A large model-based multivariate collaborative drying control method and system

Abstract

The application discloses a multivariate cooperative drying control method and system based on a large model. Working condition data of a drying cavity are collected and subjected to synchronous normalization processing to establish a working condition benchmark matrix, temperature and humidity coupling deviation and heat and mass cross disturbance characteristics are identified, and standard operation parameters are formed through variable differentiation calibration; the standard operation parameters are input into a large model to infer and obtain inference decision weights, the working condition benchmark matrix is weighted and mapped to generate a coupling response graph; a drying joint regulation benchmark is established based on the coupling response graph, multibatch difference detection and working condition disturbance elimination are performed to determine a regulation credible boundary, and an abnormal working condition is screened to form a control deviation record group; the standard operation parameters and the drying joint regulation benchmark are subjected to time sequence correlation analysis to obtain a regulation iteration factor, the control deviation record group is subjected to drying grade labeling according to the regulation iteration factor, and finally, a temperature, humidity and wind cooperative control instruction is output, so that adaptive cooperative control of a grain drying process under multivariate coupling dynamics is realized.

Description

Technical Field

[0001] This invention relates to the field of agricultural intelligent equipment control technology, and in particular to a multivariate collaborative drying control method and system based on a large model. Background Technology

[0002] In grain drying operations, the three physical quantities of temperature, humidity, and ventilation within the drying chamber are interdependent. Adjusting any one of these quantities will trigger a chain reaction in the others. Simultaneously, the migration rate of moisture content within the grain grains continuously decreases as the drying process progresses, resulting in drastically different effects of the same process control strategy in the early and later stages of drying. Existing control systems often rely on fixed thresholds or empirical parameters to independently adjust single variables, failing to capture the dynamic coupling relationship between multiple variables as moisture content changes. Poor coordination between hot air supply and dehumidification operations easily leads to problems such as localized over-drying or uneven drying.

[0003] Due to differences in variety, initial moisture content, and stacking conditions, the heat and mass transfer characteristics within the grain cavity vary significantly between batches, making it difficult to maintain the effectiveness of process control benchmarks established for a single batch in subsequent batches. Existing methods lack a systematic identification and correction mechanism for inter-batch operational deviations, resulting in a gradual deterioration in control accuracy with the accumulation of batches, ultimately making it difficult to consistently meet quality standards. Summary of the Invention

[0004] This invention discloses a multivariate collaborative drying control method and system based on a large model. By normalizing the cavity operating data and identifying multivariate coupling offsets, standard operating parameters reflecting the dynamic characteristics of heat and mass transfer are established. With the help of a large model, a comprehensive reasoning is performed on the moisture content gradient and the temperature-humidity-air coupling state to generate a control benchmark. Combined with multi-batch difference detection and operating condition disturbance elimination mechanism, control deviations are identified. Finally, temperature-humidity-air collaborative control commands are output to realize multivariate adaptive collaborative control of the grain drying process.

[0005] The first aspect of this invention proposes a multivariate collaborative drying control method based on a large model, comprising the following steps: The operating condition data of the drying chamber, including moisture content, temperature, humidity and air volume, are collected, and the operating condition data are synchronously normalized to establish an operating condition reference matrix. The temperature and humidity coupling offset of the operating condition reference matrix is ​​identified to determine the thermal mass compensation weight. Based on the operating condition reference matrix, thermal mass cross-disturbance is identified to determine the disturbance attenuation parameter. Based on the disturbance attenuation parameter and the thermal mass compensation weight, variable differentiation calibration is performed to form standard operating parameters. The standard operating parameters are input into the large model to perform comprehensive reasoning on the coupling state of moisture content gradient and temperature, humidity and wind to obtain reasoning decision weights. Based on the reasoning decision weights, the operating condition benchmark matrix is ​​weighted and mapped to generate a coupling response spectrum. A drying joint control benchmark is established based on the coupled response spectrum. Multiple batch difference detections are performed based on the drying joint control benchmark to determine the operating deviation amplitude. The control confidence boundary after eliminating operating condition disturbances is determined based on the operating deviation amplitude. Abnormal operating conditions are screened from the operating condition benchmark matrix according to the control confidence boundary to form a control deviation record group. The standard operating parameters are correlated with the drying joint control benchmark by time series analysis to obtain the control iteration factor. The control deviation record group is corrected according to the control iteration factor, and the drying level is marked and the temperature, humidity and wind coordinated control command is output.

[0006] A second aspect of this invention proposes a multivariable collaborative drying control system based on a large model, comprising: The data acquisition module is used to collect operating condition data of the drying chamber, including moisture content, temperature, humidity and air volume, and to perform synchronous normalization processing on the operating condition data to establish an operating condition reference matrix. The parameter calibration module is used to identify the temperature and humidity coupling offset of the operating condition reference matrix to determine the thermal mass compensation weight, identify thermal mass cross-disturbance based on the operating condition reference matrix to determine the disturbance attenuation parameter, and perform variable differential calibration based on the disturbance attenuation parameter and the thermal mass compensation weight to form standard operating parameters. The reasoning and decision module is used to input the standard operating parameters into the large model to perform comprehensive reasoning on the coupling state of moisture content gradient and temperature, humidity and wind to obtain reasoning and decision weights, and to perform weighted mapping on the working condition benchmark matrix according to the reasoning and decision weights to generate a coupling response spectrum. The deviation detection module is used to establish a drying joint control benchmark based on the coupled response spectrum, perform multi-batch difference detection based on the drying joint control benchmark to determine the operating deviation amplitude, determine the control confidence boundary after eliminating operating condition disturbances based on the operating deviation amplitude, and select abnormal operating conditions from the operating condition benchmark matrix to form a control deviation record group based on the control confidence boundary. The instruction output module is used to perform time-series correlation analysis between the standard operating parameters and the drying joint control benchmark to obtain the control iteration factor, and to correct the control deviation record group according to the control iteration factor, and to output the temperature, humidity and wind coordinated control instruction for drying level labeling.

[0007] The beneficial effects of this invention are reflected in the following points: 1. By eliminating dimensional differences through synchronous normalization processing, identifying the attenuation law of temperature and humidity coupling offset and heat-mass cross-disturbance, and establishing standard operating parameters by combining moisture content sensitivity ranking and stage adaptive benchmark mapping, the control parameters can accurately reflect the dynamic changes of heat and mass transfer in different drying stages, solving the problem of insufficient adaptability of fixed parameter calibration under variable moisture content conditions. 2. After the standard operating parameters are structured and encoded, they are input into a large model to perform comprehensive reasoning on the moisture content gradient and temperature-humidity-wind coupling state, obtain the reasoning decision weights of the fusion gradient response and coupling strength, generate a coupling response spectrum through weighted mapping, and establish a drying joint control benchmark, realizing an accurate description of the optimal synergistic state of heat and mass in different drying stages. 3. By eliminating systematic deviations between batches through multi-batch difference detection and thermal inertia compensation, establishing a control credibility boundary by combining material stratification parameters, extracting control iteration factors by using moisture saturation critical identification and iterative convergence analysis, and completing the collaborative output of temperature-humidity-wind coordinated control commands after hierarchical labeling of control deviation record groups, the control strategy is continuously and adaptively corrected as batches accumulate. Attached Figure Description

[0008] Figure 1 This is a schematic flowchart of a multivariate collaborative drying control method based on a large model according to the present invention.

[0009] Figure 2 This is a structural block diagram of a multivariable collaborative drying control system based on a large model according to the present invention. Detailed Implementation

[0010] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0011] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0012] The technical solutions of the embodiments of this application will be described below.

[0013] like Figure 1 As shown, this embodiment of the invention provides a multivariate collaborative drying control method based on a large model, including the following steps S110-S150: Step S110: Collect operating condition data of the drying chamber, including moisture content, temperature, humidity and air volume, and perform synchronous normalization processing on the operating condition data to establish an operating condition reference matrix.

[0014] Specifically, operating data of the drying chamber is collected. The spatial layout of the sensors determines the depth of coverage of the internal state of the chamber by the operating data. The moisture content sensor is placed on the discharge side, where the instantaneous moisture content of the grain before it leaves the machine is measured after passing through the drying zone, serving as the final basis for determining whether the target moisture content has been reached. Temperature and humidity sensors are arranged at equal intervals along the longitudinal direction of the chamber. The bottom sensors sense the intensity of hot air on the inlet side, and the top sensors sense the degree of humidity accumulation near the exhaust port. The air volume sensor is installed on the cross-section of the exhaust channel, continuously monitoring the airflow discharged from the chamber to reflect the intensity of moisture exchange between the chamber and the outside environment. The sampling interval of each sensor is set according to the difference in the dynamic response speed of the measured quantity. The sampling intervals for temperature and humidity and air volume are shorter to capture rapid fluctuations, while the sampling interval for moisture content is longer to match the slow rhythm of grain moisture migration. If a sensor continuously misses more than two sampling intervals during the operating data collection period, an interruption mark is marked for the corresponding time period. The operating data in the interrupted section is not replaced with a zero value to avoid misinterpreting the missing signal as a low value response. When the chamber is in the feeding or discharging transition phase, the introduction of a new batch of wet grain causes a sudden increase in grain moisture detected by the moisture content sensor on the discharging side. This sudden change is a normal response to the material layer switching rather than a sensor malfunction. During the transition phase, the operating data is marked with phase boundary labels. During normalization, this segment uses independent statistical parameters and is processed separately from the stable operation segment. Extremely high moisture content values ​​at the initial feeding stage are not included in the statistical calculation of the normalization benchmark for the stable operation segment, avoiding compressing the numerical resolution of the subsequent normal operation segment by instantaneously high-moisture conditions. The operating data undergoes preliminary validity verification at the acquisition side. Single samples exceeding the sensor's range limit are discarded and replaced by interpolation of adjacent sample values.

[0015] A working condition reference matrix is ​​established by synchronously normalizing the operating condition data. The core constraint of synchronous normalization is that all signals must be aligned at the same time node before they can participate in matrix integration. The time axis corresponding to the lowest sampling frequency in each operating condition data is taken as a common time reference. High-frequency signals are downsampled according to the common time reference and the mean value within the window is obtained. Low-frequency signals are linearly interpolated and completed on the common time reference. The alignment operation eliminates the time misalignment caused by the difference in sampling frequency among the operating condition data. Each aligned signal is normalized to its maximum and minimum values ​​within the batch. The normalization formula is x_norm=(x-x_min) / (x_max-x_min), where x is the original sampled value, x_min and x_max are the minimum and maximum values ​​of the corresponding channels during the effective time period of the current batch, and x_norm is the normalized value ranging from 0 to 1 with dimensions eliminated. Extreme values ​​are statistically derived from the operating data of the effective time period of the current batch. Time periods marked as missing or transitional are skipped and not included in the extreme value statistics. In winter batches, due to the low initial cavity temperature, the temperature during the preheating stage is much lower than that of the normal drying stage, which widens the temperature range of the entire batch. After normalization, the temperature values ​​of the stable drying period are concentrated in the higher quantile range, which intuitively reflects that the temperature of this segment is in the high intensity zone within the entire temperature range of the batch. The normalized values ​​are directly comparable to those of the same drying segment in summer batches. After normalization, the timing sequences of each signal are integrated side-by-side to form a baseline matrix. The numerical range of each variable in the baseline matrix is ​​compressed to the interval between 0 and 1, and the amplitudes of multiple variables are directly comparable at the matrix level, completely eliminating dimensional differences. Stage boundary markings are retained in the baseline matrix, and the row ranges of stable operation segments and transition segments are independently distinguished. The number of rows in the baseline matrix strictly corresponds to the effective acquisition duration of the current batch.

[0016] Step S120: Identify the temperature and humidity coupling offset of the operating condition reference matrix to determine the thermal mass compensation weight; identify thermal mass cross-disturbance based on the operating condition reference matrix to determine the disturbance attenuation parameter; and perform variable differentiation calibration based on the disturbance attenuation parameter and the thermal mass compensation weight to form standard operating parameters.

