Biodegradable mulching film degradation life prediction method suitable for complex time and space
Through the prediction network combining LSTM and multi-head self-attention method, the problem of predicting the degradation life of biodegraded mulch under complex space-time conditions is solved, and accurate degradation life prediction under different geographical and climatic conditions is achieved, and the development of the mulch industry is improved. efficiency.
Patent Information
- Application Number
- CN202411284137.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-13
- Publication Date
- 2025-01-03
AI Technical Summary
The prior art cannot effectively predict the field degradation life of biodegradable mulch at different times and spaces, especially under complex space-time conditions.
Using a prediction network with LSTM combined with multi-head self-attention method, global similarity calculation and feature extraction are performed by collecting material characteristic data and spatiotemporal data, combining agricultural IoT devices to obtain environmental information, build high-order spatial-temporal conduction continuity constraints, and generate degradation rate prediction model and improve prediction accuracy through iterative optimization.
It has achieved rapid prediction of the degradation life of biodegradable plastic film under different geographical and climatic conditions, and improved the material design efficiency and the development speed of the plastic film industry.
Smart Images
Figure CN119252376A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of artificial intelligence prediction, and in particular relates to a method for predicting the degradation life of biodegradable mulch suitable for complex time and space. Background Art
[0002] my country has a vast territory and a changeable environment. The degradation performance of biodegradable mulches in different time and space varies greatly. It is difficult to predict the degradation life of existing biodegradable mulches in the field and can only rely on long-term experimental evaluation. There is no available method to predict the degradation life of biodegradable mulches under complex temporal and spatial conditions in the field.
[0003] CN115711845A discloses a method for predicting the service life of a biodegradable mulch film. The method constructs the relationship between the service life of the biodegradable mulch film predicted in the laboratory and the service life in the actual field environment. The method samples the biodegradable mulch film at regular intervals during an indoor high-temperature ultraviolet accelerated aging test and during actual use in the field, tests the physical and chemical properties of the samples, and compares the differences between the two to establish a prediction model and method for the service life of the biodegradable mulch film. The method is only applicable to the prediction of the degradation life in a single specific scenario, and cannot achieve the prediction of the degradation life in complex time and space in the field under different geographical climates. Summary of the invention
[0004] Purpose of the invention: In order to solve the problem of predicting the degradation life of biodegradable mulch under complex temporal and spatial conditions in the field, this application proposes a method for predicting the degradation life of biodegradable mulch suitable for complex temporal and spatial conditions.
[0005] Technical solution: A method for predicting the degradation life of biodegradable mulch applicable to complex time and space, comprising the following steps:
[0006] Collect material property data and spatiotemporal data that affect the degradation life of biodegradable mulch in different regions, classify them according to time periods, and obtain time series data;
[0007] After the material property data and spatiotemporal data are encoded separately, the similarity matrix is obtained through global similarity calculation, and then jump-connected with the original data and sent to the decoder to extract features respectively. After fusion, spatiotemporal high-order derivative continuity constraints are performed to obtain enhanced data;
[0008] The enhanced data is used as the input of the prediction network, which adopts LSTM combined with a multi-head self-attention method. A stacked LSTM network is used in the encoder to capture the temporal global correlation in different time period sequences. For each time period, the enhanced data of material property data and spatiotemporal data are input, and then a multi-head self-attention module is constructed to learn the global features between the time series data and extract the most concerned local areas in the features. Finally, the degradation rate of each time period is predicted by the prediction head.
[0009] Furthermore, the material property data is sampled at fixed time intervals starting from the start time of application of the biodegradable mulch in the field environment, and includes: sample mass, mechanical properties, thermodynamic properties, rheology, melt index, thickness, light transmittance and UV shielding properties, water vapor and oxygen permeability, number of surface cracks and fissures, and degree of breakage.
[0010] Furthermore, the spatiotemporal data is the environmental information of the region obtained by using agricultural Internet of Things devices, and the average, maximum and minimum values of the environmental factors in each cycle are statistically calculated. The environmental factors include: temperature, humidity, radiation intensity, rainfall, wind speed, soil pH value, soil organic matter content, soil porosity, soil type, soil microbial abundance, and soil humidity.
[0011] Furthermore, the process of separately encoding the material characteristic data and the spatiotemporal data includes cleaning the data, removing abnormal data, replacing missing data with 0 using padding; normalizing all values; and cosine encoding the time series data.
