A grain temperature prediction method for grain air film storehouse by fusing gradient domain denoising and adaptive modulation convolution

Abstract

The present application belongs to the technical field of grain storage monitoring and prediction, and provides a grain temperature prediction method for grain air film silo, which is fused with gradient domain denoising and adaptive modulation convolution, and comprises sample construction, adaptive smoothing, thermal inertia time coding, local feature extraction, feature modulation, context vector construction and vector mapping. The adaptive smoothing is achieved by using the mode difference between the real signal and the measurement noise sign in the differential sequence, which is alternately the same sign and lasts for a certain time, so as to preliminarily separate the effective signal from the measurement noise in the low fluctuation sequence. The decoupled fusion of the physical prior knowledge and the data-driven learning is realized by generating a differentiated exponential decay code for each historical time step of each temperature measurement layer and splicing it to the input in the form of an independent feature channel. The local signal reliability index based on the ratio of local variance to noise reference is introduced at the output end of the causal convolution, which has the fine-tuning effect of applying supplementary smoothing at the position with low signal reliability.

Description

Technical Field

[0001] This invention relates to the field of grain storage monitoring and prediction technology, and in particular to a method for predicting grain temperature in air-supported film warehouses by integrating gradient domain denoising and adaptive modulation convolution. Background Technology

[0002] Grain temperature is a core state parameter for safe grain storage. Accurate short-term, multi-step prediction of grain temperature helps grain depots plan ventilation and refrigeration operations in advance, preventing grain piles from heating up and becoming moldy, and ensuring grain quality and safety. Grain air-supported membrane silos are a new type of silo characterized by airtight, heat-insulated reinforced concrete enclosure structures. Due to their excellent heat insulation and airtightness, they are gradually being put into use. However, the grain temperature inside these silos is less affected by external disturbances, exhibiting dynamic characteristics significantly different from traditional silo types.

[0003] Monitoring data from the air-supported membrane warehouse reveals two prominent characteristics of grain temperature in this type of warehouse, posing significant challenges to existing prediction methods: First, the daily average variation in grain temperature within the warehouse is extremely low, approaching the resolvable range of conventional temperature measurement systems. The effective daily temperature change signal and measurement noise are on the same order of magnitude in amplitude, making single-step signals easily submerged by the noise and difficult to distinguish directly. Second, the air-supported membrane warehouse has been in use for a relatively short period, resulting in limited historical monitoring data suitable for modeling, representing a typical small-sample scenario. Existing grain temperature prediction methods, such as the standard Long Short-Term Memory (LSTM) network and the attention-LSTM model with superimposed attention mechanisms, do not include specific suppression mechanisms for low-fluctuation scenarios where signals are submerged by noise. Attention mechanisms inherently assign higher weights to signals with greater information content. When the effective change and noise are on the same order of magnitude, the attention weights are easily dominated by noise rather than the true trend. Under small-sample, low-fluctuation conditions, this leads to degradation in prediction accuracy, instability in multi-step predictions, and even a decline in the accuracy of attention-LSTM compared to ordinary LSTM. Some existing solutions attempt to combine convolution with recurrent networks to enhance feature extraction capabilities, but most directly perform convolutional smoothing in the temperature domain. For low-fluctuation grain temperature sequences with signal and noise of equal magnitude, it is difficult to achieve a precise balance between noise suppression and weak trend preservation, easily leading to over-smoothing that loses true change details or insufficient denoising intensity that leaves residual noise. Other studies attempt to introduce prior physical optimization of grain thermal inertia for prediction, but most adopt a multiplicative masking method to directly scale temperature values. This method artificially distorts the monotonicity and trend of the temperature sequence, destroying the physical continuity of temperature changes. This not only interferes with the accuracy of subsequent time series modeling but also causes denoising mechanisms based on the differential sign characteristics to fail. In addition, existing denoising mechanisms are mostly global processing with fixed intensity, unable to adaptively adjust the denoising intensity according to the signal quality of local regions. This easily causes the loss of effective information in areas with clear signals and fails to provide sufficient smoothing guarantees in areas with residual noise. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a grain temperature prediction method for air-supported film silos that integrates gradient domain denoising and adaptive modulation convolution, thereby solving the problems of extremely low daily temperature fluctuations in air-supported film silos, easy noise overload of single-step signals, and limited historical data.

[0005] To achieve the above objectives, the present invention provides the following solution: A method for predicting grain temperature in an air-supported film storage facility, integrating gradient domain denoising and adaptive modulation convolution, includes: The daily grain temperature sequence, concurrent environmental variables, and equivalent heat diffusion depth corresponding to each temperature measurement layer number in the grain air film warehouse were collected. The sample was constructed with the collected data corresponding to a fixed-length historical window as input and the grain temperature of the next few days as the prediction target. The original temperature sequences of each temperature measurement layer in the sample are transformed to the gradient domain. The difference sequences are adaptively smoothed according to the sign consistency index of the difference values ​​at each time position. The smoothed difference sequences are accumulated and summed to reconstruct the temperature domain, resulting in a denoised temperature sequence. The thermal response characteristic time of each temperature measurement layer is calculated based on the equivalent thermal diffusion depth corresponding to the layer number of each temperature measurement layer in the sample and the thermal diffusion coefficient of the grain. The thermal inertia time code corresponding to each time step is generated based on the thermal response characteristic time. The thermal inertial time code and the denoised temperature sequence are concatenated along the feature dimension to obtain the enhanced feature sequence, and local feature extraction is performed on the enhanced feature sequence using a convolutional front end composed of several parallel one-dimensional causal convolutional branches. The output and input of the causal convolution branch are weighted and fused based on the local signal confidence index at each time position to obtain the modulation feature sequence; the local signal confidence index is determined based on the ratio of the local neighborhood variance to the noise baseline at each time position. The modulation feature sequence is modeled using a long short-term memory network to obtain the hidden state at each time step, and the hidden state at each time step is weighted and fused using an additive attention mechanism to obtain the context vector. The context vector is mapped to the grain temperature increment for the next few days using a fully connected layer, and the grain temperature increment is added to the grain temperature of the last time step of the historical window in the sample to obtain the predicted value of the grain temperature for the next few days.