[0017] In some embodiments, identifying the temperature and humidity coupling offset of the operating condition reference matrix to determine the heat and mass compensation weights includes: extracting temperature, humidity, and wind time-series curves from the operating condition reference matrix to construct a temperature, humidity, and wind joint response map; performing hysteresis disturbance analysis on the temperature, humidity, and wind joint response map to extract hysteresis response windows and air volume disturbance identifiers; dividing the coupling offset segments based on the air volume disturbance identifiers and the hysteresis response windows to generate an offset intensity sequence; and performing heat and mass transfer correlation analysis on the offset intensity sequence to map and generate heat and mass compensation weights.

[0018] Temperature, humidity, and airflow time-series curves were extracted from the operating condition baseline matrix to construct a joint response diagram. The normalized time series of temperature, humidity, and airflow within the stable operating range of the operating condition baseline matrix were extracted independently. These three time series were overlaid on the same time coordinate system, and the values ​​of the three variables at each moment constituted the joint state point for that moment. All state points were continuously arranged along the time axis to form the skeleton of the joint response diagram. The synchronous changes between state points were labeled according to coupling type. When the operator reduced the airflow to increase the cavity temperature, the temperature rose while the airflow decreased synchronously, labeled as a negative coupling event. When the dehumidification valve was fully open, the airflow reached its peak while the cavity humidity plummeted, labeled as a strong dehumidification linkage event. After the grain entered the slow-drying stage, the resistance to the migration of bound water increased, and the operator frequently switched between heating and dehumidification. The frequency of these two types of events alternating was significantly higher than in the initial constant-rate stage, and the thermo-mass coupling relationship of the cavity became more complex as the moisture content decreased. The temporal resolution of the temperature-humidity-air joint response diagram is consistent with the common sampling benchmark of the operating condition baseline matrix. When the resolution is insufficient to distinguish between rapid airflow disturbances and temperature lag responses, local cubic spline interpolation upsampling is performed on the temperature curves for the corresponding time periods. The interpolation is performed within the original accuracy range of the operating condition baseline matrix to avoid introducing additional measurement errors. After interpolation, the coupling category of the joint state point label is re-verified to see if the category changes due to the increased resolution. If a change occurs, the high-resolution result after interpolation shall prevail. Step changes caused by material layer switching in the transition section are not included in the scope of the temperature-humidity-air joint response diagram.

[0019] Hysteresis disturbance analysis was performed on the temperature, humidity, and wind joint response diagram to extract the hysteresis response window and the outlet air volume disturbance identifier. Abrupt change points in the outlet air volume curve of the temperature, humidity, and wind joint response diagram were located using first-order difference extrema. Points where the absolute value of the difference exceeded twice the time series standard deviation of the outlet air volume were identified as candidate disturbance points. The disturbance direction of the candidate disturbance points was determined by the sign of the difference. A positive difference corresponds to a sudden increase in outlet air volume, triggering a verification judgment of negative coupling events in the temperature, humidity, and wind joint response diagram; a negative difference corresponds to a sudden decrease, triggering a verification judgment of strong humidity release linkage events. These two types of directions trigger different temperature and humidity response direction judgment rules in the offset segment division. Starting from each candidate disturbance point, a fixed time window is extended backward within the temperature, humidity, and air combined response map. In cavities with long hot air ducts or high grain density, the airflow requires a longer transmission time from generation to affecting sensor readings. The difference between the starting time of a significant response in the temperature or humidity curve within the window and the candidate disturbance point is the measured lag time. The median of the measured lag times for all candidate disturbance points is determined as the standard width of the lag response window. In a batch of corn drying operations, the measured lag times of 12 candidate disturbance points ranged from 18 to 35 seconds. Extreme lag events usually correspond to local grain agglomeration blocking the airflow channel. The median of 24 seconds excludes this type of anomaly and is determined as the lag response window width for this batch. The airflow disturbance indicator consists of three elements: the time location of the candidate disturbance point, the direction of the disturbance, and the magnitude of the absolute value of the difference.

[0020] An offset intensity sequence is generated by dividing the coupling offset segments based on the airflow disturbance identifier and the hysteresis response window. Offset response segments are delineated forward from the time position of each disturbance point of the airflow disturbance identifier, with the width of the hysteresis response window as the span. Temperature and humidity changes within a segment are considered coupled responses driven by that airflow disturbance; temperature and humidity changes outside the segment boundary are not included in the scope of that disturbance event. When the distance between two adjacent disturbance points is less than the width of the hysteresis response window, the segment for the subsequent disturbance point starts from the end of the previous segment to avoid overlapping segments and confusion in coupling response attribution. The coupling strength of each segment is measured using the normalized cross-correlation coefficient R. Where ΔT_i and ΔH_i are the temperature and humidity changes at the i-th time point within the segment, respectively, and R represents the positive time-varying temperature and humidity. The larger the absolute value, the tighter the coupling. Candidate disturbance points with larger disturbance amplitudes in the airflow disturbance identifier usually correspond to segments with higher R values. When the airflow suddenly increases during the high-temperature and low-humidity phase of a certain batch, the corresponding segment's R value reaches 0.91, while when the airflow increases slowly during the low-temperature and high-humidity phase, the corresponding segment's R value is only 0.43. The difference between the two confirms the driving effect of disturbance intensity on the depth of heat-mass coupling. Segments with R values ​​below 0.2 are identified as weakly coupled segments. When grain is too dry or becomes damp and clumps together, the airflow bypasses and forms a dead zone. Evaporation at a certain height layer of the cavity almost stops, and temperature and humidity changes no longer respond in tandem with airflow disturbances. Weakly coupled segments are retained with effective markers in the offset intensity sequence for subsequent manual verification and are not included in the mean and variance statistics to avoid abnormal operating conditions lowering the heat-mass transfer activity assessment of the corresponding stage. The offset intensity sequence is composed of the R values ​​of all offset response segments arranged in chronological order.

[0021] Heat and mass transfer correlation analysis was performed on the offset intensity sequence to generate heat and mass transfer compensation weights. The offset intensity sequence exhibits an overall decreasing trend from high to low on the time axis. A steep slope in the offset intensity sequence corresponds to a low initial moisture content of the grain and rapid progress in the drying process, while a gentle slope corresponds to continuous and vigorous evaporation in batches with high moisture content. Abrupt changes in the slope of the offset intensity sequence usually coincide with the switching points of the drying stages. After segmenting the offset intensity sequence according to the drying stages, the mean μ_s and variance σ_s² of each segment were extracted. In the constant-rate stage segment with vigorous evaporation, the mean R value is high, indicating a large heat and mass transfer intensity, and the corresponding temperature compensation weight component w_T should be higher. After entering the deceleration stage, evaporation weakens and the dispersion of R values ​​at different times increases, and the corresponding humidity compensation weight component w_H should be reduced to reflect the deteriorating regularity of moisture transfer in this stage. The temperature compensation weight component w_T is determined by mean normalization mapping: w_T = 0.5 + 0.5 × (μ_s - μ_min) / (μ_max - μ_min), where μ_s is the mean of the current stage, and μ_min and μ_max are the minimum and maximum values ​​of the mean of each stage in the whole batch. The value of w_T ranges from 0.5 to 1.0. The humidity compensation weight component w_H is determined by variance reciprocal normalization. The normalization formula is w_H = 1 - (σ_s² - σ_min²) / (σ_max² - σ_min²), where σ_s² is the variance of the current stage, and σ_min² and σ_max² are the minimum and maximum values ​​of the variance of each stage in the whole batch. The value of w_H ranges from 0 to 1. The smaller σ_s² is, the higher w_H is. The two components together constitute the thermal mass compensation weight vector for this stage. In the mid-stage of a certain batch, μ_s was 0.78 and σ_s² was below 0.03, with w_T mapped to 0.89 and w_H to 0.94. In the final stage, μ_s decreased to 0.31, σ_s² increased to 0.11, w_T decreased to 0.65, and w_H decreased to 0.62. The difference in the thermal mass compensation weight values ​​between the two stages directly reflects the systematic decrease in the activity of thermal mass transfer during the drying process. Stages with high mean and small variance show strong and regular thermal mass transfer intensity, corresponding to both components of the thermal mass compensation weight being high. Stages with low mean or large variance show at least one component decreasing to accurately reflect the deterioration of coupling quality in that stage. The thermal mass compensation weight vectors for each stage of the entire batch are concatenated in stage order to form a complete thermal mass compensation weight. The vector component values ​​at the boundaries of each stage transition linearly to eliminate weight jumps during stage switching.

[0022] Based on the operating condition baseline matrix, thermo-mass cross-perturbation identification is performed to determine perturbation attenuation parameters. The temperature, humidity, and air volume columns in the operating condition baseline matrix are extracted to form a multi-dimensional perturbation analysis space. The first-order difference extreme value of the air volume is used to locate the abrupt change anchor point. The temperature and humidity response patterns within a hysteresis response window before and after each anchor point are classified and identified. The response patterns are divided into three categories: monotonically decaying, oscillating convergence, and step jump. When the grain moisture content is high, the cavity heat capacity is large. After the air volume is disturbed, the temperature drops smoothly without rebounding, exhibiting a monotonic decay type. The decay rate is λ=|ΔT| / Δt, where ΔT is the normalized change in temperature within the segment (dimensionless), Δt is the corresponding time in seconds, and λ is 1 / second, reflecting the normalized intensity of temperature decay per unit time. When the moisture content drops to the middle range, the cavity heat capacity decreases, but evaporation still has capacity. After a sudden increase in air volume, the temperature first drops and then rises, forming an oscillating convergence. The oscillating convergence type decay rate decreases periodically from the peak-to-valley amplitude, with the decay ratio r_k=A_{k+1} / A The _k estimate is used, where A_k is the peak-to-trough amplitude of the k-th period's oscillation (normalized dimensionless value), and A_{k+1} is the peak-to-trough amplitude of the (k+1)-th period's oscillation. A smaller r_k indicates faster oscillation decay. During the mid-stage of grain deceleration, when the airflow increases sharply by 20%, the temperature exhibits an oscillating convergence response. The decay ratios for four consecutive periods are 0.61, 0.58, 0.63, and 0.60, with an average of approximately 0.60. It takes approximately six periods for the amplitude to decay to 10% below the initial disturbance amplitude. This decay duration, together with the corresponding disturbance amplitude, constitutes the disturbance decay parameter description pair for this event. Step-jump events are marked as non-steady-state disturbances and do not participate in the decay rate statistics. The occurrence time and amplitude of step-jump events are recorded in the auxiliary field of the disturbance decay parameter. The entire batch of anchor point events are divided into three amplitude levels: slight, moderate and strong, according to the magnitude of the sudden change in air volume. The average attenuation rate and the average attenuation duration within each amplitude level are extracted as the disturbance attenuation parameters for that level. The boundary of the amplitude level division is determined by the ternary digit of the air volume column of the current batch operating condition reference matrix.

[0023] In some embodiments, the step of forming standard operating parameters by differentially calibrating variables based on the disturbance attenuation parameters and the thermal mass compensation weights includes: grouping the disturbance attenuation parameters by variable type to generate variable sensitivity weights; sorting the variable sensitivity weights by moisture content sensitivity to generate a stage switching signal; performing stage adaptive benchmark mapping based on the stage switching signal and the variable sensitivity weights to form a differential calibration benchmark; and integrating the differential calibration benchmark and the thermal mass compensation weights to form standard operating parameters.