[0012] Furthermore, after the material property data and spatiotemporal data are encoded separately, a similarity matrix is obtained by global similarity calculation, which is then jump-connected with the original data and sent to the decoder to extract features respectively; this process uses two separate encoders, and the encoding process is expressed as:
[0013] f' n =Encoder f (f n )
[0014] s' n =Encoder s (s n )
[0015] Among them, f n and n Respectively represent the material property data and spatiotemporal data in time period n, Encoder f and Encoder s There are two encoders;
[0016] The global similarity is calculated by calculating the dot product of the encoded material property data and the spatiotemporal data. The global similarity is expressed as:
[0017] G = softmax(f' n ·s' n T )
[0018] The original data is then skip-connected with the similarity matrix and then sent to the decoder for feature extraction. The formula is expressed as:
[0019] t n =Decoder([f' n ,s' n ,G])
[0020] In the formula, [f' n ,s' n ,G] means connecting the original data and the similarity matrix, and Decoder() is a decoder.
[0021] Furthermore, the space-time high-order derivative continuity constraints are specifically:
[0022] T(x,y,n)=t n , where (x, y) is the spatial position of the material in space and time, and n is the time period of data collection;
[0023] T'(x,y,n)=T'(x,y,n+1)
[0024] T”(x,y,n)=T”(x,y,n+1)
[0025] Where: T' is the first-order derivative of T, T" is the second-order derivative of T; T' and T" are continuous.
[0026] Furthermore, after obtaining the degradation rate at each time period stage, the method also includes fitting a degradation rate curve using a polynomial, obtaining the degradation cycle time by setting a degradation rate failure threshold, and subsequent experiments change the geographical area and the scope of time series coverage to perform iterative optimization and improve the model prediction accuracy.
[0027] Furthermore, the root mean square error (MSE) and mean absolute error (MAE) between the predicted value and the true value are used to characterize the goodness of fit of the model.
[0028]
[0029] In the formula, is the actual degradation rate, η n is the predicted degradation rate, and N is the number of samples.
[0030] Beneficial effects: The present invention discloses a method for predicting the degradation life of biodegradable mulch films in different time and space, which relates to the technical field of predicting the degradation life of biodegradable materials. Geographical and climate data of different regions are obtained, and various data are preprocessed to generate spatiotemporal characteristic data. A degradation rate prediction network based on global self-attention is designed to learn the divided data set to generate a degradation rate prediction model. In the model verification iteration, new data is collected in different time periods and different geographical environments, new environmental parameters are input to predict the model, and the model is iterated through the new data. Through this technology, the degradation performance verification of mulch materials for different regional environments can be quickly realized, the material design efficiency can be improved, and the development of the mulch industry can be accelerated. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 is a flow chart of an embodiment of the present invention;
[0032] Figure 2 A schematic diagram of time series data according to an embodiment of the present invention;
[0033] Figure 3 is a flow chart of a data enhancement process according to an embodiment of the present invention;
[0034] Figure 4 The flowchart of the network prediction method according to the embodiment of the present invention is shown in FIG. DETAILED DESCRIPTION
[0035] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.
[0036] like Figure 1 As shown, the technical solution provided by the present invention is a method for predicting the degradation life of biodegradable mulch applicable to complex time and space, comprising the following steps:
[0037] Collect material property data and spatiotemporal data that affect the degradation life of biodegradable mulch in different regions, classify them according to time periods, and obtain time series data;
[0038] After the material property data and spatiotemporal data are encoded separately, the similarity matrix is obtained through global similarity calculation, and then jump-connected with the original data and sent to the decoder to extract features respectively. After fusion, spatiotemporal high-order derivative continuity constraints are performed to obtain enhanced data;
[0039] Use the enhanced data as the input of the prediction network. The prediction network adopts the LSTM combined with the multi-head self-attention method. A stacked LSTM network is used in the encoder to capture the temporal global correlation in different time series. For each time period, the enhanced data of the input material properties data and spatio-temporal data are input, and then a multi-head self-attention module is constructed to learn the global features between the time series data and extract the local feature factors that need to be most attention in the features. Finally, the degradation rate of each time period is predicted through the prediction head.
[0040] Further, the material properties data starts from the starting time (time 0) when the biodegradable mulch film is applied to the field environment. Samples are taken every 10 days from 0 to 210 days, and the mass, mechanical properties, thermodynamic properties, rheology, melt index, thickness, light transmittance and UV shielding properties, water vapor and oxygen permeability, the number of surface cracks and fractures, and the degree of fragmentation of the samples are measured. The times are 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210 respectively.