[0006] The present invention discloses the following technical effects: This invention provides a method for predicting grain temperature in air-supported film storage silos by integrating gradient domain denoising and adaptive modulation convolution. It addresses the difficulty in distinguishing signal from noise in low-fluctuation temperature sequences by adaptively smoothing the difference in sign changes between the real signal and measured noise in the differential sequence, achieving temperature trend preservation and coarse-grained signal-to-noise separation. Furthermore, it solves the problem of insufficient adaptability of a unified historical window due to differences in heat conduction lag between different temperature measurement layers by constructing differentiated thermal response time codes based on the spatial distance of the temperature measurement layers and the grain's thermal diffusivity, thus achieving a fusion of physical prior and data-driven learning. Finally, it utilizes local variance and noise benchmarks... By constructing a local signal credibility index and adjusting the convolution intensity, the problem of insufficient adaptability of fixed smoothing intensity caused by uneven distribution of residual noise is solved, and local adaptive denoising is achieved. By using concatenated gradient domain denoising, thermal inertia time coding and adaptive convolution modulation, the problem of coordinating the processing of signal and noise of the same magnitude, spatial thermal inertia differences and uneven residual noise is solved, and multi-level complementary prediction is achieved. By using the increment of grain temperature relative to the input end as the prediction target and combining it with anchor point recovery of the prediction value, the training difficulty caused by the mismatch between the absolute temperature span and the actual change scale is solved, and the prediction output space is matched with the grain temperature change scale. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 A schematic diagram of the grain temperature prediction process in an air-film grain storage facility, which integrates gradient domain denoising and adaptive modulation convolution, as provided in an embodiment of the present invention. Figure 2 A network architecture diagram provided for embodiments of the present invention; Figure 3 This is a schematic diagram illustrating the gradient domain denoising principle provided in an embodiment of the present invention. Figure 4 A schematic diagram of the thermal inertia time coding principle provided in an embodiment of the present invention; Figure 5 A schematic diagram of causal convolution and adaptive modulation provided for embodiments of the present invention; Figure 6 This is a statistical chart showing the variation of RMSE with prediction step size on a grain air-supported film warehouse, provided in an embodiment of the present invention. Figure 7 This is a comparison chart of the average RMSE over 1 to 7 days on a grain air-supported film warehouse, provided as an embodiment of the present invention. Detailed Implementation

[0009] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0010] The purpose of this invention is to provide a method for predicting grain temperature in air-supported film silos that integrates gradient domain denoising and adaptive modulation convolution, thereby addressing the shortcomings of extremely low daily temperature fluctuations in air-supported film silos, easy noise overload of single-step signals, and limited historical data.

[0011] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0012] Figure 1 This is a schematic diagram of the grain temperature prediction process in an air-supported film warehouse, which integrates gradient domain denoising and adaptive modulation convolution, as provided in an embodiment of the present invention. Figure 2 The network architecture diagram provided for the embodiments of the present invention is as follows: Figure 1 and Figure 2 As shown, this invention provides a method for predicting grain temperature in a grain air-film storage facility by fusing gradient domain denoising and adaptive modulation convolution, comprising: Step 100: Collect the daily grain temperature sequence of each temperature measuring layer in the grain air film warehouse, the environmental variables at the same time, and the equivalent heat diffusion depth corresponding to the layer number of each temperature measuring layer (determined by linear interpolation according to the warehouse radius, grain loading height and the total number of temperature measuring layers). Construct a sample with the collected data corresponding to the fixed-length historical window as input and the grain temperature of the next few days as the prediction target. Step 200: Transform the original temperature sequence of each temperature measurement layer in the sample to the gradient domain, perform adaptive smoothing on the difference sequence according to the sign consistency index of the difference value at each time position, accumulate and sum the smoothed difference sequence to reconstruct it back to the temperature domain, and obtain the denoised temperature sequence. Step 300: Calculate the thermal response characteristic time of each temperature measurement layer based on the equivalent thermal diffusion depth corresponding to the layer number of each temperature measurement layer in the sample and the thermal diffusion coefficient of the grain, and generate the thermal inertia time code corresponding to each time step based on the thermal response characteristic time. Step 400: The thermal inertial time code and the denoised temperature sequence are concatenated along the feature dimension to obtain the enhanced feature sequence, and local feature extraction is performed on the enhanced feature sequence using a convolutional front end composed of several parallel one-dimensional causal convolutional branches. Step 500: The output and input of the causal convolution branch are weighted and fused according to the local signal confidence index at each time position to obtain the modulation feature sequence; the local signal confidence index is determined according to the ratio of the local neighborhood variance to the noise baseline at each time position. Step 600: Use a long short-term memory network to perform temporal dependency modeling on the modulation feature sequence to obtain the hidden state at each time step, and use an additive attention mechanism to perform weighted fusion of the hidden states at each time step to obtain the context vector; Step 700: Use a fully connected layer to map the context vector into the grain temperature increment for the next few days, and add the grain temperature increment to the grain temperature of the last time step of the historical window in the sample to obtain the predicted value of the grain temperature for the next few days.

[0013] The following is combined with Figure 2 The overall architecture of the model shown will be explained step by step, illustrating the technical solution of this embodiment.

[0014] S1 Data Acquisition and Sample Construction.

