Early warning method and system for caking in solid beverage granulation process

Abstract

The invention belongs to the field of intelligent manufacturing and industrial process control, and relates to a caking early warning method and system in a solid beverage granulation process. The method comprises the steps that the material temperature, the environment humidity, the stirring motor torque and the adhesive spraying rate are collected in real time, and the fluctuation entropy of torque time sequence data is calculated; determining reference noise distribution according to the spraying rate variance, calculating a noise adjustment coefficient in combination with the fluctuation entropy, and generating an enhanced feature set by the disturbance parameter; constructing a multi-decision tree early warning model, and setting an information gain threshold value of node splitting according to a noise adjustment coefficient to control the depth of the tree; generating a verification set by disturbing historical boundary samples, and calculating the weight of each tree according to the confusion matrix; and performing weighted combination on classification results of all the trees, and outputting a caking early warning signal. According to the invention, accurate and reliable real-time early warning of the caking risk in the solid beverage granulation process can be realized, and the production stability and the product qualification rate can be improved in an assisted manner.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of intelligent manufacturing and industrial process control, and particularly relates to a solid beverage granulation process caking early warning method and system. BACKGROUND

[0002] With the transformation of the food industry to high-end manufacturing and intelligentization, the granulation process of solid beverage as a core link, the stability of its production process directly determines the solubility, fluidity and final quality of the product. The granulation process is a complex system involving thermodynamics, fluid mechanics and mechanical stirring, which is extremely sensitive to changes in parameters such as material temperature, environmental humidity, stirring torque and binder spraying rate. In actual production, the non-linear fluctuations of the above parameters are extremely easy to cause material caking, and then cause equipment blockage, unplanned shutdown and raw material waste.

[0003] In the existing manufacturing equipment, the monitoring technology for caking mainly experienced the following two development stages, but there are still significant deficiencies in dealing with complex working conditions: First, the traditional monitoring stage based on threshold and simple statistics. The early granulation equipment mainly relies on single sensor threshold alarm or simple statistical process control (SPC). However, the granulation site environment is complex, and the sensor signal often mixes a large amount of random noise and process disturbance. The traditional method is difficult to strip the noise influence, and cannot effectively capture the nonlinear coupling relationship between multiple variables. This one-size-fits-all static threshold is extremely easy to produce false alarms or omissions when facing small dynamic changes in the production environment.

[0004] Second, the preliminary intelligent stage based on static machine learning. In order to improve the monitoring accuracy, some current devices begin to introduce decision tree, random forest and other machine learning algorithms for fault diagnosis. The existing decision tree model usually fixes the depth of the tree as a preset hyperparameter, but in the actual granulation process, the smoothness of the working condition is real-time changing. When the working condition fluctuates violently, the model is easy to learn noise; when the working condition is stable, the model cannot capture subtle features, and lacks the ability to dynamically adjust the model complexity according to the real-time working condition. In addition, in the ensemble learning model such as random forest, the same weight is usually given to all base learners. This ignores the performance difference of different decision trees in processing specific data distribution, and fails to fully utilize the advantages of expert models, resulting in a decrease in early warning accuracy in critical states. SUMMARY

[0005] Therefore, the purpose of the present application is to provide a solid beverage granulation process caking early warning method and system, to solve the technical problems that the traditional monitoring method is sensitive to noise and easy to produce false alarms and omissions, and the existing machine learning model structure and weight are fixed, and it is difficult to dynamically adjust according to the real-time working condition, resulting in overfitting or underfitting of the model in complex working conditions.

[0006] To solve the above problems, the technical scheme of the solid beverage granulation process caking early warning method provided by the present application is: The solid beverage granulation process caking early warning method comprises the following steps: Obtain real-time process parameters of the solid beverage granulation process, the real-time process parameters including material temperature, environmental humidity, stirring motor torque and binder spraying rate, and calculate fluctuation entropy based on time series data of the stirring motor torque; Determine a reference noise distribution according to the real-time variance of the binder spraying rate; calculate a noise adjustment coefficient negatively correlated with the fluctuation entropy value; combine the reference noise distribution and the noise adjustment coefficient to perturb the real-time process parameters and the fluctuation entropy, and generate an enhanced feature set; Construct an early warning model composed of multiple decision trees, and when constructing each decision tree, determine an information gain threshold according to the noise adjustment coefficient, and perform node splitting only when the information gain of the candidate split feature is greater than the information gain threshold, so as to control the depth of the decision tree; For each constructed decision tree, combine the reference noise distribution and the noise adjustment coefficient to perturb the classification boundary samples in the historical data to generate a boundary verification set, and calculate an output weight based on the classification confusion matrix of the decision tree on the boundary verification set; input the enhanced feature set into the early warning model, and combine the classification results of each decision tree according to the output weight of each decision tree to obtain a caking early warning signal.