[0024] The disturbance attenuation parameters are grouped by variable type to generate variable sensitivity weights. The mean attenuation rate of each amplitude level within the disturbance attenuation parameters is split according to the response variable. The mean attenuation rate of the temperature response triggered by the airflow disturbance is assigned to the temperature sensitivity group, and the mean attenuation rate of the humidity response triggered by the disturbance attenuation parameters is assigned to the humidity sensitivity group. A faster attenuation rate indicates a higher efficiency and stronger sensitivity of the variable in responding to airflow disturbances. The mean attenuation rate of each amplitude level temperature sensitivity group is normalized and mapped to the 0-1 interval to obtain the temperature dimension variable sensitivity weight component. The humidity sensitivity group is processed similarly to obtain the humidity dimension component. These two dimension components together constitute the variable sensitivity weight vector for that amplitude level. A higher value in the vector indicates a stronger ability of the variable to transmit thermal and mass disturbances under the corresponding disturbance intensity. In the later stages of rice batch drying, surface moisture has largely evaporated. The resistance to the migration of bound water from the interior to the surface becomes the main factor limiting evaporation. A slight increase in airflow results in a rapid temperature decrease while the humidity within the cavity remains almost unchanged, indicating that heat transfer efficiency is now higher than moisture transfer efficiency. The sensitivity weight of temperature variables is significantly higher than that of humidity variables. At this stage, blindly increasing airflow to reduce humidity is ineffective and may even lead to overheating. The control strategy should focus on temperature precision rather than dehumidification rate. The sensitivity weight of variables retains a three-level structure at the amplitude level. The sensitivity weight vectors corresponding to slight, moderate, and strong disturbances are calculated independently. The differences in the values ​​of the three-level vectors reflect the differentiation of variable response priorities under different disturbance intensities. A larger difference in sensitivity at the amplitude level indicates a more pronounced differentiation in the temperature and humidity response of the batch of grain to airflow adjustments of that intensity.

[0025] A phase switching signal is generated by ranking the sensitivity weights of variables based on moisture content sensitivity. The physical basis for ranking is that the moisture content range of the grain directly affects the relative priority of temperature and humidity responses. When the moisture content is high, the evaporation driving force is strong, and the humidity sensitivity component is higher than that of temperature. When the moisture content drops to the critical range, the relationship between the two components reverses. The width of the critical range varies depending on the grain variety. The critical range for rice crops is usually narrower and the reversal is more abrupt, while the critical range for corn crops is wider and the reversal is more gradual. The difference d_t between the temperature and humidity components in the sensitivity weight vector of variables at each time point of the entire batch is arranged by time to form a difference time series. The zero-crossing moment when d_t turns from negative to positive is located as the sensitivity priority reversal point. The distribution density of the reversal points on the time axis reflects the uniformity of the drying front advancement. Sparse reversal points indicate clear phase boundaries, while dense points indicate that the moisture content fluctuates repeatedly near the critical range. During the loading of a certain batch of rice, uneven feeding speed resulted in a thicker layer of grain on the upper part of the cavity and a thinner layer on the lower part. Initially, humidity sensitivity dominated for approximately 40 minutes, followed by three temperature and humidity sensitivity priority reversals within 15 minutes. This indicated that the grain in different layers of the cavity almost simultaneously crossed the critical moisture content, but at different times. Correspondingly, the reversal amplitude of strong disturbance levels was significantly greater than that of slight disturbance levels during this period, suggesting that the differentiation in thermo-mass response caused by uneven interlayer distribution was more pronounced under large disturbances. Fine-tuning the airflow masked the interlayer differences. Each reversal point, along with the change in d_t before and after the reversal, constitutes the stage switching signal. The larger the amplitude of the stage switching signal, the more significant the moisture content transition corresponding to that switch. When the reversal time deviation exceeds one sampling interval, the timing of the sensitivity weights of the current batch variables is locally interpolated and refined before repositioning to ensure that the accuracy of the stage switching signal reversal time is not less than half of the sampling interval.

[0026] For example, the step of forming a differentiated calibration benchmark by performing stage adaptive benchmark mapping based on the stage switching signal and the variable sensitivity weight includes: performing cavity temperature zone gradient distribution analysis on the variable sensitivity weight and the stage switching signal to locate the temperature zone attenuation critical point; performing spatial layering pattern identification on the temperature zone attenuation critical point to generate a temperature zone distribution feature set; performing batch-to-batch variation analysis on the temperature zone distribution feature set to generate batch configuration parameters; and performing stage adaptive benchmark mapping on the temperature zone attenuation critical point based on the batch configuration parameters to generate a differentiated calibration benchmark.

[0027] The temperature gradient distribution of the cavity is analyzed using variable sensitivity weights and stage switching signals to locate the critical point of temperature attenuation. The difference sequence of the normalized response amplitude of each layer of temperature in the longitudinal direction of the cavity to the air volume disturbance constitutes the longitudinal temperature sensitivity gradient. Adjacent layers with high absolute gradient values ​​indicate a significant difference in heat transfer efficiency between the two layers. This difference exhibits a systematic shift as the drying process progresses. Layers near the air inlet side usually show a high gradient in the early stage of drying, while the gradient tends to flatten out in the middle and late stages as the overall grain temperature decreases. After extracting the longitudinal gradient sequence corresponding to each reversal point of the stage switching signal, the interlayer migration of the gradient peak position between adjacent reversal points reflects the drift amplitude of the temperature interface over time. Interlayer positions with a drift exceeding 10% of the total cavity height are marked as candidate points for temperature attenuation. Candidate points that correspond to local extreme values ​​in the variable sensitivity weight difference sequence near the reversal point are confirmed as critical points for temperature attenuation. Candidate points that only meet the drift condition but do not correspond to extreme values ​​are considered false critical points and are excluded. During grain drying, hot air enters from the bottom of the cavity and passes through the grain layer. The lower grain is the first to come into contact with the hot air and its moisture content decreases first. The difference in heat and mass transfer efficiency is most significant between the longitudinal layers. The fourth to fifth layers in the middle section of the cavity all show gradient peaks at three consecutive flipping points. This position is confirmed as the critical point of temperature attenuation in this stage of this batch. The critical point drifts from the fourth layer to the sixth layer as the stage progresses, indicating that the drying front gradually moves from the bottom to the top of the cavity. When the drift rate is slow, it indicates that there is a grain layer with high moisture content in the lower part of the cavity, forming a local high moisture content retention zone. The drying efficiency in this area is significantly lower than the overall process.

[0028] Spatial stratification pattern identification is performed on the temperature attenuation critical point to generate a temperature distribution feature set. The position of the temperature attenuation critical point in the longitudinal direction of the cavity divides the cavity into a high-temperature zone above the critical point and a low-temperature zone below the critical point. The two stratification boundaries are dynamically adjusted as the temperature attenuation critical point drifts. If the critical point crosses the sensor deployment layer during the drift, the corresponding layer's classification switches from one side to the other. The longitudinal distribution pattern of the temperature sensitivity component within each layer is characterized by two statistical measures: mean and kurtosis. A high mean and low kurtosis indicate uniform heat transfer within the layer, while a low mean and high kurtosis indicate that heat transfer is concentrated in a few local layers. The difference between the mean sensitivity of the high-temperature zone and the low-temperature zone constitutes the interlayer sensitivity difference. When the difference exceeds 0.3, the thermal properties of the two layers are significantly different and independent distribution features need to be established separately. When the difference is less than 0.1, the two layers are close and can be merged into a single distribution. During the loading of a certain batch of grain, the feed chute was deflected, resulting in a relatively loose accumulation near the air inlet of the cavity. Hot air preferentially concentrated and passed through the loose area with low air resistance, bypassing the dense layer. In the upper part of the cavity, the high-temperature zone had a high heat concentration in a few sensor layers near the air inlet, while the temperature sensitivity of other layers was significantly lower, manifested as a persistently high kurtosis in the high-temperature zone. The temperature sensitivity distribution of the corresponding layers was extremely uneven. The layers through which the airflow concentrated were the layers corresponding to the kurtosis. The kurtosis field of the high-temperature zone for this batch was separately marked in the temperature distribution feature set. The temperature distribution feature set is composed of the sum of the mean and kurtosis vectors of each stage and each layer. The drift trajectory of the temperature attenuation critical point is added to the temperature distribution feature set as the dynamic coordinates of the layer boundary, and the coverage is consistent with the number of stages divided by the stage switching signal.

[0029] Batch configuration parameters are generated by analyzing the batch-to-batch variation of the temperature distribution feature set. This analysis compares the current batch's temperature distribution feature set with historical batches at corresponding stages. Differences in mean values ​​within the same stage reflect differences in initial moisture content and stacking state of the grain between batches, while kurtosis differences reflect differences in heat transfer uniformity. These two types of differences are independent. Batch pairs with similar mean values ​​but significant kurtosis differences indicate that the total heat transfer intensity of the two batches is comparable, but their distribution patterns are different, requiring different adjustments to the corresponding baseline mapping. When the mean difference exceeds twice the standard deviation of the historical batch's corresponding stage mean, a significant heat transfer baseline shift is identified in the current stage. For example, a batch of rice harvested after continuous heavy rainfall has a higher initial moisture content. Evaporation is synchronously vigorous in all layers of the high-temperature zone within the cavity, and heat is evenly diffused to each sensor layer rather than concentrated locally. The high-temperature zone mean is 0.18 higher than the historical mean, and the kurtosis is lower. Compared to previous batches, this batch has stronger evaporation drive and a more uniform distribution, requiring the baseline mapping to be adjusted towards higher intensity to match the actual heat and mass transfer intensity of this batch. After normalizing the mean difference and kurtosis difference of each stage, the batch configuration parameter sequence is formed by splicing them in the stage order. When the sign of each stage component of the batch configuration parameter is positive, the adjustment is towards the strong heat transfer end, and when the sign is negative, the adjustment is towards the weak heat transfer end. When the number of historical batches is less than 3, the corresponding stage component of the batch configuration parameter is replaced with a neutral value of 0. The neutral value corresponds to the conservative assumption that the current batch has no significant deviation from the history. After the historical batches accumulate to 3 or more, the stage component is automatically switched to the calculated value based on the measured data.

[0030] Based on batch configuration parameters, a phased adaptive benchmark mapping is implemented for the temperature attenuation critical point to generate a differentiated calibration benchmark. The phased adaptive benchmark mapping uses the interlayer position and drift trajectory of the temperature attenuation critical point as spatial positioning coordinates, and the phase components of the batch configuration parameters as adjustment amounts. Within each phase, independent benchmark stretching operations are performed on the high-temperature and low-temperature regions. The stretching amount for the high-temperature region is determined by multiplying the amplitude of the mean difference component of the corresponding phase of the batch configuration parameters by a stretching coefficient. The stretching coefficient is pre-calibrated based on the statistical mapping relationship between the mean difference and the optimal calibration effect in historical batches. The stretching direction is consistent with the sign of the mean difference. When the stretching amount exceeds 20% of the current range of the high-temperature region benchmark, it is truncated to this upper limit to prevent drastic benchmark drift caused by anomalous data from a single batch. The low-temperature region is adjusted based on the kurtosis difference component. When the kurtosis difference is positive, the weight of the concentrated area in the low-temperature region is increased while the weight of the dispersed area is decreased. When the kurtosis difference of a certain batch in the low-temperature region is 0.22, the weight of the concentrated area is increased by 15%. When the kurtosis difference is negative, the adjustment direction is reversed. When the absolute value of the kurtosis difference is less than 0.05, the corresponding stratum is considered to have no significant inter-batch kurtosis difference, and the weight adjustment step is skipped. After the two regions are adjusted, the temperature attenuation critical point is used as the splicing and stitching point to form the overall calibration benchmark. The benchmark values ​​of adjacent stages transition linearly at the switching boundary. The overall calibration benchmarks of each stage in the whole batch are spliced ​​in the stage sequence to form the differentiated calibration benchmark. The spatial resolution of the differentiated calibration benchmark is determined by the longitudinal positioning accuracy of the temperature attenuation critical point. The more accurate the critical point positioning, the higher the splicing and stitching accuracy of the high and low temperature zones. The differentiated calibration benchmark retains the independent component structure of the high and low temperature zones.

