Grain circulating drying process abnormity monitoring method and monitoring system

Abstract

The invention discloses a monitoring method and a monitoring system for abnormity in a grain circulating drying process, and belongs to the field of grain engineering. The method comprises the following steps: acquiring time sequence data acquired by a plurality of sensors in real time; performing feature extraction on the data obtained in the step 3 by using a CNN model; projecting feature vectors extracted by the CNN model to a high-dimensional feature space by adopting an SVM model, and performing modeling judgment by adopting an SVM model based on an RBF kernel; and classifying the feature vectors extracted by the CNN model by using the trained SVM model to obtain a fault classification result. According to the method, abnormity judgment and recognition in the grain postpartum circulating drying process are completed based on time sequence feature fusion, a typical process monitoring effect is achieved, the prediction result is accurate, online error reporting can be achieved by combining abnormal conditions such as actual grain drying temperature instability, humidity abnormity and equipment current and voltage faults, and the method is suitable for popularization and application. And an effective troubleshooting operation is provided for circulating drying of grains.

Description

Technical Field

[0001] The invention relates to a method and a system for monitoring abnormalities in a grain circulation drying process, and belongs to the field of grain engineering. Background Art

[0002] my country is the world's largest grain producer and consumer, with an annual output exceeding 600 million tons. According to statistics, post-harvest grain losses due to mold and mildew caused by untimely post-harvest drying reach 21 million tons annually, accounting for approximately 4.2% of China's total grain output. Therefore, the development of grain drying is urgent. Promoting drying technology and equipment and increasing the mechanization of the post-harvest grain drying process are essential for efficient grain drying and safe storage. Compared to developed countries like Japan, the United States, and Western Europe, where mechanized grain drying rates exceed 90%, my country's current level of grain drying mechanization lags significantly, particularly in the areas of refined post-harvest grain drying and system control, where intelligent technology needs to be improved. Ensuring high-quality development of post-harvest grain drying technology is crucial to my country's grain industry.

[0003] During the grain drying process, regardless of the type of drying method used, the grain's self-care, the initial conditions of the drying medium, external environmental factors, and the internal structure of the drying medium are all key factors in achieving the grain drying goal. Due to the limitations of the internal moisture diffusion rate of grain grains, a single large-scale precipitation can easily cause the grains to crack and break. Therefore, in actual application, a multi-batch drying-slow-refreshing cycle process is set up to achieve post-harvest grain drying to preserve quality. As one of the most typical and widely used post-harvest grain drying methods, the circulating drying process can be flexibly adjusted according to the characteristics of the material, the magnitude of a single precipitation, and the requirements for post-drying quality. It is one of the most representative drying processes currently commercialized. During the entire grain circulation drying process, the grain needs to be continuously dried and precipitated in batches, and it is necessary to ensure that the equipment is cyclically operated during the grain drying process to ultimately achieve a grain moisture content that meets actual requirements.

[0004] A complete cyclic drying process usually consists of key core units such as a quality control system, a hot air system, a drying medium flow channel, a conveying system, an elevator, a grain discharge system, a conveying system, and a grain tower. The drying process requires repetitive programming of grain heating, drying, heat preservation, slow thawing, and cooling. The sensors involved include temperature, humidity, wind speed, moisture, and various current and voltage sensors. In actual operation, due to long-term operation of the equipment or improper process parameter settings, it is very easy to encounter faults such as abnormal drying medium temperature, abnormal humidity, abnormal airflow speed, abnormal current and voltage. Once an abnormal state occurs, it often requires a lot of manpower and material resources to manually check one by one. This process is time-consuming and labor-intensive, and often leads to the termination of the cyclic drying process, resulting in a large amount of grain waste and loss. Summary of the Invention

[0005] In order to achieve intelligent monitoring and abnormality identification during the post-harvest grain drying process, while improving the safety and effectiveness of grain drying monitoring and reducing the manpower and material costs associated with troubleshooting, the present invention provides a method and system for monitoring abnormalities during grain drying cycles. The technical solution is as follows:

[0006] A method for monitoring abnormalities in a grain drying process according to the present invention comprises:

[0007] Step 1: Obtain time series data collected by multiple sensors in real time;

[0008] Step 2: Preprocess the collected data;

[0009] Step 3: Normalize the preprocessed time series data;