[0015] Daily grain temperature sequences were collected from various temperature monitoring layers arranged along the height of the grain pile within the grain air-supported silo. Simultaneously, four environmental variables were recorded: internal temperature, internal humidity, external temperature, and external humidity. In this embodiment, the prediction target is the average grain temperature of the effective temperature monitoring points within the same temperature monitoring layer. Therefore... It is not interpreted as the unique geometric distance to a specific actual temperature measurement point on that layer, but is defined as the first... The equivalent thermal diffusion depth corresponds to each temperature measurement layer number. The equivalent thermal diffusion depth is determined by the actual chamber radius. Grain loading height and the total number of temperature measurement layers Determined, and at the shallow endpoint With deep endpoints Linear interpolation is performed between them according to the layer number; where, and The dimensions of the chamber and the arrangement of the temperature measurement layer are determined based on the size of the chamber and the layout of the temperature measurement layer.

[0016] After adopting the above definition, This represents the equivalent parameters used to construct the hierarchical thermal inertia prior, rather than a simple replacement for the actual shortest distance between the 25 temperature measurement points within the layer. Each layer The actual radius of the storage silo, the grain loading height, and the arrangement of the temperature measuring layer are calculated based on the aforementioned linear interpolation method.

[0017] .

[0018] Using a fixed-length historical observation window as input features, and future... The temperature of grain at each level of the day was used as the target for prediction to construct samples, and the historical window length was optimized. Heaven, predicting step size optimization Day. Regarding the first [day / year] within the historical window. The temperature measuring layer is at the first The grain temperature value at each time step is recorded as follows: Environment variables are denoted as Unified regulations Indicates the earliest time step within the window. This represents the latest time step. Samples are divided into training and testing intervals in chronological order, with target dates that do not overlap. Missing values ​​are handled using linear interpolation within the training and testing intervals, respectively. Normalization parameters are calculated only for the training interval and then applied to the testing interval. Gradient domain denoising is performed in the actual temperature dimension, and thermal inertia encoding is a dimensionless channel without range normalization.

[0019] S1-1 First level: Gradient domain denoising and temperature reconstruction.

[0020] refer to Figure 3 This step is the first in all data processing and is performed before all subsequent transformations. Its core idea stems from an in-depth analysis of the signal and noise patterns in the grain temperature sequence of a grain storage facility: In the original temperature domain, the actual temperature change (e.g., a continuous, slow rise at a rate of approximately 0.03℃ per day) and the measurement noise (random fluctuations of approximately ±0.03℃ per measurement) are on the same order of magnitude in amplitude, making it extremely difficult to distinguish them directly in this domain. However, in the gradient domain, i.e., the first-order difference sequence between adjacent time steps, their directional patterns exhibit fundamental differences. The actual physical heating process is directionally consistent in the short term, and the difference sequence shows a continuous series of positive or negative values, exhibiting a characteristic of consistent sign. In contrast, the measurement noise is temporally uncorrelated, and the difference sequence shows alternating positive and negative values, exhibiting a characteristic of random sign reversal. This pattern difference makes the signal-to-noise separability significantly higher in the gradient domain than in the temperature domain.

[0021] The distinguishing property of sign persistence and sign alternation holds strictly only on the undistorted temperature sequence governed by the original physical process. Any preceding multiplicative scaling or nonlinear transformation may artificially alter the monotonicity of the sequence, causing the differential sign of the real heating signal to flip, thus presenting a pseudo-alternating pattern indistinguishable from noise in the gradient domain, leading to the failure of signal-to-noise separation. Therefore, this step must be placed at the very beginning of all data transformation operations, directly acting on the original temperature sequence constructed by S1.

[0022] Specifically, the transformation from the temperature domain to the gradient domain. For the original temperature time series of each thermometer (the superscript of the thermometer is omitted below), the first-order difference between adjacent time steps is calculated: .

[0023] in, For the first The first-order difference value at each time step and The first The and the first The original temperature values ​​at each time step. The calculated length is... Difference sequences At the same time, the initial temperature value of the sequence is preserved. This serves as the integration anchor point for subsequent temperature reconstruction. This difference operation introduces no learnable parameters and is a deterministic linear transformation.

[0024] Furthermore, a local sign consistency check is performed. This is done for each time position in the difference sequence. Take Centered on, with a total length of Local sliding window, This is the preferred value; at both ends of the sequence, the window only retains the differences within the valid range. The number of differences that are strictly greater than 0 within the statistical window is counted. The number of differences that are strictly less than 0 Zero difference is not counted in either the positive or negative case. When, positive difference ratio When all differences within the window are 0, let To make the symbol consistency index Define the symbol consistency index for this location: .

[0025] in, For time position The sign consistency index, with a value range of 100%. ; κ represents the proportion of positive differences among the non-zero differences within the window. When the number of positive and negative differences is similar, κ takes the smaller value; the minimum value of κ under the preferred window length is determined by the window length. This index is used to complete coarse-grained sign pattern discrimination and should not be interpreted as a criterion for identifying all true inflection points independently.

[0026] Furthermore, gradient-domain adaptive smoothing is performed on the difference sequences, with the smoothing strength determined by the sign consistency index. Dynamic regulation. This indicates that the arithmetic mean operation uses the same valid range as the aforementioned symbolic statistical window; when At that time, the radius of the complete window is 2 and the total length is 5, instead of a radius of 5. The smoothed difference value is given by the following formula: .

[0027] in, The difference values ​​are after smoothing. The original difference value, This is a sign consistency index. For positions close to 1 with the same sign continuation, the results primarily retain the original difference to avoid unnecessary smoothing of slow heating or cooling trends; in For alternating positive and negative positions close to 0, the results are mostly represented by the neighborhood mean to reduce random measurement jitter. Since the true trend turning point and the random sign alternation may have similar sign ratios within the local window, this step is only used as a preliminary coarse-grained processing. Turning points and changes at multiple time scales are further modeled by subsequent dual-scale causal convolution, LSTM, and attention mechanisms.