[0031] Standard operating parameters are formed by integrating differentiated calibration benchmarks and thermal mass compensation weights. The differentiated calibration benchmarks determine the reference value ranges for high and low temperature zones at each stage, while the thermal mass compensation weights reflect the activity level of thermal mass transfer at each stage. The core of their integration is to apply a dynamic offset to the benchmark values ​​based on thermal mass activity, causing the standard operating parameters to converge towards the active direction during stages of strong thermal mass transfer and towards the conservative direction during stages of weak thermal mass transfer. In the middle stage of drying for a certain batch, the thermal mass compensation weight for the temperature component reached 0.89, corresponding to a shift in the temperature calibration benchmark towards a higher value in the high-temperature zone of that stage, directly reflecting the vigorous evaporation and sufficient heat demand in the cavity at this stage. In the final stage, both components of the thermal mass compensation weights dropped below 0.65, and the temperature and humidity calibration benchmarks narrowed simultaneously, indicating that the cavity should maintain a low thermal mass input to prevent over-drying when the grain is nearing its final state. When the compensated calibration value exceeds the normalized range of 0 to 1, it is truncated to the boundary value. The compensated calibration values ​​of all variables in each stage of the differentiated calibration benchmark are arranged in the order of temperature, humidity, air volume and moisture content to form a subset of the standard operating parameters for that stage. The subsets of each stage in the whole batch are spliced ​​together in time order to form complete standard operating parameters. The continuity at the boundary of each stage is guaranteed by linear interpolation of the transition section of the differentiated calibration benchmark. The temporal resolution of the standard operating parameters is consistent with the common sampling benchmark of the operating condition benchmark matrix.

[0032] Step S130: Input the standard operating parameters into the large model to perform comprehensive reasoning on the coupling state of moisture content gradient and temperature, humidity and wind to obtain reasoning decision weights, and perform weighted mapping on the operating condition baseline matrix according to the reasoning decision weights to generate a coupling response spectrum.

[0033] In some embodiments, the step of inputting the standard operating parameters into a large model to perform comprehensive reasoning on the coupling state of moisture content gradient and temperature-humidity-wind to obtain reasoning decision weights includes: identifying the drying stage of the standard operating parameters to generate a stage type identifier; structurally encoding the stage type identifier and the standard operating parameters to generate a process state input sequence; inputting the process state input sequence into a large model to perform comprehensive reasoning on the coupling state of moisture content gradient and temperature-humidity-wind to obtain gradient response components and coupling strength evaluation results; and fusing the gradient response components and the coupling strength evaluation results to generate reasoning decision weights.

[0034] The standard operating parameters are used to identify the drying stage and generate stage type identifiers. The basis for drying stage identification is the fundamental difference in the moisture migration mechanism of grain within different moisture content ranges. In the constant-rate drying stage, the evaporation rate of free water on the grain surface is stable, and the synergistic characteristics of temperature, humidity, and wind show a regular and stable pattern. In the falling-rate drying stage, the resistance to the migration of bound water inside the grain increases, the coupling relationship of the three variables tends to be nonlinear, and the rate of moisture content decrease continuously narrows. The temporal slope of the moisture content column in the standard operating parameters is used as the main criterion for stage identification. When the absolute value of the slope remains stable and is relatively large across multiple consecutive sampling points, it is determined to be the constant-rate stage; when the absolute value of the slope continuously shrinks and the fluctuation intensifies, it is determined to be the falling-rate stage; and when the slope approaches zero and the moisture content is lower than the target final value, it is determined to be the completed stage. Precise positioning of stage boundaries is aided by abrupt changes in the joint temperature and humidity response of standard operating parameters. When the difference between the boundary time determined solely by the moisture content slope of standard operating parameters and the joint temperature and humidity abrupt change time exceeds two sampling intervals, the joint temperature and humidity abrupt change time is taken as the standard. For a certain batch, the starting time of the narrowing of the moisture content slope is approximately 18 seconds earlier than the joint temperature and humidity abrupt change time. Ultimately, the stage boundary from constant rate to deceleration for that batch is determined by the joint temperature and humidity abrupt change time. The stage type identifier consists of the type code and start and end times of each stage. The type code distinguishes between constant rate, deceleration, and completion. The number of stage type identifier segments within the same batch is usually 2 to 4. When the number of stage type identifier segments exceeds 5, it indicates abnormal fluctuations in the moisture content time series. It is necessary to recheck whether the moisture content column of the standard operating parameters for that batch is affected by acquisition noise. If noise interference is confirmed, the moisture content column is filtered by median and the stage identification is re-executed.

[0035] The process state input sequence is generated by structured encoding of stage type identifiers and standard operating parameters. Structured encoding converts multivariate time-series data into an input format that the large model can parse, while preserving the temporal relationships between variables and the semantic boundaries between stages. Without either, the large model cannot distinguish the physical meaning differences of the same variable's value at different stages during inference. Standard operating parameters are expanded row by row according to sampling time, with each row containing four normalized values: temperature, humidity, airflow, and moisture content. The type code of the stage type identifier is added to the corresponding row as an additional column. The type code uses a unique thermal method to distinguish between three categories of stage type identifiers: constant speed, decreasing speed, and completion. The constant speed segment corresponds to the code [1,0,0], the decreasing speed segment to [0,1,0], and the completion segment to [0,0,1]. Three sampling rows before and after the boundary time of the stage type identifier are appended with boundary flags. These boundary flags indicate to the large model that this region belongs to the stage transition zone. The temporal characteristics of variables within the transition zone do not represent the typical state of any stage. The large model reduces the confidence weight of the transition zone rows when inferring the moisture content gradient. The row vectors of the process state input sequence fully describe the joint features of the standard operating parameters at the corresponding time in the two dimensions of caloric quality and stage. The large model reads the process state input sequence segment by segment through a sliding window method. The window length matches the average stage duration of the batch, and the window step size is half of the window length to ensure continuous coverage of stage features between adjacent windows. If the proportion of the transition zone row in the process state input sequence window exceeds 50%, the window is skipped and not sent into inference.

[0036] The process state input sequence is fed into a large model to perform comprehensive inference on the coupling state of moisture content gradient and temperature, humidity, and wind, obtaining gradient response components and coupling strength evaluation results. After receiving a sliding window of the process state input sequence, the large model performs parallel inference on the temporal pattern of the moisture content column and the joint dynamics of the three variables of temperature, humidity, and wind within the window. The moisture content gradient inference focuses on the distribution pattern of the rate of change of moisture content within the window and its consistency with the stage type identifier, while the temperature, humidity, and wind coupling inference focuses on the strength and direction of the coordinated response of the three variables within the window. The two inference paths share the contextual representation of the process state input sequence to capture the cross-variable correlation between moisture content change and heat and mass transfer. The gradient response component is composed of the moisture content gradient intensity vector output by the large model. Each dimension of the gradient response component vector corresponds to the gradient response intensity at different time positions within the window. The higher the gradient intensity, the higher the degree of matching between the rate of moisture content decrease and the driving force of temperature, humidity, and wind at that moment. During the middle stage of the grain batch deceleration phase, the operator excessively reduced the air volume due to concerns about excessive dryness, resulting in slow removal of accumulated moisture in the cavity and a tendency for the moisture content decrease to stagnate. The intensity of the gradient response component in this period was below 0.3 for several consecutive moments. The large model determined that there was a severe mismatch between the thermo-mass driving force and the moisture content decrease during this period. The coupling strength evaluation result is composed of the pairwise coupling strength matrices of the three variables of temperature, humidity, and wind output by the large model. Each element outside the diagonal of the matrix reflects the degree of synergy between the corresponding variable pairs within the current window. The larger the absolute value of the element, the stronger the coupling. A positive sign indicates synergy in the same direction, and a negative sign indicates constraint in the opposite direction. Windows with boundary markers in the process state input sequence are marked with low confidence in the coupling strength evaluation result. The matrix elements of the transition zone window with low confidence in the coupling strength evaluation result are not included in the mean statistical range of the weighted fusion.

[0037] The inference decision weights are generated by fusing gradient response components and coupling strength assessment results. The gradient response components reflect the temporal distribution of the driving force for water content decrease, while the coupling strength assessment results indicate the type of synergistic mechanism among the three variables from which the driving force originates. Periods with strong water content gradients but weak coupling strength indicate sufficient evaporation driving force but low thermo-mass synergistic efficiency. The fusion of gradient response components and coupling strength assessment results can identify the optimal operating period that simultaneously satisfies high gradient and high coupling. The fusion weight W_i is determined by the formula W_i = α × G_i + β × C_i, where G_i is the gradient strength scalar of the gradient response component at time i, C_i is the weighted average of the absolute values ​​of each element in the coupling strength assessment result matrix at time i, and α and β are fusion coefficients whose sum is always 1. In the constant-rate stage, α is set to 0.4 and β to 0.6 to highlight the dominant role of the coupling structure in the uniform evaporation stage; in the deceleration stage, α is set to 0.7 and β to 0.3 to highlight the weight of the gradient strength in determining the internal diffusion stage. The fusion coefficients are switched stage by stage according to the stage type identifier. The fusion results W_i at each time point are arranged in chronological order and then normalized over the entire batch duration. The inference decision weights are composed of the normalized W_i time series. In the inference decision weights, the time with high W_i indicates that the large model determines that the drying efficiency is relatively high, which is jointly determined by the moisture content gradient and the temperature-humidity-wind coupling at that time. The time with low W_i indicates that the thermal-mass coupling efficiency is relatively low.

[0038] The coupled response map is generated by weighting the operating condition baseline matrix based on the inference decision weights. The weighted mapping uses the W_i time series of the inference decision weights as the weight vector, and applies a weighted product operation to the corresponding time values ​​of each column of the operating condition baseline matrix. After weighting, the numerical distribution of each column of time series is concentrated towards higher W_i times. The thermomass state in the efficient period obtains a larger amplitude proportion in the mapping result, while the values ​​of the inefficient period are compressed. Therefore, the mapping result more accurately reflects the characteristics of the state interval with high heat and mass transfer efficiency. The four columns of the operating condition baseline matrix—temperature, humidity, air volume, and moisture content—are independently weighted and mapped. The weighted time series of each column retains the original time resolution of the operating condition baseline matrix. The independent mapping preserves the response differences of each variable at different time periods. When the low value range of the inference decision weight is concentrated in the period when the response of a certain variable is abnormal, the weighted time series value of the corresponding column of that variable is strongly compressed, and the compression magnitude is significantly greater than that of other variable columns. In a certain batch of drying process, the heating supply was sufficient, but the fan speed of the exhaust system was low, and the air volume was small, which limited the moisture removal efficiency. During the high-efficiency drying period determined by the large model, the air volume failed to coordinate with the temperature. The amplitude of the weighted air volume column was significantly lower than that of the temperature and humidity columns, indicating that the driving force of the high-efficiency period of heat and mass in this batch was mainly contributed by the heating intensity and humidity gradient, rather than the air volume adjustment. The low fan frequency is the priority direction for improving the drying efficiency of this batch. The coupling response spectrum consists of four weighted time series columns arranged side by side. The relative amplitude differences between columns intuitively reflect the strength of each variable's contribution during the current batch's efficient drying period. The envelope curve of the coupling response spectrum in the time dimension is highly consistent with the time series shape of W_i of the inference decision weights. The peak time of the envelope curve corresponds to the optimal period of thermo-mass coupling, and the valley segment of the envelope curve corresponds to the period of thermo-mass mismatch.

[0039] Step S140: Establish a drying joint control benchmark based on the coupled response spectrum; determine the operating deviation amplitude by performing multiple batch difference detection based on the drying joint control benchmark; determine the control confidence boundary after eliminating operating condition disturbances based on the operating deviation amplitude; and select abnormal operating conditions from the operating condition benchmark matrix to form a control deviation record group based on the control confidence boundary.