[0010] Step 4: Use the CNN model to extract features from the data obtained in step 3, including using a three-layer convolutional neural network to obtain feature context information of time periods at different scales. The convolution process is performed according to the following formula:

[0011]

[0012] Where * represents the convolution operation, t represents the time series, i represents the convolution kernel and bias on the i-th layer, Represents the result of the i-th layer convolution. The three-layer convolution network includes: a convolution kernel of size 3*time series feature dimension T, a convolution kernel of size 3*32, and a convolution kernel of size 3*16. is the result obtained by ReLU activation function;

[0013] Step 5: training the CNN model;

[0014] Step 6: Use the SVM model to project the feature vectors extracted by the CNN model into a high-dimensional feature space, use the RBF kernel-based SVM model for modeling judgment, and use cross-validation and grid search to tune the RBF kernel parameters gamma and C;

[0015] Step 7: Train the SVM model, and use the trained SVM model to classify the feature vectors extracted by the CNN model to obtain a fault classification result.

[0016] Optionally, the multiple sensors include: a temperature sensor, an air humidity sensor, an air volume sensor, a wind speed sensor, a moisture content sensor, a current intensity sensor, and a voltage intensity sensor.

[0017] Optionally, the preprocessing process of step 2 includes:

[0018] Step 21: The collected data exceeds the critical value of the sensor or is within the critical value for 5 consecutive times, which is considered as erroneous data. The data collected by the sensor for 5 consecutive times that are exactly the same are considered as duplicate data, and the data are filled using the mean-difference method.

[0019] Step 22: If more than 20% of the data in a sensor time series dataset needs to be corrected, it is considered an invalid dataset and is directly deleted during the statistical calculation process. If less than 20% of the data needs to be corrected, data is collected from four adjacent points above and below, and the correction is performed using the mean of the difference.

[0020] Optionally, the three-layer convolutional network includes: 32 convolution kernels of size 3*the feature dimension T of the time series, 16 convolution kernels of size 3*32 and 8 convolution kernels of size 3*16.

[0021] Optionally, the method also includes an abnormal alarm. When an abnormal point occurs twice consecutively in a single operating cycle of the sensor, an alarm is triggered. The alarm content includes abnormal inlet / outlet temperature, abnormal grain moisture, abnormal drying wind speed, abnormal humidity, abnormal ventilation air volume, and abnormal voltage / current.

[0022] Optionally, step 5 adopts a cross entropy loss function to train the CNN model and uses an Adam optimizer.

[0023] Optionally, the initial learning rate during training is set to 0.001.

[0024] The present invention provides a grain circulation drying process abnormality monitoring system, the system comprising:

[0025] A sensor module including a plurality of sensors for monitoring environmental data of a grain cycle drying process;

[0026] A data acquisition module is configured to acquire time series data collected in real time by the multiple sensors;

[0027] A data preprocessing module is configured to preprocess the collected data;

[0028] The feature extraction module is configured to use the CNN model to extract features from the data, including using a three-layer convolutional neural network to obtain feature context information of time periods at different scales. The convolution process is performed according to the following formula:

[0029]

[0030] Where * represents the convolution operation, t represents the time series, i represents the convolution kernel and bias on the i-th layer, Represents the result of the i-th layer convolution. The three-layer convolution network includes: a convolution kernel of size 3*time series feature dimension T, a convolution kernel of size 3*32, and a convolution kernel of size 3*16. is the result obtained by ReLU activation function;

[0031] The fault classification module is configured to use the SVM model to classify the feature vectors extracted by the CNN model to obtain a fault classification result.

[0032] The present invention provides a device for monitoring abnormalities in a grain circulation drying process, comprising a memory and a processor;

[0033] The memory is used to store computer programs;

[0034] The processor is used to implement the method for monitoring abnormalities in a grain circulation drying process as described in any one of the above items when executing the computer program.

[0035] The present invention provides a computer-readable storage medium, characterized in that a computer program is stored on the storage medium, and when the computer program is executed by a processor, the abnormality monitoring method of the grain circulation drying process as described in any one of the above items is implemented.