[0028] Specifically, the initial temperature value is retained. Using the integration anchor point, perform cumulative summation on the denoised difference sequence to reconstruct the temperature domain step by step: .

[0029] .

[0030] in, This is the denoised temperature value after reconstruction. The initial temperature value of the sequence. For the first The difference values ​​after smoothing at each time step The length of the historical window is defined. The reconstructed temperature sequence has the same length as the original sequence, without altering the temporal dimension of subsequent networks. The aforementioned centered window is only used within the historical observation window that has been fully acquired; it does not read prediction target data after the end of the historical window, thus preventing training target leakage. The causal convolution referred to in this scheme means that the convolution output depends only on the current and earlier augmented features. As an integral anchor point used to maintain an absolute temperature reference, the gradient domain processing mainly aims to improve the variation characteristics fed into subsequent networks, and does not claim to eliminate all measurement errors that the anchor point itself may contain.

[0031] The diurnal fluctuations of the four environmental variables—indoor temperature, indoor humidity, outdoor temperature, and outdoor humidity—are typically much greater than those of grain temperature. Furthermore, these variables are collected by independent sensors with significantly higher accuracy than grain temperature measurement elements, resulting in a much lower noise-to-signal ratio compared to the grain temperature sequence. Therefore, the environmental variables are not processed through the gradient domain described above and their original values ​​are directly retained. , and the reconstructed grain temperature sequence These features are combined to form the input features for subsequent steps.

[0032] S1-2 Second level: Thermal inertial guided layered time-coded splicing.

[0033] refer to Figure 4This step is performed after gradient domain denoising and temperature reconstruction, and before the causal convolution front-end. Its design aims to introduce the unique thermal conduction physics priors of grain air-supported film storage into the model, so that temperature sensing layers at different depths can obtain differentiated time sensing capabilities that match their physical thermal response time scales, while strictly avoiding any damage to the physical form of the temperature sequence itself.

[0034] Grains are materials with extremely poor thermal conductivity; the thermal diffusivity of soybeans is typically around [value missing]. The process of external environmental temperature disturbances propagating from the grain pile wall or surface to the interior of the grain pile is extremely slow. Surface temperature measurement points, tens of centimeters from the boundary, can respond to environmental changes within days, while core temperature measurement points, several meters to tens of meters from the boundary, can have a response lag of up to several months. This means that for surface measurement points, earlier historical data within a 30-day input window may already have a weak physical correlation with the current grain temperature; while for core measurement points, historical data from the same location within the same window may still have high reference value. Existing methods neglect this spatiotemporal difference determined by the physical laws of heat diffusion.

[0035] Therefore, this step introduces thermal inertia prior by concatenating independent feature channels instead of directly scaling the temperature value using a multiplicative mask. This design choice has a key physical logic consideration: if the thermal inertia decay coefficient is directly applied to the temperature value through element-wise multiplication, even if the original sequence is monotonically increasing in temperature, multiplying it by a monotonically decreasing mask may result in a decreasing segment in the product sequence, causing the difference sign to flip from positive to negative. This artificially introduced sign flip would violate the physical premise that the true direction of change is continuous, upon which the first-level gradient domain denoising relies. By changing to feature channel concatenation, the temperature value itself is not scaled by any coefficient, and its physical form is completely preserved and passed to subsequent networks as is; the thermal inertia information exists as an independent contextual feature channel, and causal convolution and LSTM can autonomously learn the joint relationship between the temperature value and the temporal encoding in subsequent processing.

[0036] Specifically, for each temperature measurement layer The equivalent thermal diffusion depth is obtained by linear interpolation according to the layer number in S1. Based on the thermal diffusivity of grain and Calculate the thermal response characteristic time of the thermometer: .

[0037] in, For the first The thermal response characteristic time of each temperature sensing layer This represents the equivalent thermal diffusion depth corresponding to this layer. The thermal diffusivity of the grain. Used to characterize the time scale by which external thermal disturbances are transmitted to the equivalent location at this level.

[0038] Furthermore, for each temperature measurement layer Based on its thermal response characteristic time This generates a thermal inertia time code for each historical time step of the input window. The temperature sequence still uses... and For the latest time; also define a historical lag index that corresponds one-to-one with the temperature time step. Therefore, the latest moment The furthest historical moment The encoded value is calculated using the following formula: .

[0039] in, For the first Each temperature sensor layer in historical lag index Thermal inertia time encoding value at that location, and Use the same time unit. When concatenating with temperature sequences, the first... Temperature value Corresponding code This ensures the latest temperature and The encoding is paired. The surface layer corresponds to... Smaller, the encoding decays faster as historical lag increases; deeper correspondences The value is relatively large; the encoding within a 30-day window may be close to 1, indicating that the layer has a long thermal memory within the current window scale, and is not a computational failure. Thermal inertia encoding is not normalized by channel range to avoid division by zero or abnormal amplification of near-constant channels. The equivalent thermal diffusion depth is taken as the smaller of the vertical distance from the temperature-measuring layer to the grain pile surface and the horizontal distance to the side wall of the storage silo, i.e., the nearest heat transfer boundary.

[0040] Furthermore, the thermal inertia time-series encoding is treated as an independent feature channel and concatenated along the feature dimension with the denoised temperature sequence output from the first level. For the temperature measurement layer... At time step The input, and the concatenated enhanced feature vector are: .

[0041] in, This represents a concatenation operation along the feature dimension. The temperature value is the result of gradient domain denoising and reconstruction. This is a time-encoded scalar for thermal inertia. For environmental variables... Since it reflects the macroscopic meteorological conditions of the entire warehouse or outside the warehouse and does not change with the sensor depth, its raw value is directly added as an additional feature channel: the complete input feature vector of each temperature measurement layer is The enhanced features of all temperature measurement layers and all time steps are organized into input tensors in the original manner and fed into the subsequent causal convolutional front end.