[0040] Specifically, a drying joint control benchmark is established based on the coupled response spectrum. High-amplitude segments in each column of the coupled response spectrum correspond to the operating range with the highest thermo-mass coupling efficiency. The threshold for determining high-amplitude segments is determined by adding one standard deviation to the weighted average of each column. Continuous periods exceeding this threshold are marked as high-efficiency operating zones. The value ranges of the four variables within the high-efficiency operating zone define the optimal operating range for each variable. The upper and lower boundaries of the optimal operating range are taken as the 90th and 10th quantiles of the corresponding variables within the high-efficiency operating zone, respectively, excluding the interference of marginal extremes on the interval boundaries. In a certain batch's constant-rate phase, the high-efficiency operating zone lasts approximately 35 minutes. During this period, the normalized temperature values ​​are concentrated in the higher quantile range, thus determining the upper and lower boundaries of the corresponding temperature control benchmark. The drying joint control benchmark takes the optimal operating range of each variable as the core constraint. The four-variable control benchmark of temperature, humidity, air volume and moisture content jointly cover the operating conditions of optimal heat and mass synergy during the drying process. Due to the fluctuation of fuel supply pressure, the heating intensity of the hot air burner fluctuates periodically, or the uneven distribution of moisture content between grain layers causes the heat and mass transfer of different layers to alternate to reach the optimal state. In the coupled response spectrum, the high amplitude segment is divided into a fragmented distribution by the low amplitude segment. The boundary of the drying joint control benchmark established for such batches is appropriately widened to accommodate the normal drift amplitude of the heat and mass state under periodic disturbances, and to avoid misjudging normal adjustment fluctuations as abnormal deviations. The drying joint control benchmark is established in segments according to the drying stage. The four-variable control benchmarks for the constant rate stage and the falling rate stage are determined independently. The constant rate stage benchmark reflects the stable thermo-mass synergy range dominated by free water evaporation, while the falling rate stage benchmark reflects the narrowing thermo-mass synergy range dominated by internal bound water migration. The boundary connection between the two stage benchmarks is aligned with the stage boundary time according to the stage type identifier. The high-efficiency operating condition band spanning the two stages in the coupled response spectrum is split according to the stage type identifier of its row and assigned to the corresponding stage control benchmark, without merging across stages.

[0041] In some embodiments, determining the operational offset amplitude by performing multi-batch difference detection based on the drying joint control benchmark includes: performing batch registration based on the drying joint control benchmark to obtain a homogeneous operational sequence; extracting the cavity thermal inertia attenuation characteristics from the homogeneous operational sequence to generate a thermal inertia compensation coefficient; extracting the difference between adjacent batches from the homogeneous operational sequence based on the thermal inertia compensation coefficient to generate a difference sequence; and performing extreme value screening and calibration of the operational offset amplitude based on the difference sequence.

[0042] Batch registration is performed based on the drying joint control benchmark to obtain the common operating sequence. The premise of batch registration is that different batches are comparable in terms of drying stage division and thermo-mass state. Direct alignment by absolute time will result in different drying stages at the same moment due to differences in initial moisture content. For example, high-moisture-content batches harvested during the rainy season may experience a constant-rate evaporation stage lasting more than 2 hours, while batches stored concurrently and partially dehydrated may only experience a constant-rate evaporation stage lasting 40 minutes before entering a deceleration stage. If direct comparison is performed based on feed time, one batch may be in the vigorous stage of free water evaporation while another has entered the difficult stage of internal bound water migration, making their thermo-mass states inherently incomparable. This will generate a large number of spurious offsets in cross-batch difference calculations. Using the moisture content characteristic values ​​of each stage boundary in the drying joint control benchmark as anchor points for relative time alignment within each stage yields better registration accuracy than the absolute time alignment scheme. The starting point of the constant-rate segment for each batch corresponds to the initial moisture content characteristic value of the constant-rate segment in the drying joint control benchmark, and the starting point of the decreasing-rate segment corresponds to the initial moisture content characteristic value of the decreasing-rate segment. The absolute time axis of each batch is stretched or compressed to the standard time axis of the drying joint control benchmark according to the above anchor points. The stretching / compression ratio is determined by the ratio of the actual stage duration of each batch to the standard stage duration of the corresponding stage in the drying joint control benchmark. After mapping, the normalized values ​​of the four variables corresponding to the same standard time for each batch are arranged side by side in the batch dimension, forming a cross-batch variable observation set for each standard time. The observation sets of all standard times constitute a homogeneous operating sequence. If a batch has an abnormally long constant-rate segment duration due to a heating failure, and the stretching ratio exceeds 1.5 times during registration, the homogeneous operating sequence for the corresponding time period of this batch is marked with an abnormal stretching, and the weight of the data of this batch is appropriately reduced when extracting the difference. The number of rows in the cross-batch observation set of each standard time in the homogeneous operating sequence is equal to the total number of batches participating in the registration. When the total number of batches is less than 3, the statistical representativeness of the homogeneous operating sequence is insufficient.

[0043] The thermal inertia decay characteristics of the cavity are extracted from the same operating sequence to generate a thermal inertia compensation coefficient. The thermal inertia decay characteristics of the cavity reflect the systematic pattern that the initial thermal state of the cavity is higher due to residual heat between adjacent batches. In continuous production scenarios, after the previous batch of grain is dried and unloaded, the residual heat in the cavity wall bricks and interlayer has not yet dissipated. After the next batch of wet grain is fed, the cavity wall continues to release residual heat to the fed grain, which causes the temperature sensor reading of the new batch in the constant speed section to be higher than the normal drying baseline. This higher amount gradually dissipates and decreases over time. The residual heat is particularly significant when the interval between two batches is less than 30 minutes. After the production stops at night, the initial thermal inertia deviation of the first batch is close to zero because the cavity has been fully cooled. If the two types of batches are not treated differently, the calculation of the temperature difference between batches will be mixed with a large amount of residual heat interference that is unrelated to the operating conditions. The difference between the average value of the temperature series across batches in the initial stage of the constant-rate segment in the same-source operation sequence and the standard value of the temperature in the initial stage of the constant-rate segment of the drying joint control benchmark is defined as the initial thermal inertia deviation D_0 (normalized dimensionless value). The decay pattern of this deviation with the standard time axis is characterized by exponential fitting, and the fitting model is D(t)=D_0×e^(-k×t), where k is the decay coefficient in units of 1 / minute, t is the relative time from the start of the constant-rate segment in units of minutes, and k is estimated by fitting the temperature series of each batch in the same-source operation sequence. The larger the value of k, the faster the residual heat of the cavity dissipates. When the cavity insulation performance is poor or the batch interval is long, the value of k is usually larger. The thermal inertia compensation coefficient is composed of the fitted value of D(t) at each standard time. Each standard time corresponds to a compensation amount. When D(t) approaches zero in the middle and late stages of the constant-rate segment, the corresponding thermal inertia compensation coefficient is close to zero, indicating that the residual heat effect has dissipated during this period. After this time, the thermal inertia compensation coefficient is set to zero.

[0044] For the same-source operating sequence, the difference between adjacent batches is extracted based on the thermal inertia compensation coefficient to generate a difference sequence. The difference extraction is performed pairwise within the cross-batch observation set at each standard time of the same-source operating sequence. Adjacent batches are defined as two batches that are consecutive in time after registration. For adjacent batches, the four-variable observation values ​​at the same standard time are subtracted variable by variable. Before subtracting, the compensation amount corresponding to the thermal inertia compensation coefficient is subtracted from the temperature series observation values ​​of each batch to eliminate the systematic lifting effect of residual heat on the temperature difference. If the temperature difference is directly subtracted without compensation, the residual heat component will cause the subsequent operating deviation to be systematically overestimated. The humidity, air volume, and moisture content series are not affected by thermal inertia and are directly subtracted. The time series of the variable-wise differences of each adjacent batch pair are arranged on the standard time axis to form the difference time series of that batch pair. The difference time series of all adjacent batch pairs are superimposed on the batch pair dimension and the cross-batch pair mean is taken at each standard time. The mean time series constitutes the difference sequence. The standard period in the difference sequence with a continuously high value indicates that multiple batches have a systematic shift during that period, rather than random fluctuations of individual batches. Occasional disturbances of a single batch are diluted in the cross-batch mean, while systematic shifts are amplified in the mean, so the two can be effectively distinguished.

[0045] Extreme value screening is used to calibrate the operational offset magnitude based on the difference sequence. The time series values ​​of each variable column in the difference sequence reflect the intensity of inter-batch drift in thermo-mass state between adjacent batches. The local maxima in the time series correspond to the moment when the inter-batch deviation is most significant. The magnitude and duration of the maxima together measure the severity of the offset event. Extreme value screening employs a sliding window peak detection method. The window length is set to one-quarter of the standard duration of a single stage in the drying joint control benchmark. The maximum absolute value of the difference between each variable column within the window is extracted as the peak value of that window. Among all window peak values, windows exceeding 1.5 times the global standard deviation of the corresponding variable column in the difference sequence are marked as high-offset windows. When high-offset windows are continuously distributed on the time axis, they are merged into an offset event segment. The maximum value of the peak values ​​of each variable column within the segment is extracted as the variable offset amplitude of the offset event. The offset amplitudes of the four variables together constitute the operating offset amplitude vector of the offset event. In a certain batch, the PID parameters of the temperature control loop were not recalibrated after maintenance, resulting in a continuous accumulation of temperature deviations, while the humidity and air volume loops were still working normally. The offset amplitude of the temperature column in the difference sequence was significantly higher than that of the humidity column. The significant difference between the two columns indicates that the offset event was driven by the drift of the single temperature loop parameter, rather than an overall imbalance of the cavity's heat and mass. Subsequent targeted investigation of the temperature control loop is sufficient without recalibrating the entire drying strategy. In the difference sequence, the variable column without a high offset window corresponds to a variable component in the running offset amplitude vector that is set to zero. A component of zero indicates that the inter-batch deviation of the variable is within the normal fluctuation range within the current multi-batch detection range.

[0046] In some embodiments, determining the control reliability boundary after eliminating operating condition disturbances based on the operating offset amplitude includes: segmenting the operating condition period of the operating offset amplitude to obtain the disturbance change range; identifying the operating condition mode within the disturbance change range to generate material stratification parameters; extracting the normal operating condition fluctuation range based on the material stratification parameters and the disturbance change range to generate an operating condition baseline; and eliminating the change amount within the operating condition baseline from the operating offset amplitude to determine the control reliability boundary.

[0047] The operating deviation amplitude is segmented into working cycle segments to obtain the disturbance variation range. The basis for segmenting the working cycle is that the fluctuations in the working conditions during the drying process have a quasi-periodic structure related to the feeding and discharging operations and the hot air supply adjustment cycle. Directly performing global statistics on the operating deviation amplitude would misjudge the normal fluctuation peaks within the quasi-period as abnormal deviations. After segmentation, separate analysis within each cycle can effectively distinguish between periodic normal fluctuations and cross-cycle cumulative drift. The time sequence of each variable component in the operating deviation amplitude vector is statistically analyzed according to the distance between adjacent peaks. The mode of the distance distribution is defined as the dominant fluctuation cycle T_p. T_p is compared with the set cycle of the air volume adjustment operation in the drying joint control benchmark. If the difference between the two does not exceed 20%, it is confirmed that the dominant fluctuation cycle originates from the air volume adjustment operation. If the difference exceeds 20%, it indicates the presence of an unplanned disturbance source that requires manual verification. Unplanned disturbance sources are commonly caused by non-periodic pressure fluctuations due to discontinuous grain feeding or partial blockage of the hot air pipeline. The time series of operational offset amplitudes is divided into segments with a step size of T_p. The start and end times of each segment correspond to the local minimum values ​​in the operational offset amplitude time series. The minimum value is used as the dividing point to ensure that the peak values ​​of fluctuations within each segment fall completely within a single segment without being truncated. The duration of each segment varies slightly due to the small offset of the minimum value position. The preliminary division results constitute the disturbance change interval. When there is a difference between the T_p values ​​calculated by the temperature column and the humidity column, the least common multiple of the two is taken as the unified segment period. The segments that cross periods within the disturbance change interval are merged and corrected to ensure that the complete fluctuation period of each variable falls within the same disturbance change interval. Segments that cross the boundary between the constant speed segment and the deceleration segment are forcibly truncated at the boundary time according to the stage type. The short segments generated after truncation are merged into the adjacent segments to ensure the minimum sample size. The time intervals of each segment are finally integrated to form the set of disturbance change intervals.