[0036] The beneficial effects of the present invention are:

[0037] (1) The abnormal monitoring system for the grain drying process based on the deep learning algorithm constructed by the present invention can greatly improve the stability monitoring of equipment and process operation during the operation cycle, replacing the original manual judgment operation of the grain drying process, and has the characteristics of rapid response to production safety and fault repair response.

[0038] (2) The feature time series basic data preprocessing developed by the present invention can fill in the data gaps caused by non-human accidental abnormalities such as sensor failure during the grain drying process, thereby improving the online monitoring capability of the grain circulation drying process and providing an effective data preprocessing method for the subsequent machine algorithm fault troubleshooting and identification.

[0039] (3) The grain circulation drying process abnormality monitoring system developed by the present invention based on time series feature fusion can complete the abnormal judgment and identification of grain post-harvest circulation drying process, has a typical process monitoring effect, and has accurate prediction results. It can report online abnormal conditions such as unstable grain drying temperature, abnormal humidity, equipment current and voltage failure, etc., and provide an effective troubleshooting operation for grain circulation drying.

[0040] (4) The present invention uses convolutional neural networks to complete feature screening of time series data sets, obtains contextual features of multiple types of sensors at different time scales, and combines supervised training of support vector machines (SVM) to perform projection inversion on the extracted feature information, completing the judgment and prediction of abnormalities in the grain circulation drying process. It proposes to combine abnormal threshold control to achieve accurate prediction and error reporting of abnormal situations, ensuring accurate and efficient operation in the grain circulation drying process.

[0041] The abnormality monitoring method of the present invention can realize intelligent monitoring and abnormality identification of the post-harvest grain drying process, with accurate fault prediction, high fault classification accuracy, and strong environmental adaptability of the hybrid model. It greatly improves the effective safe operation time of the grain drying process, shortens the manpower and material costs consumed by fault troubleshooting, and plays an important role in promoting the high-quality development of the grain industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0043] Figure 1 This is a flow chart of the fault diagnosis method for the cyclic grain drying process of the present invention.

[0044] Figure 2 This is a schematic diagram of online monitoring of the grain circulation drying process constructed by the present invention.

[0045] Figure 3 This is a graph showing the statistical values ​​of sensor data of different types and the distribution of normalized values ​​in the first embodiment of the present invention.

[0046] Figure 4 This is a feature space distribution diagram under different abnormal conditions in Example 3 of the present invention. DETAILED DESCRIPTION

[0047] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0048] Example 1:

[0049] This embodiment provides a method for monitoring abnormalities in a grain drying cycle. Figure 1 , the method comprising:

[0050] Step 1. Raw data collection: During the grain drying process, rapid detection sensors for temperature, air humidity, air volume, wind speed, moisture content, current intensity, and voltage intensity are set up to ensure that the sensors and other equipment can input real-time data into the detection system and ensure data accuracy and timeliness. During each single operation cycle of each sensor, the data are sorted by time to form a single sensor time series data set x1, x2 to xn, forming the data set X = [x1 x2 x3…xn].

[0051] Step 2: Raw Data Cleaning: All raw signals collected by sensors are cleaned. During this process, data collected that exceeds the sensor's critical value or that is within the critical value for up to five consecutive points is considered erroneous data and corrected. Data collected from a sensor that is identical for up to five consecutive times is considered duplicate data and deleted, and missing data is corrected by comparison. If a sensor time series dataset requires correction for more than 20% of its data, it is considered invalid and directly deleted during statistical calculations. If less than 20% of the data requires correction, the data is corrected by filling in the missing data using the mean of the difference between the two. Once the missing data is filled, the complete time series dataset is considered clean.

[0052] This example uses rapid detection sensors for temperature, humidity, air volume, wind speed, moisture content, current intensity, and voltage intensity during the grain drying cycle. Data sets were collected for 30 minutes before and after a single fault, with 3-minute intervals between each run. Detailed data is shown in Table 1. In this data set, temperatures > 80.0°C and humidity > 90% were defined as exceeding critical values. Temperature data collected during the 9th and 21st minutes of operation, at 100°C and 200°C, respectively, were identified as erroneous data. The air volume and moisture content data collected during the 21st minute run were also identified as erroneous data and required cleaning based on the preceding and following run data.

