Method for predicting environmental tolerance of anti-skid slow-release nutrition ecological rod

Abstract

The invention provides a method for predicting the environmental tolerance of an anti-skid slow-release nutritional ecological rod, and particularly relates to the field of environmental tolerance prediction. The method comprises the following steps: firstly, collecting environmental tolerance data of the anti-skid slow-release nutrition ecological rod, and preprocessing the data; secondly, providing an environment tolerance performance prediction model of the anti-skid slow-release nutrition ecological rod, expanding a data set by the model through a data intensifier module, and generating a false sample similar to original data distribution based on a small amount of data; the environment feature encoder extracts potential features in the input data and maps the potential features to a high-dimensional space to represent performance indexes of the ecological rod in different environments. The features are transmitted to a performance prediction decoder, and the performance prediction decoder ensures that the generated output only depends on the generated part through a causal self-attention mechanism, so that future information leakage is avoided. Meanwhile, the context is in butt joint with the attention mechanism to dynamically adjust the generated content, the correlation between input and output is enhanced, and the accuracy and correlation of prediction are improved.

Description

Technical Field

[0001] The invention belongs to the field of tolerance performance prediction, and in particular relates to a method for predicting the environmental tolerance performance of an anti-slip slow-release nutritional ecological bar. Background Art

[0002] As a multifunctional ecological material that combines anti-skid and slow-release nutrition supply, anti-skid slow-release nutrition ecological sticks are widely used in agriculture, gardening, urban greening and environmental protection. Its main functions include providing a continuous supply of plant nutrients under different environmental conditions, while enhancing the anti-skid performance of the ground and reducing the occurrence of slip accidents through optimized surface design and material selection. In addition, the slow-release nutrient part uses biodegradable polymers such as polylactic acid (PLA) or polycaprolactone (PCL) to encapsulate the nutrient components, and the gradual release of nutrients is achieved by controlling the pore structure and chemical reaction rate of the coating. This design not only improves the efficiency of nutrient utilization and reduces the frequency of fertilization, but also extends the service life of the ecological stick, which meets the sustainable development needs of modern agriculture and environmental protection.

[0003] Environmental tolerance refers to the ability of anti-slip slow-release nutrient eco-bars to maintain their functions, structures and performances under various environmental conditions. Specifically, environmental tolerance includes anti-slip performance, stability of slow-release nutrients, integrity of physical structure, degradation rate and residual amount of nutrients. The factors affecting environmental tolerance mainly include temperature, humidity, light intensity, precipitation, wind speed and direction, soil pH, soil moisture, organic matter content and microbial activity. These environmental factors not only directly affect the material degradation rate and nutrient release rate of eco-bars, but also indirectly affect the overall performance of eco-bars by changing soil properties and microbial activity. Methods for evaluating environmental tolerance include laboratory accelerated aging tests, long-term on-site monitoring, statistical analysis, and modern machine learning and data-driven models. Laboratory tests can accurately simulate performance changes under specific environmental conditions, while field tests provide long-term data in real environments. Combining statistical analysis and machine learning methods, we can deeply understand the complex relationship between environmental factors and eco-bar performance and provide a scientific basis for performance prediction.

[0004] Existing environmental tolerance performance prediction methods mainly include statistical analysis methods based on laboratory and field data, as well as data-driven machine learning models that have emerged in recent years. Traditional statistical analysis methods rely on a large amount of experimental and field test data to establish relationship models between environmental factors and performance indicators through regression analysis and other means. However, these methods have certain limitations when dealing with complex nonlinear relationships and large-scale data. In contrast, machine learning methods such as neural networks can more effectively capture the complex nonlinear relationship between environmental factors and eco-stick performance, and improve the accuracy and robustness of prediction. Combining multi-source data integration, feature engineering and advanced algorithm optimization, a comprehensive prediction model suitable for eco-stick environmental tolerance performance is developed. In addition, integrating Internet of Things technology to achieve real-time data collection and dynamic prediction has also become an important direction for future development. With the improvement of sensor technology and data processing capabilities, environmental tolerance performance prediction methods will become more intelligent and automated, and can provide real-time support for the design optimization, quality control and use management of eco-sticks. This type of prediction method can not only enhance the market competitiveness of eco-stick products, but also promote their wide application in more fields. Summary of the invention

[0005] The main purpose of the present invention is to provide a method for predicting the environmental tolerance performance of an anti-slip slow-release nutritional ecological bar, aiming to construct a prediction model. The environmental tolerance performance prediction model of the anti-slip slow-release nutritional ecological bar is composed of a data enhancer module, an environmental feature encoder and a performance prediction decoder. The data enhancer module generates false samples with a distribution similar to that of a small amount of environmental tolerance data of anti-slip slow-release nutritional ecological bars by learning the environmental tolerance data of the anti-slip slow-release nutritional ecological bars, thereby achieving the purpose of expanding the data set. The environmental feature encoder extracts potential features from the input environmental data and maps the potential features to a high-dimensional representation space. The features represent the performance indicators of the ecological bar under different environmental conditions, and the features are passed to the performance prediction decoder for further prediction. The performance prediction decoder ensures that the output generated each time depends only on the generated part through the causal self-attention mechanism to avoid future information leakage. The generated content is dynamically adjusted through the output through the context docking attention, and the association between the input and the output is strengthened, thereby improving the accuracy and relevance of the generated results.