[0007] Further, the calculation of the fluctuation entropy based on the time series data of the stirring motor torque comprises: Take 1 second as a sampling interval, obtain the stirring motor torque data of the past 60 seconds to form a time series with a length of 60; Reconstruct the phase space of the time series, and set the embedding dimension m and the time delay τ; Calculate the number of vector pairs in the reconstructed phase space whose Chebyshev distance is less than a preset threshold r, and the preset threshold r is set to 0.2 times the standard deviation of the original time series; According to the above statistical results, the fluctuation entropy value is calculated by using the sample entropy formula.

[0008] Further, the determination of the reference noise distribution according to the real-time variance of the binder spraying rate comprises: Set the reference noise distribution as a Gaussian distribution with a mean of 0 , wherein the variance of the Gaussian distribution is equal to the variance value of the binder spraying rate data collected in the past 120 seconds.

[0009] Furthermore, the noise adjustment coefficient is calculated using an exponential decay function, and the noise adjustment coefficient is defined as follows: The calculation formula is: ;in, This represents the current fluctuation entropy value. This is the preset attenuation coefficient.

[0010] Furthermore, the method for calculating the information gain threshold is as follows: Information gain threshold The calculation formula is ;in, The noise adjustment coefficient is... The baseline information gain.

[0011] Furthermore, the method for obtaining the boundary verification set is as follows: The historical data is trained using a support vector machine, and all data points corresponding to the support vectors are identified as classification boundary samples. For each classification boundary sample, obtain the time series data of the stirring motor torque at the corresponding time and calculate the fluctuation entropy, and calculate the corresponding noise adjustment coefficient based on the fluctuation entropy; For each feature vector of a classification boundary sample, a noise vector is generated by randomly sampling from the baseline noise distribution and multiplying it by the noise adjustment coefficient corresponding to the feature vector; The generated noise vector is added to the feature vector of the original classification boundary sample to generate perturbed samples. The set of all perturbed samples constitutes the boundary validation set.

[0012] Further, the calculation of output weights based on the classification confusion matrix of the decision tree on the boundary validation set includes: Count the number of true positives in the classification results of the decision tree on the boundary validation set. Number of false positives , number of true negatives and the number of false negatives ; According to the formula Calculate precision ; According to the formula Calculate recall ; According to the formula Calculated Fractions, of which Set a preset, extremely small positive number to prevent the denominator from being zero; calculate the... The scores serve as the output weights of the decision tree.

[0013] The technical solution of the solid beverage granulation process agglomeration early warning system proposed in this invention is as follows: A clumping early warning system for solid beverage granulation process includes the following modules: The calculation module is used to obtain real-time process parameters of the solid beverage granulation process. The real-time process parameters include material temperature, ambient humidity, stirring motor torque and binder spraying rate, and calculate the fluctuation entropy based on the time series data of stirring motor torque. The generation module is used to determine the baseline noise distribution based on the real-time variance of the adhesive spraying rate; calculate the noise adjustment coefficient negatively correlated with the fluctuation entropy value using the fluctuation entropy; and combine the baseline noise distribution and the noise adjustment coefficient to perturb the real-time process parameters and fluctuation entropy to generate an enhanced feature set. The control module is used to construct an early warning model composed of multiple decision trees. When constructing each decision tree, an information gain threshold is determined based on the noise adjustment coefficient. Node splitting is only performed when the information gain of the candidate splitting feature is greater than the information gain threshold, thereby controlling the depth of the decision tree. The weighted output module is used to perturb the classification boundary samples in historical data to generate a boundary validation set for each constructed decision tree, in combination with the baseline noise distribution and noise adjustment coefficient, and to calculate the output weight based on the classification confusion matrix of the decision tree on the boundary validation set; the enhanced feature set is input into the early warning model, and the classification results of each decision tree are weighted and combined according to the output weight of each decision tree to obtain the clustering early warning signal.

