A method and device for multi-index prediction of stored grain

Abstract

The present application relates to a kind of stored grain multi-index prediction method and device, belong to grain storage quality evaluation technical field.The present application is first by uniform distribution statistical analysis method to the data enhancement and generation of stored grain data under constant temperature, constant humidity environment of small sample, solve the stored grain with label sample measurement process redundant, environmental microbiological interference, the sample scarcity problem caused by the factors such as high cost of measuring instrument;Then, using the improved SSA search algorithm optimized LSTM neural network model as prediction model, enhance the global optimization, local optimization ability of traditional SSA algorithm, using grid search to optimize the double hidden layer node number, learning rate etc.of LSTM neural network model, overcome the blindness, uncertainty and priori knowledge dependency of traditional neural network model parameter selection.

Description

Technical Field

[0001] This invention relates to a method and apparatus for predicting multiple indicators of stored grain, belonging to the technical field of grain storage quality evaluation. Background Technology

[0002] Conventional wheat quality is often evaluated based on optical, acoustic, and volatile matter characteristics. Optical methods utilize the absorption and reflection of light by stored grains to obtain spectral images for quality analysis. Acoustic methods rely on the differences in reflection and transmission caused by the grain's structure and internal state under the influence of sound waves to obtain information about the grain's internal structure. Volatile matter methods detect substances produced as gases during the transformation of organic matter in the grain to obtain information about the quality of stored grains. However, these evaluation methods are limited by factors such as the availability of testing equipment, environmental noise interference, and significant instability. Furthermore, conventional testing is mostly offline, which means that the quality evaluation results cannot reflect temporal patterns and cannot accurately predict the quality of stored grains at different storage stages, or may ignore the state of stored grains before the tested stage.

[0003] To explore more scientific, effective, simple, and accurate methods for predicting the quality of stored wheat, scholars have conducted extensive research on the physiological and biochemical indicators of wheat in recent years. They have found that the quality of stored wheat changes with storage time, which is reflected in changes in some of its physiological and biochemical indicators. The rapid development of artificial intelligence algorithms has effectively improved the efficiency and accuracy of time-series data mining and predictive analysis. For example, some researchers have proposed using terahertz time-domain spectrometers to obtain spectral data of wheat samples in specific bands, and using principal component analysis (PCA) and other analytical methods to construct discriminant models based on methods such as SVM. This allows for the classification of wheat grain mold levels into four categories: normal, mild, moderate, and severe mold, providing a detection method for grain storage. However, the spectrometers used to obtain the spectral characteristics of wheat are relatively expensive, making widespread adoption difficult. Another approach proposed incorporating backpropagation (BP) neural networks into a quality early warning model for stored wheat. This model predicts 11 quality parameters, including starch and protein values, based on temperature and moisture content. While providing an effective method for predicting the quality of stored wheat, it requires sampling data from three layers of wheat at the same location over nearly three years. This sampling process is costly, and the difficulty of obtaining the labeled data negates the predictive value. Furthermore, this prediction process ignores the storage time of wheat and the fact that wheat exhibits different physiological responses at different storage stages. To address this limitation in solving time-series problems, some researchers have focused on the correlation between multiple physiological and biochemical indicators in wheat. They constructed a support vector machine hybrid kernel function and optimized the kernel function parameters using an improved gray wolf algorithm to establish a quality prediction model for short-term stored wheat. However, because physiological indicators do not change significantly in the short term during storage, and wheat storage periods are generally long, this model cannot reflect the deterioration trend of quality parameters during long-term storage.

[0004] While the above methods help improve the accuracy of wheat storage data prediction results, in the prediction of multi-indicator wheat data, the high cost of obtaining wheat storage quality parameters and the scarcity of tagged wheat storage samples limit the data-driven quality prediction model, greatly reducing the model's generalization ability. Furthermore, the depth and temporal sequence of multi-indicator parameters of stored wheat, the relationship between indicators and storage environment, and the physicochemical coupling between indicators all have a significant impact on the accuracy of quality prediction results.

[0005] Because LSTM memory units effectively capture both short-term and long-term information in time series, LSTM models can accurately predict and analyze the changing trends of wheat indicators. The accuracy of LSTM predictions is closely related to the settings of its weights and thresholds. Bioinspired algorithms are mainly used to solve NP-hard problems, where NP refers to a nondeterministic polynomial, meaning that a feasible solution to the combinatorial optimization problem to be solved is given within an acceptable time and space. Therefore, bioinspired algorithms are an effective method for LSTM parameter optimization. The Sparrow Search Algorithm (SSA), based on the foraging and anti-predation behavior of sparrow populations, solves the problem of random changes in model input weights and thresholds, and has high-performance global search capabilities, good stability, and good convergence accuracy. The SSA algorithm has a fast convergence speed and powerful search capabilities. It uses grid search to optimize the initial weights and thresholds of the LSTM neural network model, overcoming the blindness and uncertainty of parameter selection in traditional neural network models. Currently, SSA is widely used in nonlinear time series data processing, such as wind speed forecasting and rainfall analysis and forecasting. However, SSA's local search capability is insufficient, resulting in poor local convergence of the model. Summary of the Invention

[0006] The purpose of this invention is to provide a method and apparatus for predicting multiple indicators of stored grain, so as to solve the problems of poor accuracy and poor local convergence of the model in the current process of predicting multiple indicators of stored grain.