[0042] S2 Level 3: Causal convolutional front-end feature extraction with adaptive noise-aware modulation.

[0043] refer to Figure 5 The enhanced feature sequences output by S1-1 and S1-2 have completed coarse-grained gradient domain signal-to-noise separation and thermal inertial physics prior injection. Based on this, this step firstly extracts local patterns of multiple time scales through parallel dual-scale causal convolution, and secondly introduces an adaptive modulation mechanism based on local signal credibility at the output of causal convolution to achieve fine-grained denoising and fine-tuning.

[0044] Specifically, the noise baseline variance used in the third level is individually calibrated for the sequence processed by S1-1. Within the training interval, a candidate window is defined as 14 consecutive days; the reconstructed temperature range within this window is considered to be no greater than... ℃ and the absolute value of the slope of the linear fit is not greater than When the temperature is measured in ℃ / d, it is defined as a historical stable period. The variance of the first-order difference of the reconstructed temperature within all stable periods that meet the conditions is calculated and used as the noise baseline variance for the current stage. .

[0045] in, The current noise baseline variance. This is the set of indices for stable time periods within the training interval that meet the above criteria. For the sample size, This is the first-order difference mean for that period. The model is calibrated once and kept fixed before training. (S1-1) Determined solely by the proportion of the difference signs, and independent of ; Used to describe the level of residual fluctuations in the phase following S1-1.

[0046] Furthermore, parallel dual-scale causal convolution processing is performed. The input vector at each time step of the output S1-2 is... It includes 14 denoised temperature channels, 14 thermal inertial coding channels, and 4 environmental variable channels. The enhanced feature sequence is fed into two parallel one-dimensional causal convolution branches, with the convolution kernel sizes preferably optimized. and This corresponds to two types of local receptive fields: short-term and medium-term. To ensure that the input and output correspond layer by layer in subsequent modulation, the number of output channels in each branch is preferably set to [value missing]. Each output channel corresponds one-to-one with one of the 14 temperature measurement layers. Branches at time step The convolution output is given by the following formula: .

[0047] in, For the first Branch number The weight matrix at each convolution position. For bias vectors, It is the ReLU activation function. Each output channel integrates the denoised temperature, thermal inertia coding, and four environmental variables of the corresponding layer, but the output channel number remains consistent with the temperature measurement layer number. To maintain causality, only the historical time dimension is filled. A zero vector, making the time step The convolutional output depends only on the current and past augmentation features, and the length of the output sequence is the same as the input sequence.

[0048] Specifically, for sequences that have undergone gradient-domain denoising, the residual noise levels at different time points are not uniform: at locations where the signal is strong in the original sequence, the signal-to-noise ratio after gradient-domain denoising is already high; at locations where the signal is extremely weak in the original sequence, gradient-domain denoising may leave some noise due to the lack of significant sign advantage. Therefore, applying the same intensity of convolutional smoothing to the entire sequence is unreasonable. To address this, this step introduces a local signal confidence index to evaluate the signal quality at each time step and adjust the intensity of convolutional smoothing accordingly.

[0049] For the denoised layer average temperature sequence, at each time position Take Center, radius local neighborhood , The optimal value is selected; only valid ranges are considered at the boundaries. This local statistic is also calculated only within the historical input window where observations have already been completed. The local variance and local first-order autocorrelation coefficient are calculated separately. The local variance is: .

[0050] in, For time location The local variance, For The local neighborhood centered on, The number of time steps within the neighborhood. For the neighboring region The denoising temperature value at each time step. For the neighborhood The average temperature inside. It reflects the magnitude of local temperature fluctuations at that location.

[0051] Local first-order autocorrelation coefficient: .

[0052] in, For time location The local first-order autocorrelation coefficient, and The denoised temperature values ​​are the values ​​of adjacent time steps within the neighborhood. The average temperature within the neighborhood. It is the smallest time index in the neighborhood. Reflecting the persistence of temperature changes within this neighborhood: Due to the large thermal inertia of real physical processes, there is a positive correlation between temperature values ​​of adjacent days. It is positive and deviates from zero; while the residual noise is time-independent. It is close to zero or negative.

[0053] The ratio of local variance to the noise baseline variance is multiplied by the truncated local autocorrelation coefficient and then mapped using the sigmoid function to... The local signal confidence index is obtained from the interval: .

[0054] in, For time location The local signal credibility index For local temperature variance, To reconstruct the first-order difference variance of temperature during a stationary period, The local first-order autocorrelation coefficient. This is a dimensionless steepness coefficient. Given that, for approximate white noise, the theoretical ratio of temperature variance to first-order difference variance is approximately... This embodiment does not use a fixed method. , Within the training interval, calibration is performed based on the ratio distribution, with the optimal value being... , This value is used to avoid Premature saturation results in a hard switch that approximates 0 or 1; in practical implementation, this can be determined using the training or validation interval. and However, test target data must not be used.

[0055] Furthermore, for each causal convolution branch, its convolution output vector at time step t is fused with the corresponding denoised temperature input vector using an element-wise weighted fusion method based on the local signal confidence index: .

[0056] in, For the first Branches at time step The modulated output vector, This is a local signal reliability index. for 3D denoised temperature vector For the first Branches 3D convolution output vector In the formula This indicates the identity expansion corresponding to each layer; in actual computation, both the denoising temperature and the convolution input are first converted to dimensionless standardized values ​​using the training interval parameters, thus the two have the same numerical scale. Because It is a vector containing 14 layer values ​​rather than a single scalar, even The output channels are relatively large, and each output channel still retains the values ​​of different temperature measurement layers, and will not collapse into all the same constant channels.