[0006] To achieve the above object, the technical solution of the present invention is: a method for predicting the environmental tolerance performance of an anti-slip slow-release nutritional ecological stick, the method comprising: S1. Collecting environmental tolerance data of anti-skid slow-release nutrient ecological sticks, the environmental tolerance data includes: environmental temperature, humidity, light intensity, rainfall, wind speed and direction, soil pH value, humidity, and also collecting anti-skid performance indicators of ecological sticks in the label column; S2. Preprocess the collected eco-stick environmental tolerance data to ensure that the collected raw data has high quality and is suitable for subsequent machine learning model training after systematic processing; S3. A prediction model for the environmental tolerance performance of anti-slip slow-release nutritional ecological sticks is proposed. The specific method includes: S31, constructing a data enhancer module, the data enhancer is composed of a generation structure and a discrimination structure, and the data enhancer module generates false samples with similar distribution to the environmental tolerance data by learning a small amount of environmental tolerance data of anti-slip slow-release nutritional ecological bars, thereby achieving the purpose of expanding the data set; S32, constructing an environmental feature encoder, extracting potential features from the input environmental data, and mapping the potential features to a high-dimensional representation space, wherein the features represent the performance indicators of the eco-stick under different environmental conditions, and passing the features to a performance prediction decoder for further prediction; S33. Construct a performance prediction decoder, which uses the causal self-attention mechanism to ensure that each generated output only depends on the generated part, avoiding future information leakage. It also dynamically adjusts the generated content through the output through the context docking attention, strengthens the association between input and output, and thus improves the accuracy and relevance of the generated results. S4, training the environmental tolerance performance prediction model of the anti-slip slow-release nutritional ecological stick, wherein the training is to train and optimize the model so that the prediction model can achieve better prediction results on different tasks and data; S5. Test the environmental tolerance performance prediction model of the anti-slip slow-release nutritional ecological stick to evaluate its optimization effect and generalization ability, and determine whether the prediction can be used in actual scenarios.

[0007] Furthermore, in step S31, the generation structure in the data enhancer generates a time series similar to the real anti-slip slow-release nutritional ecological stick environmental tolerance time series data from the noise, and adds the time step information t time coding realized by position coding and sine and cosine functions to the noise data. The calculation formula of the noise data is as follows: z t =concat(z input ,sin(t / 10000 2i / d ), cos(t / 10000 2i+1 / d )); In the formula, z t is the final noise data, z input is a random noise vector sampled from a standard normal distribution, with dimension z of size z dim, t is the time step, sin and cos functions are periodic codes for generating time steps, which are used to represent the position information in time, d is the dimension of the input vector, and concat(·) is a concatenation operation. In this way, the noise not only carries random information, but also carries periodic changes related to the time step, which can help generate periodic and trend characteristics in structural learning time series data. Subsequently, the generative structure models the global and local features of the data at different levels, and finally outputs more realistic time series data. Multi-level modeling gradually approaches the real time series by dividing the generation process into multiple stages, each stage generating different levels of information; Multi-level modeling first generates a rough time series, that is, the global trend, and then gradually refines the output on this basis, that is, local fluctuations and detailed information; In the first layer, the generative structure first generates a rough time series, the main goal of which is to capture the global trend and long-term dependence, by z t Input into the LSTM network, the specific process is as follows: In the formula, c t is the external condition information, is the anti-slip performance index of the eco-stick, To generate the first layer of LSTM network in the structure, is the output of the first layer of LSTM, which is a rough time series that captures the global trend and long-term dependency in the original data set, but the first layer ignores small-scale short-term fluctuations and local details; the second layer generates a rough time series based on the first layer, and the generated structure begins to refine the rough time series data, gradually adding finer-grained local fluctuations and detail information. The implementation method is realized through a layer of LSTM network, adding more local fluctuations and detail information to the generated structure. The implementation method is as follows: In the formula, To generate the second layer LSTM network in the structure, is the output of the second layer LSTM, which is the refined time series output, containing more local fluctuations and detailed information; Finally, the output of each layer is weighted and combined by weighted averaging, so that the model can adjust the influence of outputs at different levels according to the importance of each layer. The implementation process is as follows: In the formula, α1 and α2 are learnable weights, indicating the contribution of each layer output, and α1+α2=1. To generate the final output of the structure.

[0008] Furthermore, in step S31, the discriminant structure in the data enhancer determines whether the input time series data is from the original real data distribution or the false data output by the generated structure. The discriminative structure introduces multi-layer LSTM, self-attention mechanism, multi-scale modeling and residual connection method to better capture the complex local fluctuations and nonlinear relationships of time series; First, the fake data generated from the generated structure Input into LSTM, the implementation process is as follows: In the formula, LSTM1 is the first layer of LSTM network, is the output of the first layer LSTM; using the skip connection method, the false data Passed to the second layer, the implementation is as follows: In the formula, is the output of the second layer LSTM, LSTM2 is the second layer LSTM network; Subsequently, the self-attention mechanism is used to calculate the correlation with other time steps to dynamically adjust its importance, thereby capturing more complex local dependencies. The calculation formula is as follows: In the formula, Q t is the query matrix, i.e., the LSTM output at the current time step, is the transpose operation of the key matrix, representing the potential information of other time steps in the time series, V t is a value matrix, representing the output information of the current time step; Finally, multi-scale modeling is used as the output of the final discriminant structure. Multi-scale modeling converts the output of the self-attention mechanism into Attention t As input, high-frequency and low-frequency data are calculated. High-frequency data is completed using a high-pass filter, and low-frequency data is completed using a low-pass filter. The calculation formula is as follows: In the formula, For high frequency data, is low-frequency data, High_frequency(·) is the high-pass filter calculation process, Low_frequency(·) is the low-pass filter calculation process, and then the high-frequency and low-frequency data are fused using the LSTM network. The fusion process is as follows: In the formula, The final fusion result of low-frequency and high-frequency data is obtained by passing through a layer normalization layer to obtain the final output of the discriminant structure. The layer normalization calculation is as follows: In the formula, It is the final output result of the discriminant structure.

[0009] Furthermore, in step S31, position encoding enables the generative structure to perceive the position information of the time step, so as to better understand the periodicity and time dependence of long time series data; the multi-layer LSTM in the generative structure enables the generative structure to capture more complex time dependencies, multi-level features from local to global, and the conditional LSTM enables the generative structure to generate time series of different patterns according to external conditions, thereby enhancing the flexibility and adaptability of the generative structure; the multi-layer LSTM and jump connections in the discriminant structure improve the flow of information and avoid the problem of gradient disappearance, the self-attention mechanism enables the discriminant structure to dynamically focus on key time steps and enhance sensitivity to short-term fluctuations, and the multi-scale modeling allows the discriminant structure to process high-frequency and low-frequency features at the same time and capture local fluctuations at different time scales; finally, through alternating training of the generative structure and the discriminant structure, the generative structure finally generates false data with a distribution similar to that of the original sample data, thereby achieving the purpose of expanding the data set.