[0014] Furthermore, the calculation of fluctuation entropy based on time-series data of the stirring motor torque includes: With a sampling interval of 1 second, the torque data of the stirring motor over the past 60 seconds was obtained, forming a time series of length 60. The time series is reconstructed in phase space, with the embedding dimension m and time delay τ set. Calculate the number of vector pairs in the reconstructed phase space where the Chebyshev distance between any two vectors is less than a preset threshold r, where the preset threshold r is set to 0.2 times the standard deviation of the original time series; Based on the above statistical results, the fluctuation entropy value was calculated using the sample entropy formula.

[0015] Further, determining the baseline noise distribution based on the real-time variance of the adhesive spraying rate includes: The reference noise distribution is set to a Gaussian distribution with a mean of 0. Among them, the variance of the Gaussian distribution The value is equal to the variance of the adhesive spraying rate data collected over the past 120 seconds.

[0016] Furthermore, the noise adjustment coefficient is calculated using an exponential decay function, and the noise adjustment coefficient is defined as follows: The calculation formula is: ;in, This represents the current fluctuation entropy value. This is the preset attenuation coefficient.

[0017] Furthermore, the method for calculating the information gain threshold is as follows: Information gain threshold The calculation formula is ;in, The noise adjustment coefficient is... The baseline information gain.

[0018] Furthermore, the method for obtaining the boundary verification set is as follows: The historical data is trained using a support vector machine, and all data points corresponding to the support vectors are identified as classification boundary samples. For each classification boundary sample, obtain the time series data of the stirring motor torque at the corresponding time and calculate the fluctuation entropy, and calculate the corresponding noise adjustment coefficient based on the fluctuation entropy; For each feature vector of a classification boundary sample, a noise vector is generated by randomly sampling from the baseline noise distribution and multiplying it by the noise adjustment coefficient corresponding to the feature vector; The generated noise vector is added to the feature vector of the original classification boundary sample to generate perturbed samples. The set of all perturbed samples constitutes the boundary validation set.

[0019] Further, the calculation of output weights based on the classification confusion matrix of the decision tree on the boundary validation set includes: Count the number of true positives in the classification results of the decision tree on the boundary validation set. Number of false positives , number of true negatives and the number of false negatives ; According to the formula Calculate precision ; According to the formula Calculate recall ; According to the formula Calculated Fractions, of which Set a preset, extremely small positive number to prevent the denominator from being zero; calculate the... The scores serve as the output weights of the decision tree.

[0020] The beneficial effects of this invention are as follows: By utilizing the fluctuation entropy and the real-time variance of the adhesive spraying rate to controllably perturb process parameters, this invention generates an enhanced feature set, enabling the early warning model to learn noise patterns in the process, thereby enhancing the model's resistance to signal fluctuations in actual production. When constructing the decision tree, the information gain splitting threshold is correlated with the noise adjustment coefficient reflecting the process state, avoiding invalid branches caused by data perturbation and suppressing overfitting. By generating a specific boundary validation set to assess the ability of each decision tree to distinguish fuzzy and difficult-to-distinguish samples, and assigning output weights accordingly, the overall reliability of the early warning model is improved, and false alarms and false negatives in clustered early warnings are reduced. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the steps of the solid beverage granulation process agglomeration early warning method of the present invention. Figure 2 This is a schematic diagram showing the relationship between fluctuation entropy and noise adjustment coefficient. Detailed Implementation

[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0023] Specific embodiments of the clumping early warning method for the solid beverage granulation process proposed in this invention: like Figure 1 As shown, the method for early warning of clumping in the granulation process of solid beverages includes the following steps: S1, acquire real-time process parameters of the solid beverage granulation process, including material temperature, ambient humidity, stirring motor torque and adhesive spraying rate, and calculate fluctuation entropy based on the time-series data of stirring motor torque.