[0007] To solve the above-mentioned technical problems, this invention provides a multi-indicator prediction method for grain storage, which includes the following steps:

[0008] 1) Collect the physicochemical characteristics of stored grain at different storage stages and under different environmental conditions to form stored grain label samples;

[0009] 2) Using LSTM as the prediction model, the improved SSA algorithm is used to optimize the initial weights and thresholds in the prediction model. The improved SSA algorithm refers to the perturbation of the sparrow's position by the t-distribution mutation operator with the iteration number iter as the degree of freedom parameter.

[0010] 3) Input the physicochemical characteristics of the stored grain from the obtained stored grain label samples into the optimized prediction model to achieve prediction of multiple indicators of stored grain.

[0011] This invention employs an LSTM (Long Short-Term Memory) neural network as a prediction model for multiple grain storage indicators. An improved SSA (Search Algorithm for LSTM) search algorithm is used to optimize the LSTM neural network prediction model. This method introduces a t-distribution perturbation operator with the iteration number `iter` as the degree of freedom into the sparrow position update rule of the SSA optimization algorithm. This enhances the algorithm's global and local search capabilities, improves the model optimization speed, overcomes the blindness, uncertainty, and reliance on prior knowledge in traditional neural network model parameter selection, and improves search accuracy, convergence speed, and prediction stability.

[0012] Furthermore, the method also includes expanding the grain label sample based on the characteristic that the physicochemical characteristics of stored grain maintain a stable increase or decrease in the short term under the same environmental conditions.

[0013] This invention discovers that under constant temperature and humidity conditions, the respiration rate of stored grains remains stable over a short period, and the physiological reactions within the grains also proceed at a constant rate, resulting in linear and stable changes in the physicochemical characteristics of the crop. Based on this, this invention expands the sample size, solving the problem of sample scarcity caused by factors such as the cumbersome measurement process of labeled stored grain samples, interference from environmental microorganisms, and the high cost of measuring instruments.

[0014] Furthermore, the physicochemical characteristics of each stored grain in the stored grain label sample are the characteristics after screening. The screening process is as follows: analyze the correlation between each physicochemical characteristic of the stored grain and the storage time, and select the physicochemical characteristics of the stored grain with higher correlation as the input features of the prediction model.

[0015] This invention screens the physicochemical characteristics of stored grain based on their correlation with storage time, which to some extent avoids the problem of inaccurate prediction due to large differences in correlation between samples, and further improves the accuracy of prediction.

[0016] Furthermore, the position of the sparrow after disturbance in step 2) is:

[0017]

[0018] in The position of the sparrow after the disturbance. Let this be the position of the sparrow at iteration t. (iter) represents a random disturbance term.

[0019] This invention adds a random interference term. (iter) not only makes full use of the current position information, but also adds random interference information, which helps the algorithm to escape local optima.

[0020] Furthermore, when optimizing the initial weights and thresholds in the prediction model using the improved SSA algorithm, the t-distribution is adjusted according to the number of iterations, so that the t-distribution approximates the Cauchy distribution in the early stage of the algorithm operation when the number of iterations is small; in the later stage of the algorithm operation, the t-distribution approximates the Gaussian distribution; and in the middle stage of the algorithm operation, the t-distribution is between the Cauchy distribution and the Gaussian distribution.

[0021] In the early stages of the algorithm's operation, the number of iterations is relatively small, allowing the t-distribution to approximate the Cauchy distribution, resulting in good global search capabilities. In the later stages of the algorithm's operation, the t-distribution approximates the Gaussian distribution, providing good local search capabilities. In the middle stages of the algorithm's operation, the t-distribution lies between the Cauchy and Gaussian distributions, and the mutation operator of the t-distribution combines the advantages of both Gaussian and Cauchy operators, thereby improving both the global search capability and local exploitation capability of the algorithm.

[0022] Furthermore, the physicochemical characteristics of the stored grain collected in step 1) include electrical conductivity, peak viscosity, chewiness, light transmittance, water absorption rate, and fatty acid value.