[0010] Furthermore, in step S32, the environment feature encoder is stacked by multiple layers, each layer comprising key modules: input representation, context association module, feature enhancement module; First, the input representation is to project each feature data in the sample into a high-dimensional space and retain the association between each feature. For each input feature data x i，k , we will get an input representation E(x i，k ), where x i，j is the kth feature data point in the i-th sample, and a multi-layer input representation is used. The input representation of each layer represents the features at different levels. The multi-layer input representation formula is as follows: E(x i，k )=[E1(x i，1 ); E2(x i，2 );...;E k (x i，k )]; In the formula, E k (x i，k ) is the input representation of the kth data in the i-th sample data, E(x i，k) is the multi-layer input representation of the i-th sample, which is a matrix; then, a positioning code is added to enable the model to locate the position of each feature point in the entire sample. The positioning code is completed by absolute positioning and relative positioning. Absolute positioning is generated using sine and cosine functions. Absolute positioning is periodic. The absolute positioning calculation formula is as follows: In the formula, k is the position index, that is, the kth sample point in the sample, r is the dimension index of the embedding space, indicating the dimension in the position encoding vector, and d embed is the dimension, that is, the dimension of position positioning. The even dimension and the odd dimension are performed alternately, and finally form the absolute positioning representation P abs (k); Relative positioning can capture the relative distance between data features and is calculated based on the embedding of position differences. The position difference is expressed as Δp = k a -k b , the relative positioning calculation formula is as follows: P rel (k a , k b )=Embedding(Δp=k a -k b ); In the formula, k a and k b are two characteristic data points in a sample, Δp = k a -k b is the relative position difference between two data points, Embedding(·) is the embedding layer representation, P rel (k a , k b ) is k a and k b The final relative positioning; Finally, the relative positioning and absolute positioning are fused, and the fusion method adopts the weighted average fusion method. The fusion formula is as follows: P fusion (i, k) = α1·P abs (k)+α2·P rel (k a -k b ); Where α1 and α2 are hyperparameters, and α1+α2=1, P fusion (i, k) is the final positioning code value; The final input representation calculation formula is as follows: input k =E(x i，k )+P fusion (k); In the formula, input kis the input representation of the kth feature data point in a sample.

[0011] Furthermore, in step S32, the core of the context association module is a dynamic self-attention mechanism, which pays attention to the context information of other positions at each position of the input sequence and adjusts the weight according to the context information. The dynamic self-attention mechanism is implemented by introducing a learnable attention bias; For each feature data point in each sample, there is a corresponding query matrix Q k , key matrix K k , value matrix V k , Q k , K k 、V k The origin of is as follows: Q k =W Q ·Input k ; K k =W k ·Input k ; V k =W V ·Input k ; Where, is the learned weight matrix; For each query matrix Q k and key matrix K c The similarity between them uses a learnable dynamic bias Δ kc To achieve, it depends on the query Q k and key matrix K c The calculation formula is as follows: Δ kc =W Δ tanh(W1·Q k +W2·K c + b); Where W1 and W2 are the weights of the query matrix and the key matrix, b is the bias term, tanh(·) is the activation function used to enhance nonlinear characteristics, and W Δ is the learnable weight matrix that controls the learning of the bias term, Δ kc is the dynamic bias calculation result; the bias term Δ kc Added to the standard self-attention calculation, dynamically adjust the attention distribution between each feature data point. The core formula of the dynamic self-attention mechanism introduces the bias term into the standard self-attention calculation. The formula is as follows: A kc =Attention(Q k , Kc , V c )+Δ kc ; In the formula, A kc is the attention score after adding the dynamic bias, Attention(Q k , K c , V c ) is the standard self-attention calculation; by normalizing all attention weights, we get the weighted representation Z of each feature data point; finally, the feature enhancement module uses a feedforward neural network to further enhance the feature representation capability. The feature enhancement module formula is as follows: FFN(Z)=max(0,Z·W3+b3)·W4+b4; Where W3 and W4 are weight matrices, b3 and b4 are bias terms, and FFN(Z) is the output value of the final environmental feature encoder.

[0012] Furthermore, in step S32, through the combination of embedding and positioning coding, the environmental feature encoder can effectively understand the position of each feature data point in the sample in the sequence. At the same time, relative and absolute positioning provides spatial and sequential information of the context; the dynamic self-attention mechanism enables the model to adaptively adjust the attention allocation according to the input data, and more accurately capture the complex dependencies between environmental features; finally, the feature enhancement module further refines and optimizes the features so that the output of each layer is more in line with the requirements of the prediction task.

[0013] Furthermore, in step S33, the core of the performance prediction decoder is the causal self-attention mechanism and the contextual docking attention; First, the causal matrix M is introduced into the causal self-attention mechanism to control which keys the similarity between each query can be calculated. If query e can only see the information of key f, that is, e≤f, then the attention score matrix B ef If the query word e cannot see the information of the key word f, that is, e>f, then B ef is set to negative infinity, which ensures that it becomes 0 after applying Softmax; therefore, the causal matrix M can be expressed as: Then, the causal matrix M is added to the standard attention mechanism to obtain the causal attention score matrix B ef , and introduce the attention span parameter, the formula is as follows: In the formula, Q e is the query matrix of e, is the transpose of the key matrix of f, V f is the value matrix of f, is the scaling constant, Δspan e The adaptive attention span parameters learned by the network control the attention span of the feature data points in each sample; through this process, the performance prediction decoder can dynamically adjust the representation of each feature data point based on the currently generated content and context information, without relying on the future; The calculation process of contextual docking attention uses the key matrix K from the environment feature encoder k and the query matrix Q of the performance prediction decoder e , perform a dot product operation on it to obtain the similarity between each feature data point and the output of the environmental feature encoder. The calculation formula is as follows: Where V e is the value matrix of e, is the scaling constant; O ek The output representation of the current feature data point of the performance prediction decoder represents the representation generated by the context information provided by the performance prediction decoder; Finally, the output of the causal self-attention mechanism and the contextual docking attention are weighted and fused for output. The formula is as follows: O final =λ3B ef +λ4O ek ; In the formula, λ3 and λ4 are weight coefficients, O final is the causal self-attention mechanism B ef Attention O ek The weight output of .