[0024] The purpose of this step is to establish a multi-dimensional sensing system to comprehensively capture the thermodynamic, environmental, and kinetic factors affecting granulation quality, and to introduce fluctuation entropy as a core indicator for measuring system stability, quantifying the nonlinear dynamic changes during the stirring process. Fluctuation entropy can keenly reflect minute changes in the rheological properties of materials and is better able to indicate precursors to agglomeration than simple numerical values.

[0025] Temperature sensors, humidity sensors, motor current transformers, and spray flow meters are deployed on the granulation equipment to collect data on material temperature, ambient humidity, agitator motor torque, and binder spraying rate. The most recent minute's agitator motor torque time-series data is extracted as a sliding window. The data range within the sliding window is divided into several equally wide intervals. The frequency of data points falling into each interval is calculated, and the fluctuation entropy of this time-series data is calculated using the Shannon entropy formula, representing the stability of the granulation process.

[0026] In an optional embodiment, the fluctuation entropy is calculated based on time-series data of the stirring motor torque, including: With a sampling interval of 1 second, the torque data of the stirring motor over the past 60 seconds was obtained, forming a time series of length 60. The time series is reconstructed in phase space, with the embedding dimension m and time delay τ set. Calculate the number of vector pairs in the reconstructed phase space where the Chebyshev distance between any two vectors is less than a preset threshold r, where the preset threshold r is set to 0.2 times the standard deviation of the original time series; Based on the above statistical results, the fluctuation entropy value was calculated using the sample entropy formula.

[0027] Specifically, the embedding dimension is set to m=2, and the time delay τ=1. Torque readings of the stirring motor over the past minute are collected, resulting in a time series data point containing 60 values, such as 10.1 Nm, 10.2 Nm, up to 16 Nm. This one-dimensional time series is converted into a set of two-dimensional vectors: the first vector is (10.1, 10.2), the second is (10.2, 10.3), the third is (10.3, 10.4), and so on, generating a total of 59 two-dimensional vectors. The standard deviation of the entire original time series is calculated, assumed to be 0.5 Nm, and the distance threshold r is 0.1. The Chebyshev distance between any two vectors in the 59 two-dimensional vectors is calculated, which is the maximum difference between the corresponding coordinates of the two vectors, and the total number of vector pairs with a distance less than 0.1 is counted. These statistical results, along with similar statistical results in the three-dimensional reconstruction space, are substituted into the formula for calculating the sample entropy to obtain a fluctuation entropy value representing the complexity and uncertainty of the torque data.

[0028] S2, determine the baseline noise distribution based on the real-time variance of the adhesive spraying rate; calculate the noise adjustment coefficient negatively correlated with the fluctuation entropy value using the fluctuation entropy; combine the baseline noise distribution and the noise adjustment coefficient to perturb the real-time process parameters and fluctuation entropy to generate an enhanced feature set.

[0029] Specifically, the variance of the adhesive spraying rate within the most recent time window is calculated, and a Gaussian distribution with a mean of zero and a standard deviation proportional to the square root of the variance is used as the baseline noise distribution. An exponential decay function is used to calculate the noise adjustment coefficient, which is larger when the fluctuation entropy is smaller and the process is more stable, and vice versa. A noise value is randomly sampled from the baseline noise distribution, multiplied by the noise adjustment coefficient, and the result is added to the original five feature values: material temperature, ambient humidity, stirring motor torque, adhesive spraying rate, and fluctuation entropy. This process is repeated multiple times to generate a set of data containing controlled noise, forming an enhanced feature set.

[0030] In an optional embodiment, determining the baseline noise distribution based on the real-time variance of the adhesive spraying rate includes: The reference noise distribution is set to a Gaussian distribution with a mean of 0. Among them, the variance of the Gaussian distribution The value is equal to the variance of the adhesive spraying rate data collected over the past 120 seconds.

[0031] Specifically, the spraying rate of the adhesive is continuously monitored. At any given time, all rate readings from the previous 120 seconds are retrieved, for example, one data point is collected per second, resulting in a set of 120 values. The variance of these 120 rate data points is calculated, for example, if the result is 0.04. This variance is used as the variance of the baseline noise distribution. Therefore, the baseline noise distribution is determined to be a normal distribution with a mean of 0 and a variance of 0.04. The baseline noise distribution model tests the stability of the recent spraying process; when the spraying rate fluctuates greatly, the variance increases, and vice versa.