[0053] Table 1. Data collected from some sensors during the fault period

[0054]

[0055] According to the cleaning principle, if the data requiring correction is less than 20%, data can be collected from four adjacent points above and below, and corrections can be made using the mean of the differences. A total of eight data points, representing 11.43%, required correction. After correction, there were no errors in the data, making it a valid dataset for analysis. The collected data after cleaning is shown in Table 2.

[0056] Table 2 Datasets of some sensors after cleaning during the fault period

[0057]

[0058] Step 3: Convert and process the raw data. Combined with the empty bin operation, remove noise and outliers to ensure data quality. To eliminate the different sampling rates corresponding to different sensor types, data normalization is used to convert the time series data. The specific processing flow includes the following:

[0059] Because different sensors have different statistical dimensions, the data generated during the execution process needs to be dimensionless. In this embodiment, some sensor data takes both direction and unit into account. For example, the closer the temperature (°C) value is to the air inlet temperature, the better, not too high or too low. The closer the grain moisture content (%) is to the target drying moisture value, the better, not too high or too low. Indicators such as time, current, and voltage primarily have different dimensions and units, requiring dimensionless processing to make them comparable and facilitate subsequent descriptions of the stability of the entire drying cycle.

[0060] In this embodiment, normalization is used to standardize the units and dimensions of the data, so that all data are compressed within the range of [0, 1], including the two boundary numbers 0 and 1. The corresponding calculation formula is:

[0061]

[0062] in, is the time series value after data normalization, Xmax and Xmin are the maximum vector value and the minimum vector value in the time series data respectively. In this embodiment, the original statistical data of the grain cycle dryer after 150 minutes of smooth operation is as follows: Figure 3 As shown in (A), the detection values ​​of different sensor types not only have differences in dimension, but also in value, which is not conducive to the direct judgment and monitoring of abnormal points. In this regard, after normalizing the data, the scatter distribution of normal operation data is Figure 3 As shown in (B), all scattered runs are distributed in this interval, eliminating the human identification errors caused by dimension and baseline drift for subsequent modeling calculations.

[0063] Step 4: Data feature extraction: CNN convolutional neural network model is used to extract features. Considering that each time step in each operation cycle contains multi-dimensional features and there is a certain connection between different features, a three-layer convolutional neural network is used to obtain the feature context information of time periods at different scales. The convolution process is explained as follows:

[0064]

[0065] Where * represents the convolution operation, t represents the time series, and i represents the convolution kernel and bias on the i-th layer. Represents the result of the i-th layer convolution. The three-layer convolution network is set to 32 convolution kernels of size 3*time series feature dimension T, 16 convolution kernels of size 3*32, and 8 convolution kernels of size 3*16. is the result obtained by the ReLU activation function.

[0066] Step 5, CNN model training: The model training adopts the cross entropy loss function and uses the Adam optimizer. The initial learning rate is set to 0.001 and the batch size is set to 32. The number of training rounds is selected according to the dataset size and model performance of different abnormal features, and the range is controlled within 50-100 rounds.

[0067] Step 6. Construction of anomaly determination model: Use the SVM algorithm to project the feature vectors extracted by the CNN model into a high-dimensional feature space. Use the RBF kernel-based SVM model for modeling and determination. The initial regularized optimization C value is 0.9, and Tol is 1e-3. Use cross-validation and grid search to tune the RBF kernel parameters gamma and C.

[0068] Step 7: Modeling training and prediction: Using a continuous grain drying production line as the experimental analysis object, training and detection are performed on abnormal conditions such as grain drying temperature, air humidity, air volume, wind speed, moisture content, current intensity, and voltage intensity. The dataset is divided into training and validation sets with a ratio of 4:1. The training model requires a dataset of at least 80 times to ensure that the constructed model is representative.

[0069] Step 8. Abnormal threshold control: Based on the principle of setting the threshold for one continuous cycle, when the monitoring sensor has abnormal points twice in a single operating cycle, an alarm is triggered and the emergency unloading port control system of the dryer is adjusted to the open state to facilitate fault inspection and troubleshooting.

[0070] Step 9. External independent validation: External independent validation is performed using a dataset of abnormal occurrences in the drying process of multiple batches of grain cycles, and the constructed model is independently evaluated based on the accuracy of the predicted type.

[0071] Example 2:

[0072] In order to further demonstrate the effectiveness of the abnormality monitoring method for the grain drying process based on the CNN-SVM hybrid model of the present invention, this embodiment is compared with the existing prediction model.