[0023] Furthermore, during the expansion, the distribution of the physical and chemical characteristics of the stored grain is made to conform to a continuous uniform distribution in the statistical distribution.

[0024] The present invention also provides a multi-indicator prediction device for grain storage, which includes a processor that executes a program stored in a memory to implement the multi-indicator prediction method for grain storage described in the present invention.

[0025] This invention employs an LSTM (Long Short-Term Memory) neural network as a prediction model for multiple grain storage indicators. An improved SSA (Search Algorithm for LSTM) search algorithm is used to optimize the LSTM neural network prediction model. This method introduces a t-distribution perturbation operator with the iteration number `iter` as the degree of freedom into the sparrow position update rule of the SSA optimization algorithm. This enhances the algorithm's global and local search capabilities, improves the model optimization speed, overcomes the blindness, uncertainty, and reliance on prior knowledge in traditional neural network model parameter selection, and improves search accuracy, convergence speed, and prediction stability. Attached Figure Description

[0026] Figure 1 This is a flowchart of the multi-indicator prediction method for grain storage of the present invention;

[0027] Figure 2 This is a schematic diagram showing the relationship between the fatty acid value of fragrant rice varieties and storage time and temperature.

[0028] Figure 3 This is a schematic diagram of the structure of the LSTM unit in the LSTM neural network model used in the multi-index prediction method for grain storage of the present invention;

[0029] Figure 4This is a schematic diagram of the standard Gaussian distribution, standard Cauchy distribution, and t-distribution curves;

[0030] Figure 5 This is a schematic diagram illustrating the correlation coefficients between storage time and various product parameters in embodiments of the present invention;

[0031] Figure 6a This is an iterative comparison chart of the RMSE values ​​of fatty acids predicted by the present invention and the SSA-LSTM model during the experimental process;

[0032] Figure 6b This is an iterative comparison chart of the peak viscosity RMSE predicted by the present invention and the SSA-LSTM model during the experimental process;

[0033] Figure 6c This is an iterative comparison chart of the water absorption rate RMSE predicted by the present invention and the SSA-LSTM model during the experimental process;

[0034] Figure 6d This is a comparison chart of the RMSE of chewing degree predicted by the present invention and the SSA-LSTM model during the experimental process;

[0035] Figure 7a This is a comparison chart of the predicted and actual fatty acid values ​​obtained by this invention during the experimental process;

[0036] Figure 7b This is a comparison chart of the predicted peak viscosity value and the actual value obtained by this invention during the experimental process;

[0037] Figure 7c This is a comparison chart of the predicted and actual water absorption rates obtained by this invention during the experimental process;

[0038] Figure 7d This is a comparison chart of the predicted and actual chewiness values ​​obtained by this invention during the experimental process. Detailed Implementation

[0039] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0040] Method Implementation Examples

[0041] This invention first uses a uniform distribution statistical analysis method to observe the short-term fluctuations and long-term trends of physiological and biochemical indicators within stored grains under constant temperature and humidity conditions. It then augments and generates data from small samples of wheat stored under constant temperature and humidity conditions, solving the problem of sample scarcity caused by the cumbersome measurement process for labeled grain samples, interference from environmental microorganisms, and the high cost of measuring instruments. Next, it uses an optimized LSTM neural network model based on an improved SSA search algorithm as the prediction model, enhancing the global and local optimization capabilities of the traditional SSA algorithm. A grid search is employed to optimize the number of hidden layer nodes and the learning rate of the LSTM neural network model, overcoming the blindness, uncertainty, and reliance on prior knowledge in parameter selection of traditional neural network models. Finally, through Spearman correlation coefficient analysis of the physicochemical characteristics of stored grains and storage time, it selects physicochemical characteristics with highly significant correlations to storage time and other factors as predicted quality parameters. The implementation process of this method is as follows: Figure 1 As shown.

[0042] 1. Obtain time-series data of various indicators of stored grain collected under different environments to form initial label samples.

[0043] This invention selects six physiological and biochemical indicators—electrical conductivity, peak viscosity, chewiness, light transmittance, water absorption rate, and fatty acid value—as the physicochemical characteristics (also called physicochemical quality parameters) of stored grain. These physicochemical characteristics are collected at set time intervals during the storage period of the grain to form a time series of the stored grain's physicochemical characteristics. This series serves as the initial label sample. During the data collection, environmental data, primarily temperature and humidity, also needs to be recorded.

[0044] 2. Expand the initial labeled sample size.

[0045] Research revealed that when corn of the same variety was stored using various methods after harvesting in Northeast China, the fatty acid values ​​showed that, when corn of the same variety was harvested and stored at the same time and was in the same storage stage, the current storage stage under different storage conditions showed the same trend as the previous stage. Moreover, at the same temperature in a short period, the fatty acid values ​​did not show a step change.