[0048] Material stratification parameters are generated through operating condition pattern identification within the disturbance variation range. Operating condition pattern identification targets the modulation effect of the longitudinal stratification differences of grain within the cavity on the thermo-mass response patterns within each disturbance variation range. Grain layers with high moisture content and those with low moisture content exhibit fundamentally different directions and amplitudes of temperature and humidity responses to the same airflow disturbance. The high moisture content layer has a strong evaporation driving force, and the humidity response amplitude caused by the airflow disturbance is significantly greater than that of the low moisture content layer. Batches with different stratification states will exhibit different operating condition response patterns within the same disturbance variation range. The time-series mean vector and covariance matrix of the four variable components of the operational offset amplitude are extracted within each disturbance variation range. The mean vector reflects the average offset level of each variable within the range, and the covariance matrix reflects the linkage relationship between the offsets of each variable. Together, they constitute a description of the operating condition response pattern for that range. The operating condition response patterns of different disturbance change ranges are described by cluster analysis and grouped. The clustering is based on a weighted combination of the Euclidean distance of the mean vector and the Frobenius distance of the covariance matrix. The weight of the Euclidean distance is 0.4 and the weight of the Frobenius distance is 0.6. The average offset levels of the high and low water content layers in the cavity may be similar, but the linkage modes are completely different. The high water content layer has strong linkage between temperature and humidity, while the low water content layer is dominated by temperature and has a weak response to humidity. It is difficult to distinguish the two types of stratification states based on the mean distance alone. The weight allocation biased towards the covariance matrix ensures that the difference in linkage structure plays a dominant role in the stratification determination. In the clustering results, the relative levels of the centroid mean vector and the moisture content component of each cluster correspond to the stratified moisture content state within the grain cavity. Clusters with high moisture content components correspond to the operating mode dominated by the high moisture content layer, while clusters with low moisture content components correspond to the operating mode dominated by the low moisture content layer. The value of the moisture content component of each cluster directly determines three factors: the moisture content range of each layer, the operating mode category corresponding to each layer, and the moisture content gradient between layers. These three factors together constitute the material stratification parameters, and the number of layers in the material stratification parameters is consistent with the number of clusters.

[0049] For example, the step of extracting the normal operating condition fluctuation range and generating the operating condition baseline based on the material stratification parameters and the disturbance change interval includes: performing synchronous comparison registration based on the disturbance change interval to generate a batch control sequence; removing the equipment aging drift component from the batch control sequence to generate an environmental correction component; setting multiple operating condition sampling points for the batch control sequence based on the material stratification parameters to generate a normal operating condition fluctuation range; and performing statistical analysis on the normal operating condition fluctuation range and the environmental correction component to form the operating condition baseline.

[0050] Based on the perturbation variation range, a batch control sequence is generated through synchronous control registration. Synchronous control registration aligns the four-variable time series within each perturbation variation range of the current batch with the time series of historical batches at the same drying stage and with the same perturbation variation range number. Alignment is based on the standard time axis of the drying joint control benchmark, rather than an absolute clock, to eliminate the influence of differences in start time between different batches on the control results. The selection of historical batches is limited to batches with the same grain variety as the current batch and an initial moisture content difference not exceeding 2%. Same variety ensures consistent thermo-mass response mechanisms, and moisture content difference limits ensure comparable drying process rates. If the number of historical batches meeting these conditions is less than three, the upper limit of moisture content difference is relaxed to 5%, and a low-confidence label is added to the control results. When selecting historical batches, batches with equipment failure records or manual intervention operations must also be excluded, as the operating conditions of such batches are affected by abnormal factors, and including them in the control would introduce abnormal perturbation components into the batch-to-batch differences. After aligning the current batch time series with the corresponding time series of each historical batch within each disturbance variation interval, the four-variable observation values ​​of the current batch and each historical batch at the same time are arranged side by side to form a cross-batch comparison observation row for that time. If there are missing values ​​in each historical batch column in the comparison observation row due to temporary sensor absence, they are filled with the local mean of the variable within the corresponding interval of that historical batch. The filling time is not included in subsequent range statistics. The comparison observation rows for all times within each disturbance variation interval are arranged in chronological order to form a batch comparison sequence. The number of columns in the batch comparison sequence is equal to the number of historical batches participating in the comparison plus one, and the number of rows is equal to the sum of the total number of times within each disturbance variation interval. The row-by-row difference between the current batch column and each historical batch column in the batch comparison sequence reflects the deviation of the current batch from the historical batches at each time.

[0051] Environmental correction components are generated by removing equipment aging drift components from the batch control sequence. Equipment aging drift originates from two mechanisms: sensor sensitivity decay and heating element efficiency decline. Both mechanisms introduce a systematic bias in the batch control sequence that is monotonically correlated with the batch time sequence, rather than fluctuations in the operating conditions themselves. If aging drift is not removed, the actual operating condition deviation in the batch-to-batch differences will overlap with equipment performance degradation, leading to an overestimation of the subsequent operating condition baseline. Long-term polarization of the moisture content sensor electrodes leads to a systematically low output signal, and carbon buildup and blockage of the heater burner nozzles cause a simultaneous occurrence of lower temperature and higher air volume. Both mechanisms introduce a systematic bias in the batch control sequence that is monotonically correlated with the batch time sequence. In the batch control sequence, the time series of differences between the current batch column and each historical batch column are arranged in chronological order according to the historical batches. A linear trend is fitted to the time series of differences. When the trend slope passes the statistical significance test (p-value less than 0.05), an aging drift component is determined to exist. The slope estimate is multiplied by the batch time interval to obtain the aging drift amount corresponding to each historical batch. Subtracting this drift amount from the corresponding historical batch column in the batch control sequence yields the candidate sequence of environmental correction components for each historical batch. Variable columns that fail the test are considered to have insignificant aging drift and are directly retained as candidate sequences of environmental correction components. The remaining cross-batch deviations in the candidate sequences of environmental correction components for each historical batch no longer contain a monotonic aging trend component. These remaining deviations reflect batch-to-batch changes introduced by non-aging factors such as seasonal changes in environmental temperature and humidity and parameter resets caused by equipment maintenance operations. The time series of the mean of all candidate sequences of environmental correction components for historical batches constitutes the environmental correction component.

[0052] Based on the material stratification parameters, multiple operating condition sampling points are set for the batch control sequence to generate a normal operating condition fluctuation range. The operating condition sampling points are set according to the distribution position of each moisture content layer in the longitudinal direction of the cavity in the material stratification parameters. The center position of each moisture content layer corresponds to one operating condition sampling point. The sampling point position is aligned with the longitudinal sensor layout layer of the cavity, and the number of sampling points is equal to the number of layers in the material stratification parameters. After grouping the four-variable observation rows at each time point in the batch control sequence according to the sensor corresponding to the sampling point, the historical batch column values ​​within each group are extracted as the historical observation set for that sampling point. The mean and standard deviation of the historical observation set describe the historical normal fluctuation characteristics of the aquifer within the corresponding disturbance change range. When the operating condition response mode categories corresponding to each aquifer are different, the upper bound of temperature fluctuation for temperature-dominant layers and the upper bound of humidity fluctuation for humidity-dominant layers are appropriately widened. The upper bound of the fluctuation range of each layer is the mean plus twice the standard deviation, and the lower bound is the mean minus twice the standard deviation. When the number of historical batches is small, the upper and lower bounds are estimated to be moderately contracted towards the global mean of the entire batch to prevent the range from being artificially high. The fluctuation range of each layer together constitutes the normal operating condition fluctuation range of the strata. After summarizing the normal operating condition fluctuation range of all sampling points, the connection method of the fluctuation range of adjacent sampling points is determined according to the moisture content gradient between layers. When the gradient is large, obvious jumps between adjacent layers are allowed, and when the gradient is small, smooth transition is allowed. The upper and lower bounds of the normal operating condition fluctuation range of each variable at each time are taken as the envelope of the fluctuation range of all sampling points at the corresponding time. The upper layer of the cavity has high moisture content and vigorous evaporation. Under the same air volume disturbance, the humidity response amplitude can be 2 to 3 times that of the lower layer of low moisture content grain. The envelope operation includes the response amplitudes of the high and low moisture content layers into the normal operating condition fluctuation range at the same time, so as to avoid the normal operating condition being incorrectly marked as abnormal due to the difference in layer characteristics.

[0053] Statistical analysis is performed on the normal operating condition fluctuation range and environmental correction components to form the operating condition baseline. The statistical analysis aligns the normal operating condition fluctuation range and environmental correction components on the time axis and then superimposes them moment by moment. The environmental correction component is added as a bias to the upper and lower bounds of the normal operating condition fluctuation range. The correction direction is consistent with the sign of the environmental correction component; when the environmental correction component is positive, the upper and lower bounds shift upwards synchronously, and when it is negative, they shift downwards synchronously. The shift amount is equal to the mean amplitude of the environmental correction component at the corresponding moment. This shift operation incorporates the differences in environmental conditions from historical batches into the calibration of the normal fluctuation range, aligning the center position of the operating condition baseline with the actual environmental conditions of the current batch. After superposition, the upper and lower bounds at each moment are used to perform sliding smoothing across the disturbance variation intervals for each variable component. The smoothing window length is half the width of the adjacent disturbance variation interval. Smoothing eliminates the upper and lower bound jumps introduced by segmented statistics at the interval boundaries, ensuring a continuous transition of the operating condition baseline at the interval boundaries. Smoothing across stage boundaries prohibits crossing the stage type identifier's forced cutoff point to preserve the baseline difference characteristics between stages. The upper and lower bounds of the smoothed variable components together constitute the operating condition baseline. The variation of the width between the upper and lower bounds of the operating condition baseline at each time point reflects the dynamic change of the tolerance range of normal operating condition fluctuations at different drying stages. The operating condition baseline width is usually stable and relatively wide in the constant speed section. The operating condition baseline width narrows in the deceleration section due to the nonlinearity of heat and mass transfer. The operating condition baseline width is the narrowest in the final completion section. If an abnormal expansion of the baseline width occurs in the middle of the deceleration section, it usually indicates that the distribution of bound water migration path inside the batch of grain is uneven.

[0054] The control confidence boundary is determined by removing changes within the operating condition baseline from the operating offset amplitude. The removal operation compares the time series of each variable component of the operating offset amplitude with the upper and lower bounds of normal fluctuations of the corresponding variable in the operating condition baseline moment by moment. Values ​​in the operating offset amplitude time series falling within the upper and lower bounds of the operating condition baseline are identified as offsets caused by normal operating condition fluctuations and are removed from the operating offset amplitude. The remaining offset after removal is the true control deviation exceeding the normal fluctuation range. The removal ratio is estimated by the ratio of the number of times each variable component falls within the operating condition baseline range in each disturbance variation interval to the total number of times in the interval. A high removal ratio for the temperature component indicates that the batch-to-batch offset of this variable is mainly contributed by normal thermal inertia fluctuations rather than control misconduct. A low removal ratio for the humidity component indicates that normal fluctuations account for a small proportion of the humidity offset, while the true control deviation accounts for a large proportion. The difference in the removal ratios of the two types of variables reflects the stability differentiation of each control loop under the current operating conditions. When the removal ratio of a certain variable is extremely high in multiple consecutive disturbance variation intervals, the control confidence boundary in the corresponding direction can be appropriately relaxed. The upper envelope of the remaining offset of each variable after elimination is defined as the upper bound of the control confidence boundary, and the lower envelope is defined as the lower bound of the control confidence boundary. The width of the control confidence boundary reflects the reliable deviation tolerance range of the control system after eliminating normal operating condition disturbances. When the width of the temperature and humidity control confidence boundary is significantly different, it indicates that there is a systematic differentiation in the batch consistency of the two control loops. The two loops should adopt different adjustment step sizes when generating subsequent control commands.