[0073] This example uses data collected during the experimental operation of a grain recycling dryer from May 20, 2024, to July 27, 2024, as the raw data set. This data was collected 750 times, totaling 7,500 valid records. The normalized data was divided into a training set and a prediction set at a 4:1 ratio. The model training consisted of 600 runs, totaling 6,000 valid records, including data on both normal operation and failures of the grain recycling dryer. The test set consisted of 150 runs, totaling 1,500 valid records.

[0074] This embodiment uses a CNN network model for feature extraction. After training, a three-layer convolutional neural network is used to obtain feature context information of time periods at different scales. The three-layer convolutional network is set to 32 convolution kernels of size 3*the feature dimension T of the time series, 16 convolution kernels of size 3*32, and 8 convolution kernels of size 3*16. The initial learning rate is set to 0.001, the batch size is set to 32, and it is defined that the training data is passed to the model for 75 rounds of training.

[0075] For terminal fault identification and judgment, the SVM algorithm is used to project the feature vectors extracted by the CNN model into a high-dimensional feature space. The RBF kernel-based SVM model is used for modeling and judgment. The initial regularization optimization C value is 0.9, and Tol is 1e-3. Cross-validation and grid search are used to tune the RBF kernel parameters gamma and C.

[0076] The algorithms used in the comparative experiments include the partial least squares regression discriminant model PLS-DA, the linear regression discriminant model LR, the BP neural network BPNN, the support vector machine SVM, and the convolutional neural network model CNN. The same original data set and data cleaning method were used, and the prediction set and training set were constructed simultaneously in a 4:1 ratio. Finally, the overall prediction accuracy and average test time were used as the judgment basis. The abnormality prediction judgment results of different models are compared as shown in Table 3:

[0077] Table 3 Analysis of prediction effects of five different modeling models

[0078] Identification Model Normal group Abnormal group Average accuracy Average testing time PLS-DA 424 / 575 134 / 175 74.4％ 1.5min LR 429 / 575 136 / 175 75.3％ 1.8min BPNN 485 / 575 154 / 175 85.2％ 31min Support Vector Machine 480 / 575 151 / 175 84.1％ 24min CNN-SVM 573 / 575 172 / 175 99.3％ 1.1min

[0079] In terms of prediction effect, the monitoring method based on the CNN-SVM hybrid model of the present invention has the shortest average modeling test time of 1.1 minutes, indicating that during the multi-data set training process of this embodiment, the modeling efficiency of the CNN-SVM model of the present invention is relatively the best; the monitoring method based on the CNN-SVM hybrid model of the present invention has a higher accuracy than PLS-DA, LR, BPNN and SVM models, with an accuracy rate of 99.3%, which performs best in anomaly detection.

[0080] Example 3:

[0081] This example further conducted external independent verification experiments on six different abnormal conditions that occurred during the grain drying process: operating temperature, grain moisture, drying rate, drying humidity, air volume and speed, and voltage and current. A total of 149 operating data were obtained from the independent experimental samples. After data cleaning and dimension removal and normalization, the CNN-SVM hybrid model constructed in Example 1 was used to further identify and determine the abnormal type. The identification results are shown in Table 4 and Figure 4 shown.

[0082] Table 4 Analysis of recognition effects of six different abnormal situations

[0083] Identification Type Accuracy Abnormal operating temperature 29 / 29(100％) Abnormal grain moisture 22 / 24(91.67％) Abnormal drying rate 23 / 24(95.8％) Abnormal dryness and humidity 23 / 24(95.8％) Abnormal air volume and speed 24 / 24(100％) Abnormal current and voltage 24 / 24(100％)

[0084] As can be seen from the data in Table 4, this embodiment achieves 100% accurate discrimination of temperature anomalies, air volume and speed anomalies, and current and voltage during operation. The corresponding moisture anomalies in the drying process can be identified with a 91.67% accuracy rate, and drying rate and drying humidity anomalies can be accurately identified with 95.8%. The results confirm that the recognition model constructed in this embodiment has high anomaly recognition accuracy.

[0085] Some steps in the embodiments of the present invention may be implemented using software, and the corresponding software program may be stored in a readable storage medium, such as a CD or a hard disk.

