrix among multi-source sensor data within each fermentation monitoring segment and performing eigenvalue decomposition on the cross-covariance matrix to obtain an eigenvalue sequence specifically includes: For a fermentation monitoring segment, the multi-source sensor data of all sampling times contained therein are arranged into a data matrix, with the rows of the data matrix corresponding to the sensing dimensions and the columns corresponding to the sampling times; Center each row of the data matrix and subtract the mean of that row to obtain a centered data matrix; Multiply the centered data matrix by its transpose, and then divide by the number of sampling times minus one to obtain the cross covariance matrix; Perform Jacobi iterative eigenvalue decomposition on the cross covariance matrix to obtain all eigenvalues, and sort the eigenvalues ​​in descending order of value to form an eigenvalue sequence.

6. The method for monitoring fermentation in tea processing based on intelligent sensors according to claim 1, characterized in that, The calculation of the spectral entropy of the feature value sequence of each fermentation monitoring segment specifically includes: Divide each feature value in the feature value sequence by the sum of all feature values ​​to obtain the normalized weight of each feature value; Take the logarithm of each normalized weight to the base of the natural constant, and then multiply it by the normalized weight itself to obtain the entropy contribution value corresponding to each feature value. The spectral entropy of the fermentation monitoring segment is obtained by summing the negative entropy contribution values ​​corresponding to all feature values.

7. The method for monitoring fermentation in tea processing based on intelligent sensors according to claim 1, characterized in that, The step of performing dynamic time normalization on the spectral entropy of each fermentation monitoring segment and the pre-constructed fermentation baseline spectral entropy curve, and calculating the normalized distance between each fermentation monitoring segment and the fermentation baseline spectral entropy curve, specifically includes: The fermentation baseline spectral entropy curve is constructed as a time series, which records the spectral entropy reference values ​​sampled in chronological order under standard fermentation process; For the spectral entropy time series of the current fermentation monitoring segment and the fermentation baseline spectral entropy curve, a distance matrix is ​​constructed, where each element of the matrix is ​​the absolute difference of the spectral entropy values ​​at corresponding positions in the two sequences. Search for a curved path with the minimum cumulative distance from the top left corner to the bottom right corner in the distance matrix. The path search allows movement in the horizontal, vertical and diagonal directions. The cumulative distance of each step is the absolute difference of the current step plus the minimum cumulative distance of the previous step. The cumulative distance at the end of the curved path is used as the normalized distance between the current fermentation monitoring segment and the fermentation baseline spectral entropy curve.

8. The method for monitoring fermentation in tea processing based on intelligent sensors according to claim 7, characterized in that, When performing dynamic time warping, a window constraint is applied to the search of curved paths, limiting the offset of the path on both sides of the diagonal in the distance matrix to no more than the preset curved window width.

9. The method for monitoring fermentation in tea processing based on intelligent sensors according to claim 1, characterized in that, Based on the comparison between the regularized distance and the preset anomaly threshold, it is determined whether the tea corresponding to the current fermentation monitoring segment is in an abnormal fermentation state, specifically including: The mean and standard deviation of the regular distances of all completed fermentation monitoring segments were statistically analyzed. The anomaly threshold is set to the mean plus twice the standard deviation; If the regularity distance of the current fermentation monitoring segment is greater than the abnormal threshold, the tea is determined to be in an abnormal fermentation state, and the time interval of the fermentation monitoring segment is recorded. If the regularity distance of the current fermentation monitoring segment is less than or equal to the abnormal threshold, then the tea is determined to be in a normal fermentation state.

10. The method for monitoring fermentation in tea processing based on intelligent sensors according to claim 9, characterized in that, When multiple consecutive fermentation monitoring segments are identified as being in an abnormal fermentation state, the abnormality threshold is dynamically adjusted to the current mean plus three times the standard deviation to improve the specificity of the abnormality determination.

Images