Pseudo-ginseng growth state prediction irrigation method and system based on data driving

Abstract

The invention discloses a panax notoginseng growth state prediction irrigation method and system based on data driving, and belongs to the field of electronic digital data processing. According to the method, pseudo-ginseng environment data and pseudo-ginseng growth data are collected and noise reduction is carried out, a pseudo-ginseng growth state prediction model is constructed according to a deep learning algorithm Informer-LSTM-EWMA, pseudo-ginseng irrigation demand prediction is carried out through the pseudo-ginseng growth state prediction model, then an irrigation decision is made according to a prediction result, and finally an irrigation result is monitored. Therefore, the pseudo-ginseng can obtain the optimal growth condition. The system comprises a data acquisition module, a data processing module, a model training module and a watering decision making module. The precision of irrigation is improved, and the negative effects of water resource waste and excessive irrigation on the growth of panax notoginseng are effectively avoided. A scientific, efficient and sustainable irrigation solution is provided for panax notoginseng planting, the yield and quality of panax notoginseng can be improved, and meanwhile the production cost and the environmental influence are reduced.

Description

Technical Field

[0001] The present invention relates to the field of electronic digital data processing, and particularly to a data-driven prediction irrigation method and system for the growth state of Panax notoginseng. Background Art

[0002] As a highly valued medicinal plant, the management of the growth environment and irrigation control of Panax notoginseng play a crucial role in improving yield and quality. In ideal agricultural production practices, irrigation management should be closely combined with the growth environment characteristics and morphological characteristics of Panax notoginseng to optimize its growth conditions to the greatest extent. Traditional irrigation management methods for Panax notoginseng mainly rely on the empirical judgment of growers.

[0003] However, although traditional empirical judgment can maintain the growth of crops to a certain extent, it lacks precise control over irrigation water volume, intervention timing, and irrigation methods. Moreover, traditional irrigation management lacks comprehensive analysis of historical irrigation data, Panax notoginseng growth data, and environmental parameters, resulting in the lack of a scientific basis and accurate prediction for formulating irrigation plans. Therefore, there is an urgent need for a data-driven irrigation method that can intelligently predict irrigation requirements through real-time monitoring and data analysis of the growth environment of Panax notoginseng, and automatically adjust the irrigation water volume and time, so as to achieve an accurate response to the growth requirements of Panax notoginseng. The introduction of this method will undoubtedly greatly improve the yield and quality of Panax notoginseng and promote the further development of medicinal plant cultivation technology. Summary of the Invention

[0004] In view of the problems of the prior art, the present invention provides a data-driven prediction irrigation method and system for the growth state of Panax notoginseng.

[0005] To achieve the above technical solution, the specific steps are as follows:

[0006] A data-driven prediction irrigation method for the growth state of Panax notoginseng includes the following steps:

[0007] S1. Collect Panax notoginseng environmental data and Panax notoginseng growth data;

[0008] The Panax notoginseng environmental data is collected once per hour through a data remote acquisition system based on Internet of Things technology in the Panax notoginseng planting base, and includes: soil temperature, soil humidity, soil conductivity, air temperature, and air humidity;

[0009] The Panax notoginseng growth data includes: Panax notoginseng growth data and water demand data at different growth stages; among them, the Panax notoginseng growth data includes: recording the plant height and stem diameter growth data of Panax notoginseng;

[0010] The steps for collecting the Panax notoginseng environmental data are as follows:

[0011] S1.1. Install the soil temperature, humidity and electrical conductivity sensors in two layers. The first layer is installed 10 cm underground, and the second layer is installed 20 cm underground; the air temperature and humidity sensors are installed 50 cm above the ground;

[0012] S1.2. Connect each sensor to the communication daughter board and number it. The communication daughter board transmits the data to the communication main board via Bluetooth. The communication main board collects the information collected by no more than 16 communication daughter boards. Finally, the communication main board packages the total information collected and transmits it to the cloud via 4G;

[0013] S1.3. After the on-site layout is completed, the data can be viewed in real time at the terminal;

[0014] S2. Denoise the data collected in S1;

[0015] Denoise the data using Gaussian denoising. The denoising model is as follows:

[0016]

[0017] In the formula, G(x) represents the output signal after Gaussian denoising; x represents the value of the input signal, and the input signal includes: soil temperature, soil humidity, soil electrical conductivity, air temperature and air humidity; σ represents the standard deviation of the Gaussian distribution;

[0018] S3. Build a prediction model for the growth status of Panax notoginseng based on the deep learning algorithm Informer-LSTM-EWMA, and predict the irrigation demand of Panax notoginseng by inputting the results of S2 into the prediction model for the growth status of Panax notoginseng;

[0019] The construction steps are as follows:

[0020] S3.1. Divide the denoised data into a training set, a test set and a validation set according to 7:2:1;

[0021] S3.2. Build an Informer-LSTM-EWMA model; the model is divided into three layers, and the construction steps are as follows:

[0022] S3.2.1. Build the first layer of the Informer model;