[0014] Furthermore, in step S33, the causal self-attention mechanism and the contextual docking attention effectively ensure the causality of the generation process and strengthen the association between input and output; the causal self-attention mechanism ensures that the value generated each time depends only on the generated part, avoiding future information leakage, thereby maintaining the autoregressive generation order; the contextual docking attention enables the performance prediction decoder to dynamically pay attention to the relevant information of the performance prediction decoder output, thereby generating a more accurate output based on the input features; the combination of these two mechanisms enables the performance prediction decoder to maintain sequential dependency during the calculation process, and to make full use of the global context of the input, thereby improving the accuracy and relevance of the generation results.

[0015] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: In the present invention, a prediction model for the environmental tolerance performance of anti-slip slow-release nutritional ecological bars is proposed, and the prediction model for the environmental tolerance performance of anti-slip slow-release nutritional ecological bars is composed of a data enhancer module, an environmental feature encoder and a performance prediction decoder; wherein, the data enhancer module generates false samples with a distribution similar to the environmental tolerance data by learning a small amount of environmental tolerance data of anti-slip slow-release nutritional ecological bars, thereby achieving the purpose of expanding the data set; the environmental feature encoder extracts potential features from the input environmental data and maps the potential features to a high-dimensional representation space; the features represent the performance indicators of the ecological bars under different environmental conditions, and the features are passed to the performance prediction decoder for further prediction; the performance prediction decoder ensures that the output generated each time depends only on the generated part through the causal self-attention mechanism, thereby avoiding future information leakage; the generated content is dynamically adjusted through the output through the context docking attention, and the association between the input and the output is strengthened, thereby improving the accuracy and relevance of the generated results. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The present invention is a flowchart of the steps of a method for predicting the environmental tolerance performance of an anti-slip slow-release nutritional ecological stick.

[0017] Figure 2 Flow chart of the steps of the model for predicting the environmental tolerance performance of anti-slip slow-release nutritional ecological bars.

[0018] Figure 3 This is the structural diagram of the data enhancer module in the environmental tolerance performance prediction model of anti-slip slow-release nutritional ecological sticks.

[0019] Figure 4 This is the flow chart of the data enhancer module in the environmental tolerance performance prediction model of anti-slip slow-release nutritional ecological sticks.

[0020] Figure 5 This is the structural diagram of the environmental characteristic encoder in the environmental tolerance performance prediction model of anti-slip slow-release nutritional ecological sticks.

[0021] Figure 6 This is the structure diagram of the performance prediction decoder in the environmental tolerance performance prediction model of anti-slip slow-release nutritional ecological sticks.

[0022] Figure 7 This is a comparison chart of the predicted values ​​and true values ​​of the environmental tolerance performance prediction model for anti-slip slow-release nutritional ecological bars.

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0024] See also Figure 1-Figure 7 The present invention provides a technical solution: a method for predicting the environmental tolerance performance of an anti-slip slow-release nutritional ecological bar, the method comprising the steps of collecting environmental tolerance data of the anti-slip slow-release nutritional ecological bar, preprocessing the collected environmental tolerance data of the ecological bar, proposing an anti-slip slow-release nutritional ecological bar environmental tolerance performance prediction model, training the anti-slip slow-release nutritional ecological bar environmental tolerance performance prediction model, and testing the anti-slip slow-release nutritional ecological bar environmental tolerance performance prediction model.

[0025] Please refer to Figure 1 As shown, a method for predicting the environmental tolerance performance of an anti-slip slow-release nutritional ecological stick in an embodiment of the present application comprises the following specific steps:

[0026] S1. Collect environmental tolerance data of anti-skid slow-release nutrient ecological sticks, including: environmental temperature, humidity, light intensity, rainfall, wind speed and direction, soil pH value, humidity, and also collect anti-skid performance indicators of labeled ecological sticks.

[0027] Furthermore, in step S1, data is collected by deploying a variety of high-precision sensors at the installation site of the eco-stick to monitor the temperature, humidity, light intensity, rainfall, wind speed and wind direction environmental parameters in real time; at the same time, the pH value and humidity of the soil are measured by regularly collecting soil samples and using professional equipment; temperature and humidity sensors, light sensors, rain gauges, anemometers and wind vane equipment record data hourly, and transmit the data wirelessly to a central database for centralized storage and management; the soil pH value is measured once a month, and the soil humidity is recorded hourly.

[0028] S2. Preprocess the collected eco-stick environmental tolerance data to ensure that the collected raw data has high quality and is suitable for subsequent machine learning model training after systematic processing.

[0029] Furthermore, in step S2, data cleaning is first performed, including removing duplicate records and ensuring that the formats and units of all data fields are consistent to eliminate potential biases and errors; then, missing values ​​are processed by identifying the missing parts in the data, deleting samples with serious missing values ​​and using the mean method for interpolation to ensure data integrity; then, anomaly detection and processing are performed, and outliers are identified and deleted using box plots to prevent them from having a negative impact on model training; finally, data conversion is performed, and data of different scales are unified to the same range through standardization to improve the efficiency and accuracy of model training; all collected environmental data are uniformly formatted, and undergo data cleaning and quality control to ensure their accuracy and completeness. These systematic data collection methods can not only comprehensively reflect the dynamic changes in the environment in which the eco-stick is located, but also provide reliable basic data support for subsequent environmental tolerance performance predictions.

[0030] A prediction model for the environmental tolerance performance of anti-slip slow-release nutritional ecological sticks was proposed. Figure 2 As shown, the specific method includes: constructing a data enhancer module, constructing an environment feature encoder, and constructing a performance prediction decoder.

[0031] S31, build a data enhancer module, the data enhancer consists of a generation structure and a discrimination structure, such as Figure 3 As shown, the data enhancer module generates false samples with a distribution similar to that of a small amount of environmental tolerance data of anti-slip slow-release nutritional ecological bars by learning from the environmental tolerance data, thereby achieving the purpose of expanding the data set.