[0046] Using conventional fragrant rice varieties as test materials, during a 150-day storage period, it was found that the fatty acid values ​​of the two fragrant rice varieties used in the test were significantly positively correlated with storage time and storage temperature, while the aroma was significantly negatively correlated with storage time and storage temperature. Figure 2As shown, the alkali spreading value is significantly negatively correlated with storage time. Within a 30-day storage period, the aroma exhibits a linear relationship with changes in storage time, temperature, and humidity. The respiration rate of wheat under four constant-temperature conditions was detected using an SKW-3 micro-breathing pressure detector. The results showed that the wheat respiration rate exhibited a relatively stable characteristic. Under low constant-temperature conditions, the fluctuation range of the respiration rate was only 0.03–0.05 μL / (100 grains*h). Under high constant-temperature conditions, the respiration rate was slightly higher than that under low constant-temperature conditions, but remained at a stable level, indicating that high temperature contributes to the respiration efficiency of wheat. However, under constant-temperature conditions, the respiration rate generally exhibits a stable characteristic. The magnitude of the respiration rate directly affects most physiological reactions within the wheat grain. Maintaining a stable respiration rate allows most internal parameters of wheat to show a linear upward or downward trend over a short period.

[0047] The above findings indicate that under constant temperature and humidity conditions, the respiration rate of stored grains such as wheat and corn remains stable over short periods, and the physiological reactions within the grains also proceed at a constant rate, resulting in linear and stable changes in the physicochemical parameters of the crops. However, under different temperature and humidity conditions, crops stored for the same period will exhibit different linear angle coefficients for the same indicator. When generating quality-labeled stored grain samples, temperature, humidity, and storage period are all major influencing factors. Therefore, this invention expands the capacity of stored wheat label samples under constant temperature and humidity conditions in a sealed environment.

[0048] When expanding the sample of stored grain labels, the various physicochemical quality parameters of the stored grain are related to the temperature, humidity, and storage stage of the storage environment. The temperature and humidity of the storage environment can be uniformly regarded as environmental variables, because the humidity content is controlled simultaneously with the temperature. The storage stage is defined as the period from when the grain enters the storage facility, through regular measurement of physicochemical quality parameters, as well as the time interval between the current stage and the adjacent measurement period. Therefore, in the expansion of the sample of physicochemical quality parameters of stored grain, a certain physicochemical quality parameter at each stage of multiple environments can be represented as shown in formula (1).

[0049]

[0050] In the formula, x(i,j,t) represents the value of the physicochemical quality parameter of the i-th grain at storage stage t in the j-th storage environment.

[0051] As shown in formula (1), during the storage stage of crops such as wheat and corn, if the temperature and humidity remain stable, the physiological reactions inside the grains will also maintain a stable rate, and the changes in the parameter indicators caused by the physiological reactions will also maintain a stable rate of increase or decrease in the short term. However, due to the uncontrollability of other environmental variables, the increase in the indicator parameters will still have a disturbance amplitude on the existing main variation amplitude. When only the main variation amplitude is considered, the distribution of the indicator parameters approximately follows a continuous uniform distribution in the statistical distribution. The probability density function of the physicochemical quality parameters can be obtained from formula (1) as follows:

[0052]

[0053] From formula (2), the probability uniform distribution function of the physicochemical quality parameters is as follows:

[0054]

[0055] This invention expands the data for each environment. Formula (3) means that such a distribution function can be established for each environment. Here, r is the uncertainty error, which represents the error value caused by the difference in the quality parameters of stored grain during offline inspection, taking into account the fluctuations in the detection process and the sudden impact of the environment. The main value is a small random value that tends to 0.

[0056] 3. Screen the physicochemical characteristics of stored grain in the samples.

[0057] Since this invention involves six physicochemical characteristics of stored grains, directly incorporating them into subsequent model training would inevitably affect training efficiency. Furthermore, considering that the correlations between conductivity, peak viscosity, chewiness, light transmittance, water absorption rate, fatty acid value, and storage time are not uniform, physicochemical characteristics with weak correlations not only increase training complexity but also interfere with prediction results. Therefore, before using the aforementioned samples for model training, this invention first screens the physicochemical characteristics of stored grains in the samples. Specifically, it calculates the correlation coefficient between each physicochemical characteristic and storage time, and removes physicochemical characteristics with low correlation coefficients.

[0058] 4. Build a prediction model.

[0059] LSTM is a variant of traditional recurrent neural networks, suitable for processing and predicting time series with relatively long intervals and time delays. It has the advantage of long memory when processing time series predictions and is widely used in fields requiring in-depth analysis of time series data. Therefore, this invention uses LSTM as a prediction model for multi-indicator prediction of grain reserves. LSTM introduces a gate mechanism to control feature flow and loss on the basis of recurrent neural networks, solving the long-term dependency problem of recurrent neural networks. It adds memory units, forget gates, input gates, and output gates to the neural network, selectively storing information to achieve simultaneous long and short term memory. Its structure is as follows: Figure 3 As shown.