[0023] The self-attention mechanism of the Informer model is based on tuple input, namely query Q, key K and value V, and then the scaled dot product, and the expression is as follows:

[0024]

[0025] In the formula, where d represents the input dimension; represents the set of natural numbers; L Q ,LK , L V is the length of the input sequences of Q, K, and V;

[0026] Let qi, ki, and vi represent the i-th rows of Q, K, and V respectively; the attention for the m-th query is defined as kernel smoothing in the form of probability; the expression is as follows:

[0027]

[0028] In the formula, represents the probability distribution of attention; k(q i , k j ) represents the asymmetric exponential kernel j represents the row of the matrix, l represents the row of the matrix, j ∈ l;

[0029] In the Informer model, attention values are used to select important q-k pairs to describe the similarity between the n-th query and the key. q-k represents the query pairs composed of the corresponding rows in matrices Q and K; the attention value of the n-th query to the key can be expressed as a probability q(k j |q i ); considering q(k j |q i ) = 1 / L K is a uniform distribution, and the closer q is to the uniform distribution, the less important it is;

[0030] The expression for the sparsity measurement value M of the n-th query is as follows:

[0031]

[0032] In the formula, L k represents the row of the K matrix;

[0033] The above formula can be simplified by scaling to:

[0034]

[0035] Based on the proposed metric define u = c · In L Q dominant query numbers, where L Q represents the row of the Q matrix; thus, probabilistic sparse self-attention can be defined as:

[0036]

[0037] In the formula, Q represents a sparse matrix of the same size as Q and only contains the largest u queries in M, where u is controlled by the sampling factor c, and the sampling factor c should be set differently according to the actual data in different environments;

[0038] To extract the dominant features, a distillation operation is added to the Informer model. Through the distillation operation, the input dimension can be significantly reduced. The distillation operation is a serial combination of 1D convolution, ELU activation function, and max pooling;

[0039] The formula for advancing from the j-th layer to the (j + 1)-th layer is as follows:

[0040]

[0041] In the formula, [·] AB denotes the attention block, which contains multi-head ProbSparse self-attention. In Informer, j is the same as the number of encoder and decoder layers of the model;

[0042] S3.2.2. Input the result of S3.2.1 into the LSTM model to construct the second-layer LSTM model;

[0043] LSTM is a deep learning model for processing sequence data; it contains a memory cell and three gates: forget gate, input gate, and output gate; the role of these gates is to selectively control the flow of information, allowing the model to more effectively capture long-term dependencies in long sequences;

[0044] Each gate generates a state variable f t 、i t 、o t 、output unit h t and cell state C t at time t respectively, and the expressions are as follows:

[0045] f t =σ(w f [h t-1 ,x t-1 +b f )

[0046] i t =σ(w i [h t-1 ,x t-1 +b i )

[0047]

[0048] o t =σ(w o [h t-1 ,x t-1 +b o )

[0049]

[0050] h t = o t *tanh(C t )

[0051] In the formula, [w f , b f , [w i , b i , [w c , b c , [w o , b o correspond to the weight matrices and bias terms of the forget gate f t , input gate i t , cell state C t , output gate o t at time t respectively. At the beginning of training, the weight matrices and bias terms are randomly given. In the Python program, the weight matrices and bias terms can be randomly generated through the random() function. Although the weight matrices and bias terms are random at the beginning, as the model is continuously trained, the weight matrices and bias terms are iteratively updated with the training of the model. As the training of the model ends, the weight matrices and bias terms finally reach the ideal values; [h t-1 , x t-1 are the output in the hidden layer at time t - 1 and the input at time t - 1 respectively; is the input node state;

[0052] S3.2.3. Input the result of S3.2.2 into the EWMA model to construct the third-layer EWMA model;

[0053] The expression of the EWMA model is as follows:

[0054] EWMA T = α * X T + (1 - α) EWMA T

[0055] In the formula, EWMA t is the EWMA value at time t; X t is the observed value at time t; α is the smoothing parameter, (0 < α < 1), which determines the relative weight of the observed value and the previous EWMA t-1 value;

[0056] S3.2.4. Normalize the output of S3.2.3 so that the output result is within the preset range; the preset range is: [0, 1]

[0057] Specifically, when making predictions, the historical data of soil temperature, soil humidity, soil conductivity, air temperature, air humidity, as well as the plant height and stem diameter data of Panax notoginseng are used as the input of the model, and the future growth state (plant height) data of Panax notoginseng are used as the output; during the experiment, only the growth state of Panax notoginseng needs to be known, and the accurate value of the plant height morphological data of Panax notoginseng does not need to be known; therefore, when making predictions, the output data is normalized so that the output result is within the range of [0, 1], and the expression is as follows:

[0058]

[0059] In the formula, max is the maximum value in the feature data; min is the minimum value in the feature data; x' is the data before normalization, and x* is the data after normalization;

[0060] S3.2.5. Optimize the model parameters through the cross-validation method to obtain the optimal values, and substitute the optimal values into the model to obtain the optimal prediction model;