[0032] In an optional embodiment, the noise adjustment coefficient is calculated using an exponential decay function, and the noise adjustment coefficient is defined as follows: The calculation formula is: ;in, This represents the current fluctuation entropy value. This is the preset attenuation coefficient.

[0033] Specifically, it receives two inputs: one is the currently calculated fluctuation entropy value. The other is a pre-set attenuation coefficient. Assuming the current stirring process is very stable, the calculated fluctuation entropy... For a small value, such as 0.3, and the attenuation coefficient The default value is 2.0. According to the formula, the noise adjustment factor... Approximately 0.55. Conversely, if the stirring process involves violent fluctuations, the fluctuation entropy value... If it's a large value, such as 1.5, then the noise adjustment coefficient... Approximately 0.05. The more stable the production process, the lower the entropy value, and the larger the corresponding noise adjustment coefficient; the more unstable the process, the higher the entropy value, and the smaller the noise adjustment coefficient, such as... Figure 2 As shown, by allowing the baseline noise distribution model to learn from noisy data, it prevents the baseline noise distribution model from only remembering clean data features, thus remaining robust in the face of sensor jitter in real-world production.

[0034] S3. Construct an early warning model composed of multiple decision trees. When constructing each decision tree, determine an information gain threshold based on the noise adjustment coefficient. Node splitting is only performed when the information gain of the candidate splitting feature is greater than the information gain threshold, thereby controlling the depth of the decision tree.

[0035] The purpose of this step is to resolve the contradiction between overfitting and underfitting, that is, to dynamically adjust the complexity of the early warning model based on the stability of the operating conditions. When the process is unstable, the splitting threshold is increased to limit the growth of the decision tree and prevent the early warning model from fitting random noise; when the process is stable, the threshold is lowered to allow the early warning model to learn subtle features. This achieves adaptive adjustment of the model structure. Compared with traditional fixed-depth decision trees, this step ensures that the early warning model maintains optimal generalization ability under different operating conditions and effectively suppresses invalid branches caused by data perturbation.

[0036] Specifically, when training each decision tree of the early warning model, an information gain threshold is set using the noise adjustment coefficient corresponding to the current training sample. For example, the information gain threshold is proportional to the value minus the noise adjustment coefficient. When selecting the best splitting feature at each node of the decision tree, the information gain brought by that feature is calculated. Only when the information gain exceeds the set threshold is the node allowed to split; otherwise, the node becomes a leaf node. When the process is stable and the noise adjustment coefficient is high, the threshold is low, and the decision tree can grow deep; while when the process fluctuates greatly, the threshold is high, allowing pruning to prevent the early warning model from learning noise.

[0037] In an optional embodiment, the information gain threshold is calculated as follows: Information gain threshold The calculation formula is ;in, The noise adjustment coefficient is... The baseline information gain.

[0038] For example, the benchmark information gain value =0.4, this value was determined before model training. The noise adjustment coefficient calculated above is used. Adjust the aforementioned benchmark. Assume that at a certain moment, due to a stable production process, the calculated noise adjustment coefficient... The information gain threshold is 0.55. The calculated value is 0.18. If subsequent production processes become unstable, If it drops to 0.05, then the new information gain threshold... It will become 0.38. The information gain threshold is used to construct the decision tree. When production is stable, the information gain threshold is low, allowing more features to be selected; when production is unstable, the information gain threshold is high, and the selection criteria are strict.

[0039] S4. For each constructed decision tree, based on the baseline noise distribution and noise adjustment coefficient, perturb the classification boundary samples in the historical data to generate a boundary validation set, and calculate the output weights based on the classification confusion matrix of the decision tree on the boundary validation set; input the enhanced feature set into the early warning model, and perform a weighted combination of the classification results of each decision tree according to the output weights of each decision tree to obtain the clustering early warning signal.