[0060] Each LSTM unit contains a tuple, and its state at time t is C. t It contains long-term memory information from before this moment, and the hidden layer state at time t is h. t It contains short-term memory information of the sequence. The selection of long short-term memory by the LSTM unit is mainly determined by three parts: the forget gate, the input gate, and the output gate. At time t, the input of the LSTM memory unit includes: sequence input x t The state c of the memory unit at time t-1 t–1 and the state h of the hidden layer at time t–1 t–1 The forget gate is responsible for deciding which less important information to discard from the memory unit, and the update formula is as follows:

[0061] f t =σ(w fx x t +w fh h t-1 +w fc c t-1 +b f (4)

[0062] Where σ is the sigmoid activation function, x t Given the input at time t, h t Let h be the hidden layer state at time t. t-1 Let c be the hidden layer state at time t-1. t-1 The state w of memory cell at time t-1 fx w fh w fc b is the weighting coefficient in the forgetting gate. f This is the bias term. The input gate is responsible for determining which information can be stored in the memory unit, and the update formula is as follows:

[0063] i t =σ(w ix x i +w ih ht-1 +w ic c t-1 +b i (5)

[0064]

[0065]

[0066] Where w ix w ih w ic b represents the weighting coefficient of the input gate. i For the bias term of the input gate, w cx w ch b represents the weighting coefficients for memory cells and hidden layer states. o is its bias term, and tanh is the hyperbolic tangent activation function.

[0067] o t =σ(w ox x t +w oh h t-1 +w oc c t-1 +b o (8)

[0068] h t =o t tanh(c t (9)

[0069] o t To output gate information, w ox w oh w oc b represents the weighting coefficient of the output gate. o This is the bias term for the output gate.

[0070] When studying the variation patterns of multiple environmental and quality parameters in wheat, traditional LSTM networks mainly use known time-series information to predict values ​​at a specified time. However, the measurement units and numerical ranges of these wheat index sequences vary greatly, and a single physiological and biochemical index is not representative when analyzing overall quality. Therefore, this paper proposes to use the weight matrix method to integrate the LSTM prediction results of multiple physiological and biochemical index sequences.

[0071] 5. Optimize and solve the prediction model.

[0072] To address the issue of random variations in model input weights and thresholds, this invention employs the ISSA optimization algorithm to optimize the initial weights and thresholds of the LSTM neural network model, overcoming the blindness and uncertainty in parameter selection inherent in traditional LSTM neural network models. This reduces the overall error in multi-indicator prediction for stored wheat while improving the model's convergence and divergence capabilities.

[0073] The ISSA algorithm is an optimization based on SSA (Sparrow Search Algorithm), a novel swarm intelligence optimization algorithm inspired by the foraging and anti-predation behaviors of sparrows. In the Sparrow Search Algorithm, individual sparrows are classified as discoverers, followers, and watchers, with each individual's position corresponding to a solution. Watchers constitute 10%–20% of the population, while the proportions of discoverers and followers are dynamically changing. However, preliminary verification revealed that the SSA algorithm is prone to getting trapped in local optima, making it difficult to effectively balance the algorithm's global and local search capabilities. Furthermore, due to the complexity of the actual search process, it is difficult to accurately describe the actual iterative search process. Therefore, this invention proposes the ISSA algorithm, which introduces a t-distribution perturbation function into the SSA position update strategy, thereby improving the local search capability of the SSA algorithm.

[0074] SSA algorithm principle:

[0075] Assuming there are n sparrow individuals in a sparrow population, the population consisting of all individuals can be represented as:

[0076]

[0077] Where d is the dimension of the variable to be optimized, the fitness function corresponding to the current n sparrow individuals can be expressed as:

[0078]

[0079] f(x) represents the fitness value of an individual. In the division of labor, the discoverer primarily provides the foraging direction and area for the entire population, the follower follows the discoverer in foraging, and the watcher is responsible for monitoring the foraging area. The position update rules and strategies for the three are as follows:

[0080]

[0081] Formula (12) is the update rule for the position of the discoverer in the population, where t represents the current iteration number. This represents the position of the i-th sparrow in the j-th dimension during the t-th generation, where α is a random number ∈ (0, 1), and iter maxis the maximum number of iterations, R2 represents the early warning value, ST represents the safety threshold, Q is a random number following a normal distribution, L is a 1×dim matrix of all ones, and dim represents the dimension. The discoverer position update rule can be summarized as follows: when R2<ST, it means there are no predators around the foraging area, and the discoverers can search for food extensively; when R2≥ST, it means predators appear, and all discoverers need to fly to a safe area. If a predator appears, Equation (14) is executed, otherwise Equation (13) is executed.