[0057] The adaptive modulation and gradient domain denoising in S1-1 form a clear division of labor: gradient domain denoising utilizes the sign consistency of the difference sequences to achieve coarse-grained signal-to-noise separation globally, while adaptive modulation utilizes the comparison between local variance and the noise baseline to achieve fine-grained tuning locally. After each of the two convolutional branches completes modulation fusion, their outputs are concatenated along the feature dimension to form a combined local feature. .

[0058] in, For time step The combined local feature vectors, and These are the modulated outputs of the first and second convolutional branches, respectively. This indicates a concatenation operation along the feature dimension. This represents the number of output channels for each branch. This combined feature... ( This is the final output of the causal convolution front end, which is fed into the subsequent LSTM layer for temporal dependency modeling.

[0059] S3: LSTM time-series dependency modeling.

[0060] The combined local feature sequence output from the front end of the S2 causal convolution The data is fed into a single-layer Long Short-Term Memory (LSTM) network, and its gating mechanism is used to progressively update the hidden states and cell states, modeling long-range temporal dependencies in the grain-temperature sequence. Standard LSTM, at time steps... The forward computation process is as follows: .

[0061] .

[0062] .

[0063] .

[0064] .

[0065] .

[0066] in, For the sigmoid function, It is the hyperbolic tangent function. This indicates element-wise multiplication. and These are the weight matrices for each gate corresponding to the current input and the previous hidden state, respectively. The bias vector is used, and the optimal number of hidden units is 64. Initial hidden state. and initial cell state All vectors are initialized to zero. The LSTM layer outputs all... Hidden state sequence at each time step This information is then used for weighted fusion by subsequent attention mechanisms.

[0067] S4 additive attention-weighted fusion.

[0068] Within the 30-day input window of the grain temperature sequence in the air-supported warehouse, the contribution of different historical time steps to the prediction of future grain temperature varies. To automatically identify and reinforce key periods with high predictive contributions, an additive attention mechanism is applied over all hidden states.

[0069] For each time step Calculate the attention score. (The hidden state is then considered.) After linear transformation and tanh nonlinear activation, it is coupled with a learnable context query vector. Do the inner product: .

[0070] in, For attention projection matrix, For the bias vector, the preferred dimension of the attention space is 64. During training, backpropagation is used for learning, which can be understood as the model learning to focus on an ideal temporal state template and matching it with the projection of the hidden state at each time step to obtain a score. A higher value indicates a stronger correlation between that time step and the prediction target.

[0071] Then, the scores at all time steps are normalized using the softmax function to obtain the attention weights for each time step: .

[0072] in, For time step Attention weights, satisfying and Finally, the hidden states at all time steps are summed using attention weights to obtain the context vector: .

[0073] in, The context vector is an attention-weighted compressed representation of all historical information in the input window. During the weighting process, time steps with higher relevance receive larger weights, while time steps with lower relevance are suppressed. Pre-stage gradient domain processing, hot inertial coding, and causal convolutional modulation together provide the attention mechanism with structured temporal features.

[0074] S5 residual anchor-based multi-step prediction output.

[0075] context vector Input a fully connected layer and output a future layer. Forecast of daily grain temperature increase: .

[0076] in, This is the output layer weight matrix. This is the output layer bias vector. The Each component This represents the first day's food temperature relative to the end of the input sequence. Forecast value of temperature increase in grain production.

[0077] Take the grain temperature value at the last time step in the input sequence. Using this as an anchor point, the predicted increment is added to the anchor point daily to obtain the predicted absolute grain temperature values ​​for each future day: .

[0078] in, For the future Forecast value of grain temperature for the day, The actual grain temperature measured on the day of the final step in the historical window. To predict the increment, For the prediction step size, this embodiment uses the original grain temperature at the last time step as the residual anchor point, as per the claims. This value represents the latest actual observation obtained at the prediction starting point, used to maintain consistency between the predicted absolute temperature and the current warehouse condition. The S1-1 reconstructed sequence is used to extract trend and local change features. Using the original anchor point does not negate the effect of the previous denoising stage, but rather avoids directly inputting the phase lag or reference offset that may be caused by filtering into the absolute prediction value. The remaining statements regarding incremental learning reducing the output scale and improving training stability remain unchanged.

[0079] The model uses mean squared error as the loss function and employs the Adam optimizer for end-to-end training. .

[0080] in, The value of the loss function. The number of training samples, To predict the step size, For the first The sample at the th Predict the actual grain temperature value in the step. This corresponds to the predicted value. During training, the gradient domain transformation and reconstruction steps in S1-1, the thermal inertia time encoding generation steps in S1-2, and the difference, cumulative summation, exponential encoding, local variance and autocorrelation, sigmoid mapping, and other operations in the local statistics calculation steps in S2 are all deterministic forward calculations without trainable parameters, and do not introduce additional learnable parameters; the trainable parameters of the model come only from the causal convolution kernel weights and biases in S2, the LSTM gate weights and biases in S3, the attention parameters in S4, and the output layer parameters in S5.

[0081] Optionally, evaluation metrics: Prediction accuracy is assessed using root mean square error (RMSE) and mean absolute error (MAE), with errors calculated in actual degrees Celsius. .

[0082] .

[0083] in, For the sample size, For the first The actual grain temperature value of each sample For the first Predicted grain temperature values ​​for each sample.

[0084] Further, performance evaluation and analysis.

[0085] Experimental setup: All models use the same basic training protocol: input window Day, the initial learning rate of the Adam optimizer is... The learning rate was scheduled using cosine annealing with a batch size of 64, and training lasted for 120 epochs with 64 hidden units. To reduce the impact of randomness, each learnable model was trained under multiple random seeds, and the average results were reported. The training and test sets were divided chronologically, and missing values ​​were interpolated within each data interval. Normalization parameters were calculated only from the training interval. The comparative models in Table 1 used the same input window, residual objective, and evaluation criteria.