[0040] This step avoids using ordinary samples, instead selecting the most difficult-to-distinguish boundary samples and adding noise to test each decision tree, evaluating its ability to discriminate in ambiguous regions. It breaks away from the traditional ensemble learning mechanism of average voting, giving greater weight to decision trees that perform well on difficult samples (high precision and recall), thus improving the reliability of early warnings. By focusing on the accuracy of identifying critical states, it effectively reduces the false alarm and false negative rates when the granulation process is at the critical point between agglomeration and non-agglomeration.

[0041] The Support Vector Machine (SVM) algorithm is used to identify samples closest to the classification hyperplane in historical data as classification boundary samples. Using the same perturbation method as for generating the augmented feature set, these boundary samples are perturbed multiple times to generate a boundary validation set specifically for testing the model's performance in ambiguous regions. This validation set is then input into a pre-trained single decision tree for classification. The number of true positives, false positives, true negatives, and false negatives in the prediction results are counted to construct a confusion matrix, and based on this matrix, the decision tree's performance is calculated. The score is used as the output weight of the decision tree.

[0042] The enhanced feature set generated under real-time conditions is input into all decision trees in the early warning model. Each decision tree outputs a classification result: normal or blocky. Two accumulators are set up, one for the normal class and one for the blocky class. All decision trees are traversed. If a decision tree predicts a blocky result, the output weight of that decision tree is added to the blocky accumulator; otherwise, it is added to the normal accumulator. The values ​​of the two accumulators are compared; the class with the larger value is the early warning signal output. For example, if the blocky accumulator value is 3.8 and the normal accumulator value is 2.1, then a blocky early warning signal is output.

[0043] In an optional embodiment, the boundary verification set is obtained as follows: The historical data is trained using a support vector machine, and all data points corresponding to the support vectors are identified as classification boundary samples. For each classification boundary sample, obtain the time series data of the stirring motor torque at the corresponding time and calculate the fluctuation entropy, and calculate the corresponding noise adjustment coefficient based on the fluctuation entropy; For each feature vector of a classification boundary sample, a noise vector is generated by randomly sampling from the baseline noise distribution and multiplying it by the noise adjustment coefficient corresponding to the feature vector; The generated noise vector is added to the feature vector of the original classification boundary sample to generate perturbed samples. The set of all perturbed samples constitutes the boundary validation set.

[0044] Specifically, a support vector machine (SVM) classifier is trained on a historical dataset containing both normal and abnormal operating conditions. All data points identified as support vectors are considered boundary samples because they are closest to the decision boundary and are the most difficult to distinguish. For example, data point X is a boundary sample with feature vectors of temperature 25.5°C and pressure 101.3°C. The stirring torque data corresponding to data point X at that time is found, the fluctuation entropy is calculated to be 0.8, and the noise adjustment coefficient is then calculated from this. The value is 0.2. A noise vector with the same feature dimension as X is randomly generated from a baseline noise distribution with a mean of 0 and a variance of 0.04, for example, (-0.2, 0.5). This noise vector is then multiplied by a noise adjustment coefficient. We obtain (-0.04, 0.1). The adjusted noise vector is added to the original feature vector to generate a new perturbation sample with features (25.46, 101.4). This process is repeated for all boundary samples; all the generated new samples together form the boundary validation set, used to test the model's ability to discriminate in ambiguous regions.

[0045] In an optional embodiment, calculating the output weights based on the classification confusion matrix of the decision tree on the boundary validation set includes: Count the number of true positives in the classification results of the decision tree on the boundary validation set. Number of false positives , number of true negatives and the number of false negatives ; According to the formula Calculate precision ; According to the formula Calculate recall ; According to the formula Calculated Fractions, of which Set a preset, extremely small positive number to prevent the denominator from being zero; calculate the... The scores serve as the output weights of the decision tree.

[0046] For example, a decision tree to be evaluated is used to classify a boundary validation set containing, for example, 200 perturbed samples. After classification, the confusion matrix is ​​statistically analyzed by comparing the results with the true labels of the samples; for example, 80 true positives, 20 false positives, 90 true negatives, and 10 false negatives. Performance metrics are calculated based on these statistics, with a very small value ε, such as 0.00001. Precision is calculated to be approximately 0.8, and recall is calculated to be approximately 0.889. These two values ​​are then used to calculate... The score is approximately 0.842. The score is specified as the output weight of the decision tree, which comprehensively reflects the completeness of the decision tree in handling difficult samples.