[0032] Furthermore, in step S31, the generation structure in the data enhancer generates a time series similar to the real anti-slip slow-release nutritional ecological stick environmental tolerance time series data from the noise, and adds the time step information t time coding realized by position coding and sine and cosine functions to the noise data. The calculation formula of the noise data is as follows: z t =concat(z input ,sin(t / 10000 2i / d ), cos(t / 10000 2i+1 / d )); In the formula, z t is the final noise data, z input is a random noise vector sampled from a standard normal distribution, with dimension z of size z dim, t is the time step, sin and cos functions are periodic codes for generating time steps, which are used to represent the position information in time, d is the dimension of the input vector, and concat(·) is a concatenation operation. In this way, the noise not only carries random information, but also carries periodic changes related to the time step, which can help generate periodic and trend characteristics in structural learning time series data. Subsequently, the generative structure models the global and local features of the data at different levels, and finally outputs more realistic time series data. Multi-level modeling gradually approaches the real time series by dividing the generation process into multiple stages, each stage generating different levels of information; Multi-level modeling first generates a rough time series, that is, the global trend, and then gradually refines the output on this basis, that is, local fluctuations and detailed information; In the first layer, the generative structure first generates a rough time series, the main goal of which is to capture the global trend and long-term dependence, by z t Input into the LSTM network, the specific process is as follows: In the formula, c t is the external condition information, is the anti-slip performance index of the eco-stick, To generate the first layer of LSTM network in the structure, is the output of the first layer of LSTM, which is a rough time series that captures the global trend and long-term dependency in the original data set, but the first layer ignores small-scale short-term fluctuations and local details; the second layer generates a rough time series based on the first layer, and the generated structure begins to refine the rough time series data, gradually adding finer-grained local fluctuations and detail information. The implementation method is realized through a layer of LSTM network, adding more local fluctuations and detail information to the generated structure. The implementation method is as follows: In the formula, To generate the second layer LSTM network in the structure, is the output of the second layer LSTM, which is the refined time series output, containing more local fluctuations and detailed information; Finally, the output of each layer is weighted and combined by weighted averaging, so that the model can adjust the influence of outputs at different levels according to the importance of each layer. The implementation process is as follows: In the formula, α1 and α2 are learnable weights, indicating the contribution of each layer output, and α1+α2=1. To generate the final output of the structure.

[0033] Furthermore, in step S31, the discriminant structure in the data enhancer determines whether the input time series data is from the original real data distribution or the false data output by the generated structure. The discriminative structure introduces multi-layer LSTM, self-attention mechanism, multi-scale modeling and residual connection method to better capture the complex local fluctuations and nonlinear relationships of time series, such as Figure 4 As shown; First, the fake data generated from the generated structure Input into LSTM, the implementation process is as follows: In the formula, LSTM1 is the first layer of LSTM network, is the output of the first layer LSTM; using the skip connection method, the false data Passed to the second layer, the implementation is as follows: In the formula, is the output of the second layer LSTM, LSTM2 is the second layer LSTM network; Subsequently, the self-attention mechanism is used to calculate the correlation with other time steps to dynamically adjust its importance, thereby capturing more complex local dependencies. The calculation formula is as follows: In the formula, Q t is the query matrix, i.e., the LSTM output at the current time step, is the transpose operation of the key matrix, representing the potential information of other time steps in the time series, V t is a value matrix, representing the output information of the current time step; Finally, multi-scale modeling is used as the output of the final discriminant structure. Multi-scale modeling converts the output of the self-attention mechanism into Attention t As input, high-frequency and low-frequency data are calculated. High-frequency data is completed using a high-pass filter, and low-frequency data is completed using a low-pass filter. The calculation formula is as follows: In the formula, For high frequency data, is low-frequency data, High_frequency(·) is the high-pass filter calculation process, Low_frequency(·) is the low-pass filter calculation process, and then the high-frequency and low-frequency data are fused using the LSTM network. The fusion process is as follows: In the formula, The final fusion result of low-frequency and high-frequency data is obtained by passing through a layer normalization layer to obtain the final output of the discriminant structure. The layer normalization calculation is as follows: In the formula, It is the final output result of the discriminant structure.

[0034] S32. Construct an environmental feature encoder to extract potential features from the input environmental data and map the potential features to a high-dimensional representation space. The features represent the performance indicators of the eco-stick under different environmental conditions and pass the features to a performance prediction decoder for further prediction.

[0035] Furthermore, in step S32, the environment feature encoder is stacked by multiple layers, each layer comprising key modules: input representation, context association module, feature enhancement module; First, the input representation is to project each feature data in the sample into a high-dimensional space and retain the association between each feature. For each input feature data x i，k , we will get an input representation E(x i，k ), where x i，j is the kth feature data point in the i-th sample, and a multi-layer input representation is used. The input representation of each layer represents the features at different levels. The multi-layer input representation formula is as follows: E(x i，k )=[E1(x i，1 ); E2(x i，2 );...;E k (x i，k )]; In the formula, E k (x i，k ) is the input representation of the kth data in the i-th sample data, E(x i，k ) is the multi-layer input representation of the i-th sample, which is a matrix; then, a positioning code is added to enable the model to locate the position of each feature point in the entire sample. The positioning code is completed by absolute positioning and relative positioning. Absolute positioning is generated using sine and cosine functions. Absolute positioning is periodic. The absolute positioning calculation formula is as follows: In the formula, k is the position index, that is, the kth sample point in the sample, r is the dimension index of the embedding space, indicating the dimension in the position encoding vector, and d embed is the dimension, that is, the dimension of position positioning. The even dimension and the odd dimension are performed alternately, and finally form the absolute positioning representation P abs(k); Relative positioning can capture the relative distance between data features and is calculated based on the embedding of position differences. The position difference is expressed as Δp = k a -k b , the relative positioning calculation formula is as follows: P rel (k a , k b )=Embedding(Δp=k a -k b ); In the formula, k a and k b are two characteristic data points in a sample, Δp = k a -k b is the relative position difference between two data points, Embedding(·) is the embedding layer representation, P rel (k a , k b ) is k a and k b The final relative positioning; Finally, the relative positioning and absolute positioning are fused, and the fusion method adopts the weighted average fusion method. The fusion formula is as follows: P fusion (i, k) = α1·P abs (k)+α2·P rel (k a -k b ); Where α1 and α2 are hyperparameters, and α1+α2=1, P fusion (i, k) is the final positioning code value; The final input representation calculation formula is as follows: input k =E(x i，k )+P fusion (k); In the formula, input k is the input representation of the kth feature data point in a sample.