[0055] Abnormal operating conditions are selected from the operating condition baseline matrix based on the control confidence boundary to form a control deviation record group. After aligning the upper and lower bounds of the control confidence boundary with the time axis of the operating condition baseline matrix, each time step in the time sequence of each column of the operating condition baseline matrix is ​​compared with the upper and lower bounds of the corresponding variables of the control confidence boundary. When the normalized value of a variable exceeds the upper bound or falls below the lower bound of the control confidence boundary at a certain time, that time is determined to be an abnormal operating condition moment for that variable. The determination of abnormal operating condition moments is performed independently among the four variables. At a certain time, only the temperature may exceed the limit while other variables are normal, or multiple variables may exceed the limit simultaneously. When a single variable exceeds the limit, it is marked as a single variable abnormal event. When multiple variables exceed the limit simultaneously, it is marked as a multivariate joint abnormal event. The danger level of joint abnormal events is higher than that of single variable abnormalities. In the middle of the deceleration section of a certain batch, a joint abnormality occurred in which the temperature and humidity simultaneously exceeded the upper bound of the control confidence boundary, lasting for about 6 minutes. During this period, the cavity hot air temperature rose sharply and was accompanied by a decrease in dehumidification efficiency. The combination of these two factors indicates that there is a miscoordination between heating and dehumidification control during this period. The control deviation record group consists of three elements: the four variable observations at all abnormal operating conditions, the direction of the out-of-bounds error, and the magnitude of the out-of-bounds error. The magnitude of the out-of-bounds error is defined as the absolute value of the difference between the variable value at that moment and the corresponding boundary value of the control confidence boundary. The magnitude of the out-of-bounds error of the joint abnormal event in the control deviation record group is taken as the weighted average of the magnitudes of the variables involved in the out-of-bounds error. The weight is determined by the reciprocal of the boundary width of each variable in the control confidence boundary. The narrower the boundary, the higher the contribution weight of the variable to the joint abnormal magnitude. The control deviation record group is arranged in descending order of the magnitude of the out-of-bounds error.

[0056] Step S150: Perform time-series correlation analysis between the standard operating parameters and the drying joint control benchmark to obtain the control iteration factor. Based on the control iteration factor, correct the control deviation record group, label the drying level, and output the temperature, humidity and wind coordinated control command.

[0057] In some embodiments, the step of performing time-series correlation analysis between the standard operating parameters and the drying joint control benchmark to obtain the control iteration factor includes: aligning the standard operating parameters and the drying joint control benchmark in time to generate a correlated state sequence; identifying the saturation trigger threshold by performing moisture saturation critical identification on the correlated state sequence; performing iterative convergence analysis based on the saturation trigger threshold to generate a convergence error distribution; and extracting correction coefficients from the convergence error distribution to form the control iteration factor.

[0058] The standard operating parameters and the drying joint control benchmark are time-series aligned to generate a correlated state sequence. The core constraint of this time-series alignment is that both the standard operating parameters and the drying joint control benchmark are established based on the same batch's stage type identifier, and their stage boundary times naturally correspond on the standard time axis. The alignment operation uses each stage boundary time as an anchor point to correlate each column of the standard operating parameters' time sequence with the corresponding four-variable control benchmark interval of the drying joint control benchmark for each stage. After correlation, each time point forms a joint observation row, containing twelve columns of values: the measured four-variable values ​​of the standard operating parameters at that time, and the upper and lower bounds of the four-variable values ​​of the corresponding time point in the drying joint control benchmark. Time points in the standard operating parameters that are not covered by a corresponding stage in the drying joint control benchmark are marked as uncovered rows. Uncovered rows typically appear during the shutdown cooling stage after the completion phase at the end of the batch, when heat and mass transfer is no longer controllable, and these uncovered rows do not participate in subsequent correlation analysis. The difference between the measured value of the standard operating parameter in the joint observation row and the center value of the upper and lower bounds of the drying joint control reference is defined as the state deviation vector at that moment. The sign of each component of the deviation vector reflects whether the measured value is higher or lower than the center of the control reference. In the middle section of the deceleration segment of a certain batch, the temperature state deviation is continuously positive and the absolute value gradually increases. Correspondingly, the temperature component of the deviation vector continuously exceeds the humidity and air volume components during this period. All joint observation rows are arranged in chronological order to form a related state sequence. The time length of the related state sequence is consistent with the effective drying time of the current batch, and the inter-row time resolution is the same as the common sampling reference of the standard operating parameters.

[0059] The dehumidification saturation threshold is determined by identifying the dehumidification saturation threshold in the associated state sequence. Dehumidification saturation threshold corresponds to the state where the accumulated humidity within the cavity reaches the upper limit of the dehumidification capacity. At this point, further increasing the airflow cannot further improve the dehumidification rate, and the sensitivity of the humidity response to changes in airflow significantly decreases. A typical temporal characteristic of dehumidification saturation is that the humidity column state deviation in the associated state sequence tends from negative to zero or even turns into a sustained positive value. The temporal slope of the humidity column state deviation in the associated state sequence is calculated using a sliding window first-order difference. The window length is one-eighth of the standard duration of a single stage in the drying joint control benchmark. When the slope turns from negative to positive within three consecutive sliding windows and the absolute value exceeds 0.5 times the temporal standard deviation of the humidity deviation, the corresponding moment is identified as a candidate dehumidification saturation threshold. When the operator manually and significantly reduces the airflow, the cavity's dehumidification capacity drops sharply, and the humidity deviation slope also reverses from negative to positive, exhibiting the same timing characteristics as dehumidification saturation. However, this is caused by human intervention rather than the cavity reaching its dehumidification limit. At this point, the airflow significantly deviates from the drying joint control benchmark. When the absolute value of the airflow deviation exceeds 20% of the boundary width, the exclusion condition is triggered. This candidate point is not confirmed as an effective critical point to avoid misjudging the humidity reversal caused by operational intervention as dehumidification saturation criticality. The saturation trigger threshold is determined by the average absolute value of the difference between the measured humidity value of the associated state sequence at the effective critical point and the upper limit of the drying joint control benchmark humidity. The average value is taken as the difference corresponding to all effective critical points in the entire batch. A batch identified four effective critical points with corresponding humidity differences of 0.07, 0.09, 0.08, and 0.10, respectively. The average value of approximately 0.085 is the saturation trigger threshold for this batch. Batches with a low saturation trigger threshold indicate that the cavity's dehumidification capacity margin is small, and the adjustable space for humidity deviation by adjusting the airflow is limited.

[0060] Iterative convergence analysis based on the saturation trigger threshold generates a convergence error distribution. The iterative convergence analysis simulates the convergence process of the control system under the constraints of current standard operating parameters and the drying joint control benchmark, using the saturation trigger threshold as the trigger condition to gradually correct the four-variable deviations. The analysis objective is to quantify the convergence speed and residual error distribution characteristics of each variable deviation during iterative adjustment. The iterative process uses the state deviation vector at each moment of the associated state sequence as the initial input. When the absolute value of the cavity humidity deviation exceeds the saturation trigger threshold, it indicates that the pressure in the dehumidification channel has reached its limit, and the control system must increase the airflow to force dehumidification. The airflow adjustment is determined by multiplying the difference between the humidity excess and the saturation trigger threshold by the adjustment gain coefficient. The adjustment gain coefficient is obtained by normalizing the inverse of the airflow boundary width of the drying joint control benchmark. Increasing the airflow removes some heat while dehumidifying, causing a decrease in cavity temperature. The airflow adjustment is fed back to the temperature and humidity deviation components to update the state deviation vector, with the update rule being δT. _new = δT_old - g_T × ΔQ, δH_new = δH_old - g_H × ΔQ, where δT_old and δH_old are the temperature and humidity deviation components before this iteration (normalized dimensionless values), δT_new and δH_new are the updated temperature and humidity deviation components, ΔQ is the airflow adjustment amount (normalized dimensionless value), and g_T and g_H are the dimensionless response gain coefficients of temperature and humidity to airflow adjustment, respectively. The response gain is statistically estimated from the measured temperature and humidity responses after multiple historical airflow disturbances in the associated state sequence. When the humidity deviation does not exceed the saturation trigger threshold, only the temperature deviation is proportionally adjusted, and the airflow linkage is not triggered. The state deviation vector at each time step is updated according to the above iteration rules until the absolute value of the deviation is less than 1 / 4 of the width of the confidence boundary of each variable control or the number of iterations exceeds the upper limit of 10. The residual deviation of each variable at the time of stopping is recorded as the convergence error at that time. When the number of iterations reaches the upper limit of 10 and still fails to converge, the convergence error is usually significantly higher than that of the normal convergence time. Such times form a local high value cluster in the convergence error distribution. The convergence errors of all times are arranged on the time axis according to the variable components to form the convergence error distribution. If the convergence error of a certain time in the convergence error distribution is consistently high, it indicates that the control system cannot converge the deviation to the confidence range under the current parameters at that time.

[0061] Correction coefficients are extracted from the convergence error distribution to form the control iteration factor. The temporal amplitude of each variable component in the convergence error distribution reflects the strength of the control convergence ability of that variable under the current batch calorific value. Periods with persistently high amplitudes indicate insufficient response of the variable to the control action during that period, while periods with persistently low amplitudes indicate sufficient control margin. The control system should apply different intensities of iterative correction for the two types of periods. The correction coefficient F_i is determined by the formula F_i=1+γ×(E_i-E_ref) / E_ref, where E_i is the root mean square value of each variable component at time i of the convergence error distribution, E_ref is the median of the root mean square value of the convergence error distribution of the current batch over the entire period, and γ is a sensitivity coefficient taken as 1.5. When F_i is greater than 1, it indicates that the control intensity needs to be increased at that time, and when F_i is less than 1, it indicates that the control intensity can be appropriately reduced. The calculation of F_i is performed independently in each drying stage to avoid confusion of the correction direction due to differences in error amplitude across stages. The control iteration factor is composed of the correction coefficient F_i time series of all batches. The time series where F_i exceeds 1.8 is marked as a strong correction time. When strong correction times are concentrated in a certain stage, it indicates that the drying joint control benchmark setting is too tight in that stage, causing the control system to frequently enter the strong correction state. In the second half of the deceleration phase of the wheat batch, the resistance to migration of bound water in the starch granules inside the grain is much higher than that of rice. The expected rate of moisture content decrease of the drying joint control benchmark established by the control system based on rice experience is significantly higher than the actual value of wheat. The control system continuously increased the adjustment intensity for 12 consecutive minutes but still could not bring the deviation within the benchmark. The corresponding F_i continuously exceeded 1.8. The proportion of strong correction times in the whole batch exceeded 20%, which clearly indicates that the drying joint control benchmark of this batch is too tight for the wheat variety as a whole. When establishing the benchmark for the next batch, the control interval boundary of the second half of the deceleration phase should be appropriately widened with reference to the historical data of wheat heat and mass transfer.