[0041] Further, the spatio-temporal data is to obtain the environmental information of the region using agricultural Internet of Things devices, and calculate the average value, maximum value, and minimum value of the environmental factors in each period. The environmental factors include: temperature, humidity, radiation intensity, rainfall, wind speed, soil pH value, soil organic matter content, soil porosity, soil type, soil microbial abundance, soil humidity.
[0042] As Figure 2 shown, the data is stored in the database and stored in time series: t0, t1, t2,... t21, where the data of tn (0 < n ≤ 21) contains material properties data and spatio-temporal data.
[0043] Further, the process of separately encoding the material properties data and spatio-temporal data includes cleaning the data, removing abnormal data, using padding to replace the missing data with 0; normalizing all numerical values; and performing cosine encoding on the time series data.
[0044] As Figure 3 shown, after the material properties data and spatio-temporal data are separately encoded, a similarity matrix is obtained through global similarity calculation, and then it is skip-connected with the original data and sent to the decoder to extract features respectively; two separate encoders are used in this process, and the encoding process is expressed as:
[0045] f' n = Encoder f (f n )
[0046] s' n =Encoder s (s n )
[0047] Among them, f n and n Respectively represent the material property data and spatiotemporal data in time period n, Encoder f and Encoder s There are two encoders;
[0048] The global similarity is calculated by calculating the dot product of the encoded material property data and the spatiotemporal data. The global similarity is expressed as:
[0049] G = softmax(f' n ·s' n T )
[0050] The original data is then skip-connected with the similarity matrix and then sent to the decoder for feature extraction. The formula is expressed as:
[0051] t n =Decoder([f' n ,s' n ,G])
[0052] In the formula, [f' n ,s' n ,G] means connecting the original data and the similarity matrix, and Decoder() is a decoder.
[0053] Furthermore, the space-time high-order derivative continuity constraints are specifically:
[0054] T(x,y,n)=t n , where (x, y) is the spatial position of the material in space and time, and n is the time period of data collection;
[0055] T'(x,y,n)=T'(x,y,n+1)
[0056] T”(x,y,n)=T”(x,y,n+1)
[0057] Where: T' is the first-order derivative of T, T" is the second-order derivative of T; T' and T" are continuous.
[0058] like Figure 4 As shown, the prediction network uses a stacked LSTM (Long Short-Term Memory) network to capture the temporal global correlation in the periodic sequence from t0 to t21. For each time period tn, the input includes physicochemical properties and environmental spatiotemporal data. The encoder can be expressed as:
[0059] h n =LSTM(t n ,h n-1 )
[0060] Where, t n is the input feature vector at time n, h n is the hidden state of the LSTM cell.
[0061] Then a multi-head self-attention module is constructed to learn the global features between time series data and extract the most concerned local areas in the features.
[0062] MultiHead(Q,K,V)=Concat(head1,…,head n )W
[0063] head n =Attention(QW n Q ,KW n K ,VW n V )
[0064]
[0065] Among them, W n Q , W n K , VW n V , and W is the learnable parameter matrix. is the dimension of the key vector. Finally, the degradation rate η of each stage is predicted by the prediction head.
[0066] Furthermore, after obtaining the degradation rate at each time period stage, the method also includes fitting a degradation rate curve using a polynomial, obtaining the degradation cycle time by setting a degradation rate failure threshold, and subsequent experiments change the geographical area and the scope of time series coverage to perform iterative optimization and improve the model prediction accuracy.
[0067] Furthermore, the root mean square error (MSE) and mean absolute error (MAE) between the predicted value and the true value are used to characterize the goodness of fit of the model.
[0068]
[0069] In the formula, is the actual degradation rate, η n is the predicted degradation rate, and N is the number of samples.
[0070] The hyperparameters of the above model are as follows: 8 samples are trained in each batch, and the total number of iterations is 300, that is, batchsize=8,epoch=300; the loss function indicator is mean square error MSE; the optimizer is AdamW; the optimizer step size is set to 0.001, that is, learningrate=0.001.