[0086] Basic network prediction results for grain air-supported film warehouses: Table 1 shows the RMSE data for each prediction step size for four grain air-supported film warehouses from the next 1 to 7 days. Persistence is the non-learnable reference baseline.

[0087] Table 1

[0088] When the daily variation of the sequence is approximately 0.03℃, directly extrapolating from the previous day's observations is already close to the measurement noise level, and any trainable model will find it difficult to significantly surpass this in terms of single-step error. The technical advantage of this method lies in two points that the baseline cannot provide: the error does not increase rapidly with the prediction step size. As shown in Table 1, the error of the baseline increases by 3.68 times from day 1 to day 7, while in this embodiment it is 2.07 times; all four chambers can operate using the same set of parameters, without the need for chamber-by-chamber parameter tuning.

[0089] Based on the 28 cells × prediction step size listed in Table 1, the average RMSE of the dual-scale base network across the entire time domain is approximately 0.2159℃, the standard LSTM is approximately 0.2607℃, the GRU is approximately 0.4033℃, the LSTM-Attention is approximately 0.2668℃, and the SVR is approximately 0.2108℃. The average RMSE of the base network is reduced by approximately 19.1% compared to the LSTM-Attention in the table. The SVR is 0.2108℃, which differs from the 0.2159℃ of this embodiment by 0.005℃, lower than the 0.1℃ resolution of the temperature measurement system, and does not constitute a discernible difference in accuracy. Furthermore, the cross-cell stability of the SVR is significantly poor, with a mean range of 0.3034℃ across the four cells, compared to 0.1404℃ in this embodiment.

[0090] Furthermore, the relative performance of LSTM-Attention compared to ordinary LSTM varies across different bins: in bin 2, the prediction step size errors of LSTM-Attention are higher than those of LSTM, but this is not generally the case in the other bins. This indicates that the effectiveness of the attention mechanism under low volatility and small sample conditions is related to the specific bin data distribution. Figure 6 and Figure 7 It is still used to present the error changes of the basic network and comparison method corresponding to Table 1.

[0091] The beneficial effects of this invention are as follows: (1) Gradient domain adaptive smoothing utilizes the persistence of differential sign to perform coarse-grained processing on low-fluctuation grain temperature sequences, which can reduce the interference of random alternation of positive and negative signs on subsequent feature extraction, while retaining the persistent trend of the same sign as the main input information.

[0092] (2) Calculate the thermal response characteristic time based on the equivalent thermal diffusion depth and grain thermal diffusion coefficient corresponding to each temperature measurement layer number, and generate an exponential decay code corresponding to the historical lag time, which can provide the hierarchical thermal inertia difference to the model without directly scaling the temperature value.

[0093] (3) By using the thermal inertia decay coefficient as an independent feature channel and splicing it with the temperature value instead of directly scaling the temperature value, the physical and logical conflict of temperature sequence monotonicity being distorted and differential sign being artificially flipped in the multiplicative masking scheme leading to gradient domain denoising failure is resolved. The introduction of physical prior does not disrupt the normal operation of the subsequent signal-to-noise separation mechanism.

[0094] (4) The local signal credibility index is determined by local variance, stage noise benchmark and local first-order autocorrelation coefficient, and the denoised temperature vector corresponding to the layer is weighted and fused with the convolution output vector so that the intensity of convolution feature extraction can be adjusted according to the local fluctuation state within the historical window.

[0095] (5) By re-statistically calculating the noise baseline variance from the historical stable period that meets the explicit judgment conditions after gradient domain processing, the local confidence calculation adopts the residual fluctuation scale corresponding to the current processing stage.

[0096] (6) By organizing the three layers of gradient domain denoising, thermal inertial time coding and adaptive convolution modulation in sequence, each layer can solve the problem independently in its most suitable dimension without interfering with each other.

[0097] (7) By changing the model prediction target from the absolute temperature value to the increment of the grain temperature relative to the end of the input sequence on the current day, and then adding the increment to the anchor point to obtain the final prediction value, the output spatial scale is matched with the change of the prediction target.

[0098] (8) By setting coarse-grained gradient domain processing and local adaptive convolution modulation in sequence, the two processing levels are divided in terms of operation domain, processing granularity and control basis, and maintain a continuous data flow with the subsequent LSTM, additive attention and residual anchor output.

[0099] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0100] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting grain temperature in an air-supported film storage facility by integrating gradient domain denoising and adaptive modulation convolution, characterized in that, include: The daily grain temperature sequence, concurrent environmental variables, and equivalent heat diffusion depth corresponding to each temperature measurement layer number in the grain air film warehouse were collected. The sample was constructed with the collected data corresponding to a fixed-length historical window as input and the grain temperature of the next few days as the prediction target. The original temperature sequences of each temperature measurement layer in the sample are transformed to the gradient domain. The difference sequences are adaptively smoothed according to the sign consistency index of the difference values ​​at each time position. The smoothed difference sequences are accumulated and summed to reconstruct the temperature domain, resulting in a denoised temperature sequence. The thermal response characteristic time of each temperature measurement layer is calculated based on the equivalent thermal diffusion depth corresponding to the layer number of each temperature measurement layer in the sample and the thermal diffusion coefficient of the grain. The thermal inertia time code corresponding to each time step is generated based on the thermal response characteristic time. The thermal inertial time code and the denoised temperature sequence are concatenated along the feature dimension to obtain the enhanced feature sequence, and local feature extraction is performed on the enhanced feature sequence using a convolutional front end composed of several parallel one-dimensional causal convolutional branches. The output and input of the causal convolution branch are weighted and fused based on the local signal confidence index at each time position to obtain the modulation feature sequence; the local signal confidence index is determined based on the ratio of the local neighborhood variance to the noise baseline at each time position. The modulation feature sequence is modeled using a long short-term memory network to obtain the hidden state at each time step, and the hidden state at each time step is weighted and fused using an additive attention mechanism to obtain the context vector. The context vector is mapped to the grain temperature increment for the next few days using a fully connected layer, and the grain temperature increment is added to the grain temperature of the last time step of the historical window in the sample to obtain the predicted value of the grain temperature for the next few days.