[0082]

[0083] Equation (13) is the follower position update rule, represents the position of the individual with the worst fitness in the t-th generation, represents the position of the individual with the best fitness in the (t+1)-th generation. A represents a 1×dim matrix, each element in the matrix is randomly preset to -1 or 1, and A + =A T (AA T ) -1 , the follower position update method can be summarized as: when i>n / 2, it means the i-th follower has low fitness and is not qualified to compete with the discoverers for food, and needs to fly to other areas to forage; when i≤n / 2, the follower will forage near the optimal individual x p .

[0084]

[0085] Equation (14) is the vigilante position update rule, represents the global optimal position in the t-th generation, β is the control step size, which follows a normal distribution with a mean of 0 and a variance of 1, k is a random number ∈[-1,1], and ε is set as a constant to avoid the denominator being 0. f i represents the fitness value of the current individual, f g and f w represent the current fitness values of the global optimal and worst individuals. The vigilante position update method can be summarized as: when f i ≠f g , it means the individual is at the periphery of the population, and needs to take anti-predator behavior, constantly changing positions to obtain higher fitness; when f i =f g , it means the individual is at the center of the population, and it will continuously approach nearby companions to move away from the dangerous area.

[0086] Although the SSA algorithm has a fast convergence speed and strong search capability, and can optimize the initial weights and thresholds of LSTM neural network models, its local search capability is insufficient, resulting in poor local convergence of the model. Therefore, to improve the local search capability, this invention introduces the t-distribution, also known as the Student's distribution. The shape of the t-distribution curve is related to the degree of freedom (v). Compared with the standard normal distribution curve, the smaller the degree of freedom (v), the flatter the t-distribution curve, the lower the middle of the curve, and the higher the tails on both sides; the larger the degree of freedom (v), the closer the t-distribution curve is to the normal distribution curve. When the degree of freedom (v) approaches infinity, the t-distribution curve is a standard normal distribution curve. t(n→∞) approximately approaches the Gaussian distribution N(0,1), and t(n=1) approximately approaches the Cauchy distribution c(0,1). That is, the standard Gaussian distribution and the Cauchy distribution are two boundary special cases of the t-distribution. The function distribution is as follows: Figure 4 As shown.

[0087] This invention employs a t-distribution mutation operator with the iteration number `iter` as the degree of freedom parameter to perturb the sparrow's position. In the early stages of the algorithm, when the number of iterations is small, the t-distribution approximates a Cauchy distribution, giving the algorithm good global search capability. In the later stages, the t-distribution approximates a Gaussian distribution, exhibiting good local exploration capability. In the middle stages, the t-distribution falls between the Cauchy and Gaussian distributions. The t-distribution mutation operator combines the advantages of both Gaussian and Cauchy operators, simultaneously improving the algorithm's global search capability and local exploration capability. The specific position update method for the sparrow after introducing adaptive t-distribution perturbation is as follows:

[0088]

[0089] in The position of the sparrow after the disturbance. Let be the position of the sparrow at iteration t. The proposed update formula is... Based on this, a random disturbance term was added. This approach fully utilizes current location information while incorporating random interference, which helps the algorithm escape local optima. The positions of the discoverer, tracker, and monitor are all updated according to formula (15).

[0090] The expression for t(iter) is shown in formula (16), t n (x) is the probability density expression of the t-distribution function, where f is the t-distribution with n degrees of freedom (maximum number of iterations), and x is the number of iterations (iter). As the number of iterations (iter) increases, the t-distribution gradually approaches a Gaussian distribution, which helps to improve the convergence speed of the algorithm.

[0091]

[0092] Device Examples

[0093] The present invention also provides a multi-indicator prediction device for grain storage. The control device includes a processor and a memory. The processor executes a program stored in the memory to implement the multi-indicator prediction method for grain storage of the present invention. The working process of the device has been described in detail in the above method embodiments and will not be repeated here.

[0094] Experimental verification

[0095] The effectiveness of the present invention will be further verified through simulation experiments below.

[0096] This experiment used Ruixing No. 1, a high-gluten wheat variety from Henan Ruixing Seed Industry Co., Ltd., to test the predictive model (ISSA-LSTM) of this invention. Six physiological and biochemical indicators were selected: electrical conductivity, peak viscosity, chewiness, light transmittance, water absorption, and fatty acid value. The main criteria for quality inspection were based on factors such as color and odor during wheat storage, taste scores, and gluten water absorption. Data was tested every 60 days during the storage period from 0 to 300 days, and the labeled sample was expanded. This yielded time-series data for the six indicators of Ruixing No. 1 wheat under the following conditions: 15℃, 50% RH (condition A); 20℃, 65% RH (condition B); 28℃, 75% RH (condition C); and 35℃, 85% RH (condition D).