[0036] Furthermore, in step S32, the core of the context association module is a dynamic self-attention mechanism. The dynamic self-attention mechanism pays attention to the context information of other positions at each position of the input sequence and adjusts the weight according to the context information. The dynamic self-attention mechanism is implemented by introducing a learnable attention bias, such as Figure 5 As shown; For each feature data point in each sample, there is a corresponding query matrix Q k , key matrix K k , value matrix V k , Qk , K k 、V k The origin of is as follows: Q k =W Q ·Input k ; K k =W k ·Input k ; V k =W V ·Input k ; Where, is the learned weight matrix; For each query matrix Q k and key matrix K c The similarity between them uses a learnable dynamic bias Δ kc To achieve, it depends on the query Q k and key matrix K c The calculation formula is as follows: Δ kc =W Δ tanh(W1·Q k +W2·K c + b); Where W1 and W2 are the weights of the query matrix and the key matrix, b is the bias term, tanh(·) is the activation function used to enhance nonlinear characteristics, and W Δ is the learnable weight matrix that controls the learning of the bias term, Δ kc is the dynamic bias calculation result; the bias term Δ kc Added to the standard self-attention calculation, dynamically adjust the attention distribution between each feature data point. The core formula of the dynamic self-attention mechanism introduces the bias term into the standard self-attention calculation. The formula is as follows: A kc =Attention(Q k , K c , V c )+Δ kc ; In the formula, A kc is the attention score after adding the dynamic bias, Attention(Q k , K c , V c ) is the standard self-attention calculation; by normalizing all attention weights, we get the weighted representation Z of each feature data point; finally, the feature enhancement module uses a feedforward neural network to further enhance the feature representation capability. The feature enhancement module formula is as follows: FFN(Z)=max(0,Z·W3+b3)·W4+b4; Where W3 and W4 are weight matrices, b3 and b4 are bias terms, and FFN(Z) is the output value of the final environmental feature encoder.

[0037] Furthermore, in step S33, the core of the performance prediction decoder is the causal self-attention mechanism and the contextual docking attention, such as Figure 6 As shown; First, the causal matrix M is introduced into the causal self-attention mechanism to control which keys the similarity between each query can be calculated. If query e can only see the information of key f, that is, e≤f, then the attention score matrix B ef If the query word e cannot see the information of the key word f, that is, e>f, then B ef is set to negative infinity, which ensures that it becomes 0 after applying Softmax; therefore, the causal matrix M can be expressed as: Then, the causal matrix M is added to the standard attention mechanism to obtain the causal attention score matrix B ef , and introduce the attention span parameter, the formula is as follows: In the formula, Q e is the query matrix of e, is the transpose of the key matrix of f, V f is the value matrix of f, is the scaling constant, Δspan e The adaptive attention span parameters learned by the network control the attention span of the feature data points in each sample; through this process, the performance prediction decoder can dynamically adjust the representation of each feature data point based on the currently generated content and context information, without relying on the future; The calculation process of contextual docking attention uses the key matrix K from the environment feature encoder k and the query matrix Q of the performance prediction decoder e , perform a dot product operation on it to obtain the similarity between each feature data point and the output of the environmental feature encoder. The calculation formula is as follows: Where V e is the value matrix of e, is the scaling constant; O ek The output representation of the current feature data point of the performance prediction decoder represents the representation generated by the context information provided by the performance prediction decoder; Finally, the output of the causal self-attention mechanism and the contextual docking attention are weighted and fused for output. The formula is as follows: O final =λ3B ef +λ4O ek ; In the formula, λ3 and λ4 are weight coefficients, O final is the causal self-attention mechanism B ef Attention O ek The weight output of .

[0038] S4. Training the environmental tolerance performance prediction model of the anti-slip slow-release nutritional ecological stick. The training is to train and optimize the model so that the prediction model can achieve better prediction results on different tasks and data.

[0039] Furthermore, in step S4, the process of training the environmental tolerance performance prediction model of the anti-slip slow-release nutritional eco-bar involves the reasonable selection of hyperparameters to optimize the model performance. The key hyperparameters include a learning rate of 0.01 to control the training speed; a batch size of 256; an optimizer of Adam to provide an adaptive learning rate, a Dropout rate of 0.3 to prevent overfitting; and a number of epochs of 2000 to ensure sufficient training. By appropriately adjusting these hyperparameters and combining the cross entropy loss and the evaluation indicators accuracy and F1 score, the generalization ability of the model can be improved to ensure accurate prediction of the performance of the eco-bar based on environmental characteristics.

[0040] S5. Test the environmental tolerance performance prediction model of the anti-slip slow-release nutritional ecological stick to evaluate its optimization effect and generalization ability, and determine whether the prediction can be used in actual scenarios.

[0041] Furthermore, in step S5, Figure 7 As shown, the prediction model uses the time data of the previous 30 hours to predict the environmental tolerance of the eco-bar in the next 24 hours. It can be seen from the figure that the predicted value of the prediction model is close to the real data value, that is, the prediction model can predict the environmental tolerance performance of the anti-slip slow-release nutritional eco-bar in the future.