2. The method for predicting grain temperature in a grain air-film storage facility by fusing gradient domain denoising and adaptive modulation convolution as described in claim 1, characterized in that, The original temperature sequences of each thermometer layer in the sample are transformed to the gradient domain. The difference sequences are then adaptively smoothed based on the sign consistency index of the difference values ​​at each time point. The smoothed difference sequences are accumulated and summed to reconstruct the temperature domain, resulting in a denoised temperature sequence, including: The original temperature sequences of adjacent time steps of each temperature measurement layer are calculated using first-order difference to obtain the difference sequence, and the initial temperature value of the sequence is set as the reconstruction anchor point. A fixed-length local sliding window is extracted at each time position of the difference sequence, and the proportion of positive difference values ​​within the window is counted. The sign consistency index is determined based on the proportion. The difference sequence is adaptively smoothed according to the symbol consistency index; Using the initial temperature value of the sequence as an anchor point, the smoothed difference sequence is cumulatively summed step by step over time to obtain the denoised temperature sequence.

3. The method for predicting grain temperature in a grain air-film storage facility by fusing gradient domain denoising and adaptive modulation convolution as described in claim 2, characterized in that, The expression for the symbol consistency index is: The expression for the adaptive smoothing is: ;in, For the consistency of symbols; This represents the proportion of positive differences among the non-zero differences within the window. These are the smoothed difference values; The original difference value; Indicated by It calculates the arithmetic mean within a local neighborhood centered on the locator.

4. The method for predicting grain temperature in a grain air-film storage facility by fusing gradient domain denoising and adaptive modulation convolution as described in claim 1, characterized in that, The expression for the characteristic time of the thermal response is: The expression for thermal inertia time coding is: ;in, For the first The thermal response characteristic time of each temperature sensing layer; The equivalent heat diffusion depth is the smaller of the vertical distance from the temperature measuring layer to the surface of the grain pile and the horizontal distance to the side wall of the storage unit. The thermal diffusivity of the grain; For the first The thermal inertia time encoding value of each temperature sensing layer; This is the index number.

5. The method for predicting grain temperature in an air-supported film storage facility by fusing gradient domain denoising and adaptive modulation convolution as described in claim 1, characterized in that, Before weighted fusion of the output and input of the causal convolution branch based on the local signal confidence index at each time position to obtain the modulation feature sequence, the process further includes: Select the historical stationary time period sequence after gradient domain denoising and reconstruction, and set the variance of the first difference of the historical stationary time period sequence as the noise reference variance of the current stage. Determine the local variance and local first-order autocorrelation coefficient of the local neighborhood corresponding to each time position in the denoised temperature sequence; The local signal reliability index is determined based on the noise baseline variance, the local variance, and the local first-order autocorrelation coefficient.

6. The method for predicting grain temperature in a grain air-film storage facility by fusing gradient domain denoising and adaptive modulation convolution as described in claim 5, characterized in that, The expression for the local variance is: The expression for the local signal reliability index is: ;in, For time position Local variance; For The local neighborhood centered on the center; For the neighboring region The denoising temperature value at each time step; For the neighborhood Average temperature inside; For time position The local signal reliability index; This is a dimensionless steepness coefficient; This represents the current noise baseline variance. The local first-order autocorrelation coefficient; This is the threshold parameter.

7. The method for predicting grain temperature in a grain air-film storage facility by fusing gradient domain denoising and adaptive modulation convolution as described in claim 1, characterized in that, The output and input of the causal convolution branch are weighted and fused based on the local signal confidence index at each time point to obtain a modulation feature sequence, including: The convolutional output of the causal convolutional branch at a given time position is fused with the denoised temperature input at the corresponding position using an element-wise weighted fusion method based on the local signal confidence index; the expression for the element-wise weighted fusion method is as follows: ;in, For the first Branches at time step The modulated output vector; This is a local signal reliability index; This is the denoised temperature vector; It is a vector consisting entirely of 1s; For the first Branches 3D convolution outputs a vector; The modulated outputs of all parallel branches are concatenated along the feature dimension to obtain the modulated feature sequence.

8. The method for predicting grain temperature in a grain air-film storage facility by fusing gradient domain denoising and adaptive modulation convolution as described in claim 1, characterized in that, The convolutional front end employs two parallel branches; the two parallel branches correspond to local receptive fields at two time scales; each of the two parallel branches employs causal padding; the causal padding is used to fill the input sequence with zero values ​​at the historical time dimension matching the size of the convolutional kernel, controlling the convolutional output to depend on the input at the current and historical time points.

9. The method for predicting grain temperature in an air-supported film storage facility by fusing gradient domain denoising and adaptive modulation convolution as described in claim 1, characterized in that, The concurrent environmental variables include: indoor temperature, indoor humidity, outdoor temperature, and outdoor humidity; these concurrent environmental variables, together with the denoised temperature sequence and the thermal inertia time code, constitute the input features.

10. The method for predicting grain temperature in a grain air-film storage facility by fusing gradient domain denoising and adaptive modulation convolution as described in claim 1, characterized in that, The Long Short-Term Memory (LSTM) network is a single-layer LSM network; the additive attention mechanism is a Bahdanau-style attention mechanism; the additive attention mechanism is used to extract the attention score of the hidden state at each time step after linear transformation and hyperbolic tangent activation by the learnable context query vector, and then perform softmax normalization on the score and weighted summation on the hidden state at each time step to obtain the context vector.

Images