[0097] Because the internal physiological activities and deterioration rate of wheat accelerate under high temperature conditions, training with data from higher temperature environments helps the model to acquire the differential characteristics of each indicator more quickly. The characteristic values ​​of each indicator are shown in Table 1. In order to quantitatively evaluate the effectiveness and accuracy of the ISSA-LSTM optimized neural network prediction model, RMSE is used for model evaluation, as shown in Equation (17), which represents the root mean square error between the predicted value and the actual value of the trained LSTM model.

[0098]

[0099] Table 1

[0100]

[0101]

[0102] Spearman's correlation coefficients between electrical conductivity, peak viscosity, chewiness, transmittance, water absorption, fatty acid value, and storage time show that the four physicochemical quality parameters of water absorption, chewiness, peak viscosity, and fatty acid value are highly correlated with storage time, with correlation coefficients all above 0.9, while the correlation between electrical conductivity and transmittance and storage time is relatively weak.

[0103] The correlation between storage time and various quality parameters, such as Figure 5 As shown in Table 2, conductivity and transmittance, except for their weak correlation with storage time, showed correlations with other quality parameters all below 0.8. Considering the correlation among various physiological and biochemical indicators of stored wheat, and to improve training efficiency and reduce the interference of weakly correlated factors on the experimental results, conductivity and transmittance were removed, and fatty acid value, water absorption rate, chewiness, and peak viscosity were selected as the predicted wheat quality parameters. The standardized Euclidean distance was used to measure the overall distance between each indicator, and the results are shown in Table 2. It was also found that the overall distances of fatty acid value, water absorption rate, chewiness, and peak viscosity were smaller than those of conductivity and transmittance, indicating that the selected wheat physiological and biochemical indicators were reasonable and, to a certain extent, avoided the influence of large differences in correlation between samples on the experimental results.

[0104] Table 2

[0105]

[0106]

[0107] Extensive experiments have shown that the structural parameters of LSTM networks can significantly impact training results. Therefore, this invention's ISSA-LSTM model primarily analyzes the effects of the number of neurons in the LSTM double hidden layer, the learn rate, and the droupout ratio on the efficiency and accuracy of model training. Using the same SSA-LSTM model structure (population size 30, number of hidden layers 2, loss function MSE, epoch 100, etc.), the impact of introducing t-distribution perturbations on model training and prediction is compared. The iteration curves for introducing and not introducing t-distributions are shown below. Figure 6a , Figure 6b , Figure 6c and Figure 6d As shown.

[0108] Depend on Figure 6a , Figure 6b , Figure 6c and Figure 6dIt can be seen that among the four indicators, the ISSA-LSTM model with t-distribution perturbation has a better RMSE iteration value than the SSA-LSTM model. From the perspective of global convergence ability, the SSA-LSTM model tends to stabilize in the later stages of training, while the ISSA-LSTM model with t-distribution perturbation exhibits stronger optimization ability in the early and middle stages of training due to the addition of t-distribution perturbation. It can quickly break the local optimum within a shorter iteration cycle, and the RMSE value also decreases. In terms of optimization ability, the ISSA-LSTM model optimizes faster, requiring generally fewer iterations than the original SSA-LSTM model. Looking at the RMSE values ​​of the model iterations, the final RMSE value of ISSA-LSTM is lower than that of SSA-LSTM, indirectly reflecting higher prediction accuracy and smaller prediction error.

[0109] By optimizing the structural parameters of the combined model, the comprehensive error of ISSA-LSTTM prediction of multi-indicator time-series data can be obtained. Further comparative analysis of the performance of wheat indicators at different storage stages is still needed. Therefore, taking Ruixing No. 1 as an example, the actual values ​​and predicted values ​​of multiple indicators are compared as follows: Figure 7a , Figure 7b , Figure 7c and Figure 7d As shown. By Figure 7a , Figure 7b , Figure 7c and Figure 7d It is evident that under certain specific conditions, such as the fatty acid value of batches 20-25, water absorption rate of batches 10-20, and chewiness of batches 20-30, the corresponding predicted values ​​show significant errors. This indicates that the wheat quality deteriorated too rapidly during these periods, leading to actual values ​​differing from predicted values. Therefore, in multi-indicator prediction, the impact of storage stages and environmental changes on various indicator data should be fully considered. Furthermore, because the differences in chewiness between samples in different environments are already greater than those of the other three physiological and biochemical parameters, the absolute value of the error between its predicted and actual values ​​is larger compared to the errors of the predicted and actual values ​​of the other three physiological and biochemical parameters, to the same degree.