Claims

1. A method for predicting the environmental tolerance performance of an anti-slip slow-release nutritional ecological stick, characterized in that: The following steps are involved: S1. Collecting environmental tolerance data of anti-skid slow-release nutrient ecological sticks, the environmental tolerance data includes: environmental temperature, humidity, light intensity, rainfall, wind speed and direction, soil pH value, humidity, and also collecting anti-skid performance indicators of ecological sticks in the label column; S2. Preprocess the collected eco-stick environmental tolerance data to ensure that the collected raw data has high quality and is suitable for subsequent machine learning model training after systematic processing; S3. A prediction model for the environmental tolerance performance of anti-slip slow-release nutritional ecological sticks is proposed. The specific method includes: S31, constructing a data enhancer module, the data enhancer is composed of a generation structure and a discrimination structure, and the data enhancer module generates false samples with similar distribution to the environmental tolerance data by learning a small amount of environmental tolerance data of anti-slip slow-release nutritional ecological bars, thereby expanding the data set; S32, constructing an environmental feature encoder, extracting potential features from the input environmental data, and mapping the potential features to a high-dimensional representation space, wherein the features represent the performance indicators of the eco-stick under different environmental conditions, and passing them to a performance prediction decoder for further prediction; S33. Construct a performance prediction decoder, which uses the causal self-attention mechanism to ensure that each generated output only depends on the generated part, avoiding future information leakage. It also dynamically adjusts the generated content through the output through the context docking attention, strengthens the association between input and output, and thus improves the accuracy and relevance of the generated results. S4. Train the environmental tolerance performance prediction model of the anti-slip slow-release nutritional ecological stick and optimize the model so that it can obtain good prediction results on different tasks and data sets; S5. Test the environmental tolerance performance prediction model of the anti-slip slow-release nutritional ecological stick, estimate its optimization effect and generalization ability, and ensure its effectiveness in actual application scenarios.

2. The method for predicting the environmental tolerance performance of an anti-slip slow-release nutritional ecological stick according to claim 1, characterized in that: In step S1, data is collected by deploying a variety of high-precision sensors at the eco-stick installation site to monitor temperature, humidity, light intensity, rainfall, wind speed and wind direction environmental parameters in real time; at the same time, the pH value and humidity of the soil are measured by regularly collecting soil samples and using professional equipment; temperature and humidity sensors, light sensors, rain gauges, anemometers and wind vane equipment record data hourly, and transmit the data wirelessly to a central database for centralized storage and management; soil pH is measured once a month, and soil humidity is recorded hourly.

3. The method for predicting the environmental tolerance performance of an anti-slip slow-release nutritional ecological stick according to claim 2, characterized in that: In step S31, the generating structure in the data enhancer generates a time series similar to the real anti-slip slow-release nutritional ecological stick environmental tolerance time series data from the noise, and adds the time step information t time coding realized by position coding and sine and cosine functions to the noise data. The calculation formula of the noise data is as follows: z t =concat(z input ,sin(t / 10000 2i / d ),cos(t / 10000 2i+1 / d )); In the formula, z t is the final noise data, z input is a random noise vector sampled from a standard normal distribution, with dimension z of size z dim , t is the time step, sin and cos functions are the periodic codes for generating time steps, which are used to represent the position information in time, d is the dimension of the input vector, and concat((·) is the concatenation operation. In this way, the noise not only carries random information, but also carries the periodic changes related to the time step, which can help generate periodic and trend characteristics in structural learning time series data. Subsequently, the generative structure models the global and local features of the data at different levels, and finally outputs more realistic time series data. Multi-level modeling gradually approaches the real time series by dividing the generation process into multiple stages, each stage generating different levels of information; Multi-level modeling first generates a rough time series, that is, the global trend, and then gradually refines the output on this basis, that is, local fluctuations and detailed information; In the first layer, the generative structure first generates a rough time series, the main goal of which is to capture the global trend and long-term dependence, by z t Input into the LSTM network, the specific process is as follows: In the formula, c t is the external condition information, is the anti-slip performance index of the eco-stick, To generate the first layer of LSTM network in the structure, is the output of the first layer of LSTM, which is a rough time series that captures the global trend and long-term dependencies in the original data set, but the first layer ignores small-scale short-term fluctuations and local details; The second layer generates a rough time series based on the first layer. The generated structure begins to refine the rough time series data and gradually add finer-grained local fluctuations and detail information. This is achieved through a layer of LSTM network to add more local fluctuations and detail information to the generated structure. The implementation is as follows: In the formula, To generate the second layer LSTM network in the structure, is the output of the second layer LSTM, which is the refined time series output, containing more local fluctuations and detailed information; Finally, the output of each layer is weighted and combined by weighted averaging, so that the model can adjust the influence of outputs at different levels according to the importance of each layer. The implementation process is as follows: In the formula, α1 and α2 are learnable weights, indicating the contribution of each layer output, and α1+α2=1. To generate the final output of the structure.

4. The method for predicting the environmental tolerance performance of the anti-slip slow-release nutritional ecological stick according to claim 3, characterized in that: In step S31, the discriminant structure in the data enhancer determines whether the input time series data is from the original real data distribution or the false data output by the generated structure. The discriminative structure introduces multi-layer LSTM, self-attention mechanism, multi-scale modeling and residual connection method to better capture the complex local fluctuations and nonlinear relationships of time series; First, the fake data generated from the generated structure Input into LSTM, the implementation process is as follows: In the formula, LSTM1 is the first layer of LSTM network, is the output of the first layer LSTM; using the skip connection method, the false data Passed to the second layer, the implementation is as follows: In the formula, is the output of the second layer LSTM, LSTM2 is the second layer LSTM network; Subsequently, the self-attention mechanism is used to calculate the correlation with other time steps to dynamically adjust its importance, thereby capturing more complex local dependencies. The calculation formula is as follows: In the formula, Q t is the query matrix, i.e., the LSTM output at the current time step, is the transpose operation of the key matrix, representing the potential information of other time steps in the time series, V t is a value matrix, representing the output information of the current time step; Finally, multi-scale modeling is used as the output of the final discriminant structure. Multi-scale modeling converts the output of the self-attention mechanism into Attention t As input, high-frequency and low-frequency data are calculated. High-frequency data is completed using a high-pass filter, and low-frequency data is completed using a low-pass filter. The calculation formula is as follows: In the formula, For high frequency data, is low-frequency data, High_frequency(·) is the high-pass filter calculation process, Low_frequency(·) is the low-pass filter calculation process, and then the high-frequency and low-frequency data are fused using the LSTM network. The fusion process is as follows: In the formula, The final fusion result of low-frequency and high-frequency data is obtained by passing through a layer normalization layer to obtain the final output of the discriminant structure. The layer normalization calculation is as follows: In the formula, It is the final output result of the discriminant structure.