[0062] The control deviation record group is corrected based on the adjustment iteration factor to output the temperature, humidity and air coordinated control command for drying level labeling. The abnormal operating conditions in the control deviation record group are arranged in descending order of the out-of-bounds amplitude. The F_i value of the adjustment iteration factor at the corresponding time is multiplied by the out-of-bounds amplitude at that time to obtain the corrected deviation amplitude. If the corrected deviation amplitude is greater than the original out-of-bounds amplitude, it indicates high convergence difficulty in control at that time; if the corrected deviation amplitude is less than the original out-of-bounds amplitude, it indicates that the control system has strong self-correction capability at that time. The corrected deviation amplitude replaces the original out-of-bounds amplitude as the scoring basis for drying level labeling. The drying level labeling is mainly based on the ratio of the corrected deviation amplitude to the control confidence boundary width. Below 50% is considered a slight deviation, between 50% and 100% is considered a moderate deviation, and above 100% is considered a severe deviation. The level of a joint abnormal event is determined by the variable with the largest corrected deviation amplitude among the variables involved in the out-of-bounds event. When temperature and humidity jointly exceed the limits in a certain batch, the humidity corrected amplitude reaches 110% of the control confidence boundary width; this joint event is judged as a severe deviation. The temperature, humidity, and air coordinated control commands are generated based on the dryness level. Slight deviations correspond to single-variable fine-tuning commands, which only apply small corrections to variables exceeding the limits without linking other variables to avoid introducing additional disturbances. Moderate deviations correspond to temperature, humidity, and air coordinated process control commands, where variables exceeding the limits and associated response variables are adjusted synchronously to maintain thermo-mass balance. Severe deviations correspond to temperature, humidity, and air coordinated adjustment commands, where all three variables are adjusted synchronously to quickly pull back the deviation. The adjustment amount at each level is determined by multiplying the corrected deviation amplitude by the reciprocal of the corresponding variable's control confidence boundary width. The air volume adjustment amount is limited to the control confidence boundary width. Temperature, humidity, and air coordinated control commands are issued in descending order of the corrected deviation amplitude.

[0063] To implement the above-described method embodiments, a multivariate collaborative drying control method based on a large model is proposed to achieve the corresponding functions and technical effects. See also... Figure 2 , Figure 2 This paper presents a structural block diagram of a multivariable collaborative drying control system 200 based on a large model, according to an embodiment of this application, including: Data acquisition module 201 is used to collect operating condition data of the drying chamber, including moisture content, temperature, humidity and air volume, and to perform synchronous normalization processing on the operating condition data to establish an operating condition reference matrix. The parameter calibration module 202 is used to identify the temperature and humidity coupling offset of the operating condition reference matrix to determine the thermal mass compensation weight, identify thermal mass cross-disturbance based on the operating condition reference matrix to determine the disturbance attenuation parameter, and perform variable differential calibration based on the disturbance attenuation parameter and the thermal mass compensation weight to form standard operating parameters. The reasoning and decision module 203 is used to input the standard operating parameters into the large model to perform comprehensive reasoning on the coupling state of water content gradient and temperature, humidity and wind to obtain reasoning and decision weights, and to perform weighted mapping on the working condition benchmark matrix according to the reasoning and decision weights to generate a coupling response spectrum. Deviation detection module 204 is used to establish a drying joint control benchmark based on the coupled response spectrum, perform multi-batch difference detection based on the drying joint control benchmark to determine the operating deviation amplitude, determine the control confidence boundary after eliminating operating condition disturbances based on the operating deviation amplitude, and select abnormal operating conditions from the operating condition benchmark matrix to form a control deviation record group based on the control confidence boundary. The instruction output module 205 is used to perform time-series correlation analysis between the standard operating parameters and the drying joint control benchmark to obtain the control iteration factor, and to correct the control deviation record group according to the control iteration factor to output the temperature, humidity and wind coordinated control instruction for drying level labeling.

[0064] The aforementioned multivariable collaborative drying control system 200 based on a large model can implement a multivariable collaborative drying control method based on a large model as described in the above method embodiments. The options in the above method embodiments are also applicable to this embodiment, and will not be detailed here. The remaining content of this application's embodiments can be referred to the content of the above method embodiments, and will not be repeated in this embodiment.

[0065] The above embodiments are not an exhaustive list based on the present invention, and there may be many other embodiments not listed. Any substitutions and improvements made without departing from the concept of the present invention are within the protection scope of the present invention.

Claims

1. A multivariate collaborative drying control method based on a large model, characterized in that, include: The operating condition data of the drying chamber, including moisture content, temperature, humidity and air volume, are collected, and the operating condition data are synchronously normalized to establish an operating condition reference matrix. The temperature and humidity coupling offset of the operating condition reference matrix is ​​identified to determine the thermal mass compensation weight. Based on the operating condition reference matrix, thermal mass cross-disturbance is identified to determine the disturbance attenuation parameter. Based on the disturbance attenuation parameter and the thermal mass compensation weight, variable differentiation calibration is performed to form standard operating parameters. The standard operating parameters are input into the large model to perform comprehensive reasoning on the coupling state of moisture content gradient and temperature, humidity and wind to obtain reasoning decision weights. Based on the reasoning decision weights, the operating condition benchmark matrix is ​​weighted and mapped to generate a coupling response spectrum. A drying joint control benchmark is established based on the coupled response spectrum. Multiple batch difference detections are performed based on the drying joint control benchmark to determine the operating deviation amplitude. The control confidence boundary after eliminating operating condition disturbances is determined based on the operating deviation amplitude. Abnormal operating conditions are screened from the operating condition benchmark matrix according to the control confidence boundary to form a control deviation record group. The standard operating parameters are correlated with the drying joint control benchmark by time series analysis to obtain the control iteration factor. The control deviation record group is corrected according to the control iteration factor, and the drying level is marked and the temperature, humidity and wind coordinated control command is output.

2. The method according to claim 1, characterized in that, The step of identifying the temperature and humidity coupling offset of the operating condition reference matrix to determine the thermal mass compensation weights includes: The temperature and humidity wind time series curves are extracted from the aforementioned operating condition baseline matrix to construct a temperature and humidity wind joint response diagram. Hysteresis disturbance analysis was performed on the temperature, humidity and wind joint response diagram to extract the hysteresis response window and the air volume disturbance identifier; Based on the air volume disturbance identifier and the hysteresis response window, a coupled offset segment division is performed to generate an offset intensity sequence; The offset intensity sequence is subjected to heat and mass transfer correlation analysis and mapping to generate heat and mass compensation weights.

3. The method according to claim 1, characterized in that, The step of forming standard operating parameters by differentially calibrating variables based on the disturbance attenuation parameters and the thermal mass compensation weights includes: The disturbance attenuation parameters are grouped according to variable type to generate variable sensitivity weights; The sensitivity weights of the variables are sorted by moisture content sensitivity to generate a stage switching signal; A differentiated calibration benchmark is formed by performing stage adaptive benchmark mapping based on the stage switching signal and the variable sensitivity weight; The differentiated calibration benchmark and the thermal mass compensation weight are integrated to form standard operating parameters.

4. The method according to claim 1, characterized in that, The step of inputting the standard operating parameters into the large model to perform comprehensive reasoning on the coupling state of moisture content gradient and temperature, humidity and wind to obtain reasoning decision weights includes: The standard operating parameters are used to identify the drying stage and generate a stage type identifier; The process state input sequence is generated by structurally encoding the stage type identifier and the standard operating parameters; The process state input sequence is input into a large model to perform comprehensive reasoning on the coupling state of moisture content gradient and temperature, humidity and wind, and to obtain the gradient response component and the evaluation result of coupling strength. The inference decision weights are generated by fusing the gradient response components with the coupling strength evaluation results.

5. The method according to claim 1, characterized in that, The determination of operational deviation based on the drying joint control benchmark through multi-batch difference detection includes: Based on the aforementioned drying joint control benchmark, batch registration is performed to obtain homogeneous operating sequences; The thermal inertia attenuation characteristics of the cavity are extracted from the homogeneous running sequence to generate a thermal inertia compensation coefficient; The difference sequence is generated by extracting the difference between adjacent batches of the same operating sequence based on the thermal inertia compensation coefficient; The extreme value screening and calibration of the running offset range are performed based on the difference sequence.

6. The method according to claim 1, characterized in that, The determination of the control reliability boundary after removing the operating condition disturbance based on the operating offset amplitude includes: The disturbance variation range is obtained by segmenting the operating offset amplitude into working condition periods; Within the disturbance variation range, operating mode identification is performed to generate material stratification parameters; Based on the material stratification parameters and the disturbance change range, the normal operating condition fluctuation range is extracted to generate the operating condition baseline; The control confidence boundary is determined by removing the changes within the operating condition baseline from the operational offset magnitude.

7. The method according to claim 1, characterized in that, The step of performing a time-series correlation analysis between the standard operating parameters and the drying joint control benchmark to obtain the control iteration factor includes: The standard operating parameters are time-series aligned with the drying joint control benchmark to generate a correlated state sequence; The saturation trigger threshold is determined by identifying the critical moisture saturation level of the associated state sequence. Based on the saturation trigger threshold, iterative convergence analysis is performed to generate a convergence error distribution; The convergence error distribution is modified by extracting correction coefficients to form the adjustment iteration factor.

8. The method according to claim 3, characterized in that, The step of forming a differentiated calibration benchmark by performing stage-adaptive benchmark mapping based on the stage switching signal and the variable sensitivity weights includes: Perform cavity temperature gradient distribution analysis on the sensitivity weights of the variables and the stage switching signal to locate the critical point of temperature attenuation; Spatial hierarchical pattern identification is performed on the temperature attenuation critical point to generate a temperature distribution feature set; The batch configuration parameters are generated by analyzing the batch variation of the temperature distribution feature set. Based on the batch configuration parameters, a phased adaptive benchmark mapping is performed on the temperature zone attenuation critical point to generate a differentiated calibration benchmark.

9. The method according to claim 6, characterized in that, The step of extracting the normal operating condition fluctuation range and generating the operating condition baseline based on the material stratification parameters and the disturbance change range includes: Based on the perturbation variation range, a batch control sequence is generated by synchronous control registration; The equipment aging drift component is removed from the batch control sequence to generate an environmental correction component; Based on the material stratification parameters, multiple operating condition sampling points are set for the batch control sequence to generate a normal operating condition fluctuation range; A baseline operating condition is formed by statistically analyzing the normal operating condition fluctuation range and the environmental correction component.

10. A multivariable collaborative drying control system based on a large model, characterized in that, include: The data acquisition module is used to collect operating condition data of the drying chamber, including moisture content, temperature, humidity and air volume, and to perform synchronous normalization processing on the operating condition data to establish an operating condition reference matrix. The parameter calibration module is used to identify the temperature and humidity coupling offset of the operating condition reference matrix to determine the thermal mass compensation weight, identify thermal mass cross-disturbance based on the operating condition reference matrix to determine the disturbance attenuation parameter, and perform variable differential calibration based on the disturbance attenuation parameter and the thermal mass compensation weight to form standard operating parameters. The reasoning and decision module is used to input the standard operating parameters into the large model to perform comprehensive reasoning on the coupling state of moisture content gradient and temperature, humidity and wind to obtain reasoning and decision weights, and to perform weighted mapping on the working condition benchmark matrix according to the reasoning and decision weights to generate a coupling response spectrum. The deviation detection module is used to establish a drying joint control benchmark based on the coupled response spectrum, perform multi-batch difference detection based on the drying joint control benchmark to determine the operating deviation amplitude, determine the control confidence boundary after eliminating operating condition disturbances based on the operating deviation amplitude, and select abnormal operating conditions from the operating condition benchmark matrix to form a control deviation record group based on the control confidence boundary. The instruction output module is used to perform time-series correlation analysis between the standard operating parameters and the drying joint control benchmark to obtain the control iteration factor, and to correct the control deviation record group according to the control iteration factor, and to output the temperature, humidity and wind coordinated control instruction for drying level labeling.

Images