[0047] Specific embodiments of the solid beverage granulation process clumping early warning system proposed in this invention: A clumping early warning system for solid beverage granulation process includes the following modules: The calculation module is used to obtain real-time process parameters of the solid beverage granulation process. The real-time process parameters include material temperature, ambient humidity, stirring motor torque and binder spraying rate, and calculate the fluctuation entropy based on the time series data of stirring motor torque. The generation module is used to determine the baseline noise distribution based on the real-time variance of the adhesive spraying rate; calculate the noise adjustment coefficient negatively correlated with the fluctuation entropy value using the fluctuation entropy; and combine the baseline noise distribution and the noise adjustment coefficient to perturb the real-time process parameters and fluctuation entropy to generate an enhanced feature set. The control module is used to construct an early warning model composed of multiple decision trees. When constructing each decision tree, an information gain threshold is determined based on the noise adjustment coefficient. Node splitting is only performed when the information gain of the candidate splitting feature is greater than the information gain threshold, thereby controlling the depth of the decision tree. The weighted output module is used to perturb the classification boundary samples in historical data to generate a boundary validation set for each constructed decision tree, in combination with the baseline noise distribution and noise adjustment coefficient, and to calculate the output weight based on the classification confusion matrix of the decision tree on the boundary validation set; the enhanced feature set is input into the early warning model, and the classification results of each decision tree are weighted and combined according to the output weight of each decision tree to obtain the clustering early warning signal.

[0048] It should be noted that the processing procedures of the calculation module, generation module, control module, and weighted output module are the same as those of steps S1, S2, S3, and S4 in the various embodiments of the above-mentioned method for early warning of clumping in the solid beverage granulation process, and will not be described in detail here.

[0049] While various embodiments of the invention have been shown and described in this specification, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention.

Claims

1. A method for caking early warning in a solid beverage granulation process, characterized in that, The method comprises the following steps: obtaining real-time process parameters of a solid beverage granulation process, the real-time process parameters comprising material temperature, ambient humidity, stirring motor torque and binder spraying rate, and calculating fluctuation entropy based on time series data of the stirring motor torque; determining a reference noise distribution according to real-time variance of the binder spraying rate; calculating a noise adjustment coefficient negatively correlated with the fluctuation entropy value using the fluctuation entropy; perturbing the real-time process parameters and the fluctuation entropy in combination with the reference noise distribution and the noise adjustment coefficient to generate an enhanced feature set; constructing a pre-warning model composed of multiple decision trees, and determining an information gain threshold according to the noise adjustment coefficient when constructing each decision tree, and performing node splitting only when the information gain of a candidate split feature is greater than the information gain threshold, so as to control the depth of the decision tree; for each constructed decision tree, perturbing classification boundary samples in historical data in combination with the reference noise distribution and the noise adjustment coefficient to generate a boundary verification set, and calculating output weights based on a classification confusion matrix of the decision tree on the boundary verification set; inputting the enhanced feature set into the pre-warning model, and weighting and combining classification results of each decision tree according to the output weights of the decision trees to obtain a lumping pre-warning signal.

2. The solid beverage granulation process caking early warning method of claim 1, wherein, The calculation of the fluctuation entropy based on the time series data of the stirring motor torque comprises: obtaining stirring motor torque data of the past 60 seconds at a sampling interval of 1 second to form a time series with a length of 60; performing phase space reconstruction on the time series, and setting embedding dimension m and time delay τ; calculating the number of vector pairs in the reconstructed phase space whose Chebyshev distance is less than a preset threshold r, and the preset threshold r is set to 0.2 times the standard deviation of the original time series; calculating the fluctuation entropy value using a sample entropy formula according to the above statistical result.

3. The solid beverage granulation process caking early warning method of claim 1, wherein, The determination of the reference noise distribution according to the real-time variance of the binder spraying rate comprises: Setting the reference noise distribution to be a Gaussian distribution with mean 0 where the variance of the Gaussian distribution is equal to the variance of the adhesive spray rate data collected over the past 120 seconds. where the variance of the Gaussian distribution is equal to the variance of the adhesive spray rate data collected over the past 120 seconds.