[0110] Comparison of prediction errors with other models

[0111] To compare the accuracy of our model with other time-series prediction models, we used the same multi-indicator wheat time-series data to construct prediction results based on three commonly used prediction models: BP, LSTM, and SSA-LSTM. The prediction results of different models were obtained by optimizing the model parameters, as shown in Table 3. The SSA-LSTM and ISSA-LSTM models used the same population initialization parameters during training; the BP, LSTM, SSA-LSTM, and ISSA-LSTM models all underwent the same number of training iterations.

[0112] Table 3

[0113] Quality Indicators ISSA-LSTM SSA-LSTM LSTM BP Fatty acid value 0.213 0.259 0.274 0.288 Peak viscosity 0.249 0.271 0.290 0.364 Water absorption rate 0.284 0.299 0.322 0.365 Chewability 0.193 0.236 0.245 0.260 Comprehensive error 0.939 1.065 1.131 1.277

[0114] As shown in Table 3, the ISSA-LSTM model achieved the smallest overall error of 0.939 in predicting multiple quality parameters. This is 0.126, 0.192, and 0.338 lower than the overall errors of the SSA-LSTM model (1.065), LSTM model (1.131), and BP model (1.277), respectively, and the accuracy is improved by 11.83%, 16.98%, and 26.5%. This is because the model used in this invention predicts the overall trend of wheat quality changes based on long-term time-series observations and combines LSTM to reduce the overall error of multiple index sequences. In contrast, the parameter settings in the classic LSTM algorithm rely on manual settings, and the SSA-LSTM algorithm has poor local optimization capabilities, both of which lead to significant errors in the prediction results. Therefore, the ISSA-LSTM model of this invention is more suitable for predicting and analyzing multiple quality parameters of wheat and is also suitable for other types of stored grains, such as corn.

Claims

1. A multi-indicator prediction method for grain storage, characterized in that, The prediction method includes the following steps: 1) Collect the physicochemical characteristics of stored grain at different storage stages and under different environmental conditions to form stored grain label samples; 2) Using LSTM as the prediction model, the improved SSA algorithm is used to optimize the initial weights and thresholds in the prediction model. The improved SSA algorithm refers to the perturbation of the sparrow's position by the t-distribution mutation operator with the iteration number iter as the degree of freedom parameter. 3) Input the physicochemical characteristics of the stored grain from the obtained stored grain label samples into the optimized prediction model to achieve prediction of multiple indicators of stored grain.

2. The multi-indicator prediction method for grain storage according to claim 1, characterized in that, The method also includes expanding the grain label sample based on the characteristic that the physicochemical characteristics of stored grain maintain a stable increase or decrease in the short term under the same environmental conditions.

3. The multi-indicator prediction method for grain storage according to claim 1, characterized in that, The physicochemical characteristics of the stored grains in the sample of stored grain labels are the characteristics after screening. The screening process is as follows: the correlation between each physicochemical characteristic of stored grains and storage time is analyzed, and the physicochemical characteristics of stored grains with higher correlation are selected as the input features of the prediction model.

4. The method for predicting grain storage using multiple indicators according to any one of claims 1-3, characterized in that, The position of the sparrow after disturbance in step 2) is: X i t,+ j 1 =X i t ,j +X i t ,j t(iter) Where X i t,+ j 1 Let X be the position of the sparrow after the disturbance. i t ,j Let X be the position of the sparrow in iteration t. i t,+ j 1 t(iter) is a random disturbance term.

5. The multi-indicator prediction method for grain storage according to claim 4, characterized in that, When optimizing the initial weights and thresholds in the prediction model using the improved SSA algorithm, the t-distribution is adjusted according to the number of iterations so that the t-distribution approximates the Cauchy distribution in the early stages of the algorithm operation and when the number of iterations is small. In the later stages of the algorithm's operation, the t-distribution approximates a Gaussian distribution; in the middle stages of the algorithm's operation, the t-distribution lies between the Cauchy distribution and the Gaussian distribution.

6. The method for predicting multiple indicators of stored grain according to any one of claims 1-3, characterized in that, The physicochemical characteristics of stored grain collected in step 1) include electrical conductivity, peak viscosity, chewiness, light transmittance, water absorption rate, and fatty acid value.

7. The multi-indicator prediction method for grain storage according to claim 2, characterized in that, During the expansion, the distribution of the physical and chemical characteristics of the stored grain is made to conform to a continuous uniform distribution in the statistical distribution.

8. A multi-indicator prediction device for grain storage, characterized in that, The device includes a processor that executes a program stored in a memory to implement the multi-indicator prediction method for grain storage as described in any one of claims 1-7.

Images