[0086] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for monitoring abnormality in a grain drying process, characterized in that: The method comprises: Step 1: Obtain time series data collected by multiple sensors in real time; Step 2: Preprocess the collected data; Step 3: Normalize the preprocessed time series data; Step 4: Use the CNN model to extract features from the data obtained in step 3, including using a three-layer convolutional neural network to obtain feature context information of time periods at different scales. The convolution process is performed according to the following formula: Where * represents the convolution operation, t represents the time series, i represents the convolution kernel and bias on the i-th layer, Represents the result of the i-th convolution. The three-layer convolution network includes: a convolution kernel of size 3*time series feature dimension T, a convolution kernel of size 3*32, and a convolution kernel of size 3*16. is the result obtained by ReLU activation function; Step 5: training the CNN model; Step 6: Use the SVM model to project the feature vector extracted by the CNN model into a high-dimensional feature space, use the RBF kernel-based SVM model for modeling determination, and use cross-validation and grid search to tune the RBF kernel parameters gamma and C; Step 7: Train the SVM model, and use the trained SVM model to classify the feature vectors extracted by the CNN model to obtain a fault classification result.

2. The method for monitoring abnormality in grain drying process according to claim 1, characterized in that: The multiple sensors include: a temperature sensor, an air humidity sensor, an air volume sensor, a wind speed sensor, a moisture content sensor, a current intensity sensor, and a voltage intensity sensor.

3. The method for monitoring abnormality in grain drying process according to claim 1, characterized in that: The pretreatment process of step 2 includes: Step 21: The collected data exceeding the critical value of the sensor or the collected data within the critical value for 5 times are regarded as erroneous data; the collected data of the sensor that are exactly the same for 5 times are regarded as duplicate data, and the supplementary mean method is used to fill them; Step 22: If the data in a sensor time series data set that needs to be corrected exceeds 20%, it is considered an invalid data set; during the statistical calculation process, it is directly deleted; if the data that needs to be corrected is less than 20%, data is collected from 4 adjacent points above and below, and the correction is filled by using the mean of the difference.

4. The method for monitoring abnormality in grain drying process according to claim 1, characterized in that: The three-layer convolutional network includes: 32 convolution kernels with a size of 3*the feature dimension T of the time series, 16 convolution kernels with a size of 3*32 and 8 convolution kernels with a size of 3*16.

5. The method for monitoring abnormality in grain drying process according to claim 1, characterized in that: The method also includes an abnormal alarm. When an abnormal point occurs twice in a row within a single operating cycle of the sensor, an alarm is triggered. The alarm content includes abnormal inlet / outlet temperature, abnormal grain moisture, abnormal drying wind speed, abnormal humidity, abnormal ventilation air volume, and abnormal voltage / current.

6. The method for monitoring abnormality in grain drying process according to claim 1, characterized in that: The step 5 adopts the cross entropy loss function to train the CNN model and uses the Adam optimizer.

7. The method for monitoring abnormality in grain drying process according to claim 6, characterized in that: The initial learning rate during training was set to 0.

001.

8. A grain cycle drying process abnormality monitoring system, characterized in that: The system comprises: A sensor module including a plurality of sensors for monitoring environmental data of a grain cycle drying process; A data acquisition module is configured to acquire time series data collected by the multiple sensors in real time; A data preprocessing module is configured to preprocess the collected data; The feature extraction module is configured to use the CNN model to extract features from the data, including using a three-layer convolutional neural network to obtain feature context information of time periods at different scales. The convolution process is performed according to the following formula: Where * represents the convolution operation, t represents the time series, i represents the convolution kernel and bias on the i-th layer, Represents the result of the i-th convolution. The three-layer convolution network includes: a convolution kernel of size 3*time series feature dimension T, a convolution kernel of size 3*32, and a convolution kernel of size 3*16. is the result obtained by ReLU activation function; The fault classification module is configured to classify the feature vector extracted by the CNN model using the SVM model to obtain a fault classification result.

9. A device for monitoring abnormality during grain drying cycle, characterized in that: including memory and processor; The memory is used to store computer programs; The processor is used to implement the method for monitoring abnormalities in a grain circulation drying process as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by the processor, the method for monitoring abnormalities in a grain circulation drying process as described in any one of claims 1 to 7 is implemented.

Images