Claims
1. A method for predicting the degradation life of biodegradable mulch applicable to complex time and space, characterized in that: The following steps are involved: Collect material property data and spatiotemporal data that affect the degradation life of biodegradable mulch in different regions, classify them according to time periods, and obtain time series data; After the material property data and spatiotemporal data are encoded separately, the similarity matrix is obtained through global similarity calculation, and then jump-connected with the original data and sent to the decoder to extract features respectively. After fusion, spatiotemporal high-order derivative continuity constraints are performed to obtain enhanced data; The enhanced data is used as the input of the prediction network. The prediction network adopts LSTM combined with a multi-head self-attention method. A stacked LSTM network is used in the encoder to capture the temporal global correlation in different time period sequences. For each time period, the enhanced data of material property data and spatiotemporal data are input, and then a multi-head self-attention module is constructed to learn the global features between time series data and extract the local feature factors that need the most attention in the features. Finally, the degradation rate of each time period is predicted by the prediction head.
2. The method for predicting the degradation life of biodegradable mulch applicable to complex time and space according to claim 1, characterized in that: The material property data are sampled at fixed time intervals starting from the start time when the biodegradable mulch is applied in the field environment, and include: sample mass, mechanical properties, thermodynamic properties, rheology, melt index, thickness, light transmittance and UV shielding properties, water vapor and oxygen permeability, number of surface cracks and fissures, and degree of breakage.
3. The method for predicting the degradation life of biodegradable mulch applicable to complex time and space according to claim 1, characterized in that: The spatiotemporal data is obtained by using agricultural Internet of Things devices to obtain environmental information of the region, and the average, maximum and minimum values of environmental factors in each cycle are statistically calculated. The environmental factors include: temperature, humidity, radiation intensity, rainfall, wind speed, soil pH value, soil organic matter content, soil porosity, soil type, soil microbial abundance, and soil humidity.
4. The method for predicting the degradation life of biodegradable mulch applicable to complex time and space according to claim 1, characterized in that: The process of separately encoding the material characteristic data and the spatiotemporal data includes cleaning the data, removing abnormal data, replacing missing data with 0 using padding, and normalizing all values; Cosine encode time series data.
5. The method for predicting the degradation life of biodegradable mulch applicable to complex time and space according to claim 1, characterized in that: After the material characteristic data and the spatiotemporal data are encoded respectively, a similarity matrix is obtained by global similarity calculation, and then jump-connected with the original data and sent to the decoder to extract features respectively; This process uses two separate encoders and the encoding process is expressed as: f' n =Encoder f (f n ) s' n =Encoder s (s n ) Among them, f n and n Respectively represent the material property data and spatiotemporal data in time period n, Encoder f and Encoder s There are two encoders; The global similarity is calculated by calculating the dot product of the encoded material property data and the spatiotemporal data. The global similarity is expressed as: G=softmax(f' n ·s' n T ) The original data is then skip-connected with the similarity matrix and then sent to the decoder for feature extraction. The formula is expressed as: t n =Decoder([f' n ,s' n ,G]) In the formula, [f' n ,s' n ,G] means connecting the original data and the similarity matrix, and Decoder() is a decoder.
6. The method for predicting the degradation life of biodegradable mulch applicable to complex time and space according to claim 1, characterized in that: The specific constraints of space-time high-order derivative continuity are: T(x,y,n)=t n , where (x, y) is the spatial position of the material in space and time, and n is the time period of data collection; T'(x,y,n)=T'(x,y,n+1) T”(x,y,n)=T”(x,y,n+1) Where: T' is the first-order derivative of T, T" is the second-order derivative of T; T' and T" are continuous.
7. The method for predicting the degradation life of biodegradable mulch applicable to complex time and space according to claim 1, characterized in that: After obtaining the degradation rate of each time period stage, the method also includes fitting a degradation rate curve using a polynomial, obtaining the degradation cycle time by setting a degradation rate failure threshold, and performing iterative optimization by changing the geographical area and the scope of time series coverage in subsequent experiments.
8. The method for predicting the degradation life of biodegradable mulch applicable to complex time and space according to claim 1, characterized in that: The root mean square error (MSE) and mean absolute error (MAE) between the predicted value and the true value are used to characterize the goodness of fit of the model. In the formula, is the actual degradation rate, η n is the predicted degradation rate, and N is the number of samples.
9. A computing device, characterized in that include: One or more processors, one or more memories, and one or more programs, wherein the one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing any of the methods according to claims 1 to 8.
10. A computer-readable storage medium storing one or more programs, characterized in that: The one or more programs include instructions which, when executed by a computing device, cause the computing device to perform any one of the methods according to claims 1 to 8.