5. The method for predicting the environmental tolerance performance of the anti-slip slow-release nutritional ecological stick according to claim 4, characterized in that: Furthermore, in step S32, the environment feature encoder is stacked by multiple layers, each layer comprising key modules: input representation, context association module, feature enhancement module; First, the input representation is to project each feature data in the sample into a high-dimensional space and retain the association between each feature. For each input feature data x i，k , we will get an input representation E(x i,k ), where x i，j is the kth feature data point in the i-th sample, and a multi-layer input representation is used. The input representation of each layer represents the features at different levels. The multi-layer input representation formula is as follows: E(x i，k )=[E1(x i，1 );E2(x i，2 );...;E k (x i,k )]; In the formula, E k (x i，k ) is the input representation of the kth data in the i-th sample data, E(x i，k ) is the multi-layer input representation of the i-th sample, which is a matrix; then, a positioning code is added to enable the model to locate the position of each feature point in the entire sample. The positioning code is completed by absolute positioning and relative positioning. Absolute positioning is generated using sine and cosine functions. Absolute positioning is periodic. The absolute positioning calculation formula is as follows: In the formula, k is the position index, that is, the kth sample point in the sample, r is the dimension index of the embedding space, indicating the dimension in the position encoding vector, and d embed is the dimension, that is, the dimension of position positioning. The even dimension and the odd dimension are performed alternately, and finally form the absolute positioning representation P abs (k); Relative positioning can capture the relative distance between data features and is calculated based on the embedding of position differences. The position difference is expressed as △p = k a -k b , the relative positioning calculation formula is as follows: P rel (k a ,k b )=Embedding(△p=k a -k b ); In the formula, k a and k b are two characteristic data points in a sample, △p=k a -k b is the relative position difference between two data points, Emnedding(·) is the embedding layer representation, P rel (k a , k b ) is k a and k b The final relative positioning; Finally, the relative positioning and absolute positioning are fused, and the fusion method adopts the weighted average fusion method. The fusion formula is as follows: P fusion (i,k)=α1·P abs (k)+α2·P rel (k a -k b ); Where α1 and α2 are hyperparameters, and α1+α2=1, P fusion (i, k) is the final positioning code value; The final input representation calculation formula is as follows: input k =E(x i，k )+P fusion (k); In the formula, input k is the input representation of the kth feature data point in a sample.

6. The method for predicting the environmental tolerance performance of the anti-slip slow-release nutritional ecological stick according to claim 5, characterized in that: Furthermore, in step S32, the core of the context association module is a dynamic self-attention mechanism, which pays attention to the context information of other positions at each position of the input sequence and adjusts the weight according to the context information. The dynamic self-attention mechanism is implemented by introducing a learnable attention bias; For each feature data point in each sample, there is a corresponding query matrix Q k , key matrix K k , value matrix V k , Q k , K k 、V k The origin of is as follows: Q k =W Q ·input k ; K k =W k ·input k ; V k =W V ·input k ; Where, is the learned weight matrix; For each query matrix Q k and key matrix K c The similarity between them uses a learnable dynamic bias △ kc To achieve, it depends on the query Q k and key matrix K c The calculation formula is as follows: △ kc =W Δ ·tanh(W1·Q k +W2·K c +b); Where W1 and W2 are the weights of the query matrix and the key matrix, b is the bias term, tanh(·) is the activation function used to enhance nonlinear characteristics, and W Δ is a learnable weight matrix that controls the learning of the bias term, △ kc is the result of dynamic bias calculation; the bias term △ kc Added to the standard self-attention calculation, dynamically adjust the attention distribution between each feature data point. The core formula of the dynamic self-attention mechanism introduces the bias term into the standard self-attention calculation. The formula is as follows: A kc =Attention(Q k ,K c ,V c )+△ kc ; In the formula, A kc is the attention score after adding the dynamic bias, Attention(Q k , K c , V c ) is the standard self-attention calculation; by normalizing all attention weights, we get the weighted representation Z of each feature data point; Finally, the feature enhancement module uses a feedforward neural network to further enhance the ability of feature representation. The feature enhancement module formula is as follows: FFN(Z)=max(0,Z·W3+b3)·W4+b4; Where W3 and W4 are weight matrices, b3 and b4 are bias terms, and FFN(Z) is the output value of the final environmental feature encoder.

7. The method for predicting the environmental tolerance performance of the anti-slip slow-release nutritional ecological stick according to claim 6, characterized in that: Furthermore, in step S33, the core of the performance prediction decoder is the causal self-attention mechanism and the contextual docking attention; First, the causal matrix M is introduced into the causal self-attention mechanism to control which keys the similarity between each query can be calculated. If query e can only see the information of key f, that is, e≤f, then the attention score matrix B ef If the query word e cannot see the information of the key word f, that is, e>f, then B ef is set to negative infinity, which ensures that it becomes 0 after applying Softmax; therefore, the causal matrix M can be expressed as: Then, the causal matrix M is added to the standard attention mechanism to obtain the causal attention score matrix B ef , and introduce the attention span parameter, the formula is as follows: In the formula, Q e is the query matrix of e, is the transpose of the key matrix of f, V f is the value matrix of f, is the scaling constant, △span e The adaptive attention span parameters learned by the network control the attention span of the feature data points in each sample; through this process, the performance prediction decoder can dynamically adjust the representation of each feature data point based on the currently generated content and context information, without relying on the future; The calculation process of contextual docking attention uses the key matrix K from the environment feature encoder k and the query matrix Q of the performance prediction decoder e , perform a dot product operation on it to obtain the similarity between each feature data point and the output of the environmental feature encoder. The calculation formula is as follows: Where V c is the value matrix of e, is the scaling constant; O ek The output representation of the current feature data point of the performance prediction decoder represents the representation generated by the context information provided by the performance prediction decoder; Finally, the output of the causal self-attention mechanism and the contextual docking attention are weighted and fused for output. The formula is as follows: O final =λ3B ef +λ4O ek ; In the formula, λ3 and λ4 are weight coefficients, O final is the causal self-attention mechanism B ef Attention O ek The weight output of .

Images