4. The solid beverage granulation process caking early warning method of claim 2, wherein, The noise adjustment coefficient is calculated by using an exponential decay function, and the noise adjustment coefficient is defined as , and the calculation formula is ; wherein, is a current fluctuation entropy value, is a preset decay coefficient.

5. The solid beverage prilling process caking early warning method of claim 4, wherein, The calculation method of the information gain threshold is: Information gain threshold The calculation formula is ; wherein, is the noise adjustment coefficient, is the reference information gain.

6. The solid beverage granulation process caking early warning method of claim 1, wherein, The acquisition method of the boundary verification set is: training historical data using a support vector machine, and identifying data points corresponding to all support vectors as classification boundary samples; for each classification boundary sample, obtaining stirring motor torque time series data at the corresponding time and calculating fluctuation entropy, and calculating a corresponding noise adjustment coefficient according to the fluctuation entropy; for the feature vector of each classification boundary sample, generating a noise vector by randomly sampling from the reference noise distribution and multiplying the noise vector by the noise adjustment coefficient corresponding to the feature vector; adding the generated noise vector to the feature vector of the original classification boundary sample to generate a perturbed sample, and all perturbed sample sets constitute the boundary verification set.

7. The method of claim 1, wherein the method further comprises: The calculation of the output weights based on the classification confusion matrix of the decision tree on the boundary verification set comprises: counting the number of true positives of the classification result of the decision tree on the boundary verification set , the number of false positives , the number of true negatives , and the number of false negatives ; According to the formula Calculate the precision ; According to the formula The recall is calculated ; According to the formula Calculated Fractions, of which Set a preset, extremely small positive number to prevent the denominator from being zero; calculate the... The scores serve as the output weights of the decision tree.

8. A solid beverage granulation process caking early warning system characterized by, The method comprises the following modules: a calculation module, configured to obtain real-time process parameters of a solid beverage granulation process, the real-time process parameters comprising material temperature, ambient humidity, stirring motor torque and binder spraying rate, and calculate fluctuation entropy based on time series data of the stirring motor torque; The generating module is configured to determine a reference noise distribution according to a real-time variance of the adhesive spraying rate; The noise adjustment coefficient negatively correlated with the fluctuation entropy value is calculated by using the fluctuation entropy; The reference noise distribution and the noise adjustment coefficient are combined to perturb the real-time process parameters and the fluctuation entropy, and an enhanced feature set is generated; The control module is configured to construct an early warning model composed of multiple decision trees, and to determine an information gain threshold according to the noise adjustment coefficient when constructing each decision tree, so that node splitting is performed only when the information gain of a candidate splitting feature is greater than the information gain threshold, thereby controlling the depth of the decision tree; The weighting output module is configured to, for each constructed decision tree, combine the reference noise distribution and the noise adjustment coefficient to perturb the classification boundary samples in the historical data to generate a boundary verification set, and to calculate an output weight based on a classification confusion matrix of the decision tree on the boundary verification set; The enhanced feature set is input into the early warning model, and the classification results of each decision tree are weighted and combined according to the output weight of each decision tree, so as to obtain a lumping early warning signal.

9. The solid beverage prilling process cake-out warning system of claim 8, wherein, The fluctuation entropy is calculated based on the time series data of the stirring motor torque, including: The stirring motor torque data of the past 60 seconds is obtained at a sampling interval of 1 second to form a time series with a length of 60; The time series is reconstructed in a phase space, and the embedding dimension m and the time delay τ are set; The number of vector pairs in the reconstructed phase space whose Chebyshev distance is less than a preset threshold r is calculated, and the preset threshold r is set to 0.2 times the standard deviation of the original time series; The fluctuation entropy value is calculated according to the statistical result by using a sample entropy formula.

10. The solid beverage prilling process caking early warning system of claim 8, wherein, The reference noise distribution is determined according to the real-time variance of the adhesive spraying rate, including: The reference noise distribution is set as a Gaussian distribution with a mean of 0 where the variance of the Gaussian distribution is equal to the variance of the adhesive spray rate data collected over the past 120 seconds.

Images