[0061] S3.3. Use the training data set to train the Informer-LSTM-EWMA algorithm model to obtain a training model, use the test data set to test the trained model, and finally use the validation set to validate the model;

[0062] S3.4. Take the collected data as the input and input it into the validated model, and then irrigate according to the predicted growth state of Panax notoginseng; at the same time, determine the irrigation amount according to the humidity data collected by the current sensor;

[0063] S4. Make irrigation decisions according to the prediction results;

[0064] Combined with the growth requirements and irrigation principles of Panax notoginseng, make scientific and reasonable irrigation decisions;

[0065] The irrigation decision includes: linearly fitting the intervention timing, the irrigation amount of the soil, and the soil humidity;

[0066] The steps to make irrigation decisions are as follows:

[0067] S4.1. According to the results predicted by the model, the intervention timing T itt , the intervention timing T itt The expression is as follows:

[0068] T itt = index(max([y 1 , y 2 , …y k ))

[0069] Wherein, [y1, y2,..., yk] represents the model prediction result, and k is a preset value;

[0070] S4.2. Perform linear fitting on the irrigation amount of the soil and the soil humidity to obtain the relationship between the irrigation amount of the soil and the soil humidity;

[0071] Specifically, let the humidity of the sensor before irrigation be x 1 , irrigate the soil with x 2 liters of volumetric water, wait for a period of time, and record the soil humidity x 3 after the water infiltrates downward; the expression is as follows:

[0072] x 3 = sx 2 + x 1

[0073] Wherein, s represents the water irrigation coefficient of the current soil;

[0074] In addition, it should be noted that due to the differences in soil texture, altitude, and other various environmental factors in the planting area, the s values under various environmental conditions will also change accordingly;

[0075] S4.3. Combine the prediction result of the model, the intervention timing, and the soil humidity for irrigation;

[0076] S5. Execute the irrigation and monitoring strategy.

[0077] A data-driven prediction irrigation system for the growth status of Panax notoginseng;

[0078] Including: a data acquisition module, a data processing module, a model training module, and an irrigation decision-making module;

[0079] The data acquisition module is used to perform the following steps:

[0080] S1. Collect the environmental data of Panax notoginseng and the growth data of Panax notoginseng;

[0081] The data processing module is used to perform the following steps:

[0082] S2. Perform noise reduction processing on the data;

[0083] The model training module is used to perform the following steps:

[0084] S3. Construct a prediction model for the growth status of Panax notoginseng according to the deep learning algorithm Informer-LSTM-EWMA, and predict the irrigation demand of Panax notoginseng by inputting the result of S2 into the prediction model for the growth status of Panax notoginseng;

[0085] The irrigation decision-making module is used to perform the following steps:

[0086] S4. Make irrigation decisions based on the prediction results;

[0087] S5. Implement irrigation and monitoring strategies.

[0088] Advantages of the present invention:

[0089] The method of the present invention is applicable to the planting environment of agricultural greenhouses. It makes a reasonable plan for the irrigation scheme by predicting the growth state of crops in real time, and is highly sensitive to the change of the growth trend of crops under different irrigation treatment conditions. This method adopts a data-driven irrigation regulation strategy, which can comprehensively consider the environmental characteristics and the morphological characteristics of crops, and accurately adjust the irrigation water volume, making up for the difficulty of traditional empirical irrigation in meeting the specific needs of crops in different growth stages and environmental conditions. This can not only ensure that crops are in a healthy growth state, but also effectively guarantee the crop yield. Brief description of the drawings

[0090] Figure 1 is a flow chart of the present invention;

[0091] Figure 2 is a block diagram of data acquisition;

[0092] Figure 3 is a schematic diagram of the installation of soil sensors, where part (a) is the installation schematic diagram at 10 cm, and part (b) is the installation schematic diagram at 20 cm;

[0093] Figure 4 is a schematic diagram of the installation of air sensors;

[0094] Figure 5 is the communication daughter board of the acquisition system;

[0095] Figure 6 is the communication main board of the acquisition system;

[0096] Figure 7 is a data acquisition flow chart;

[0097] Figure 8 is for the user to view data in real time;

[0098] Figure 9 is a schematic diagram of the height acquisition of Panax notoginseng plants;

[0099] Figure 10 is a schematic diagram of the stem diameter acquisition of Panax notoginseng;

[0100] Figure 11 is a comparison chart before and after processing data using Gaussian noise reduction, where part (a) is the effect diagram before filtering; part (b) is the effect diagram after filtering;

[0101] Figure 12It is the flow chart of the Informer-LSTM-EWMA algorithm;

[0102] Figure 13 It is the intervention timing diagram;

[0103] Figure 14 It is the main interface of the operation interface;

[0104] Figure 15 It is the view graph of feature data. Among them, part (a) is the result graph of sensor 1; part (b) is the result graph of sensor 2; part (c) is the result graph of sensor 3; part (d) is the result graph of sensor 4;

[0105] Figure 16 It is the experimental site prepared before the start of the experiment;

[0106] Figure 17 It is the result of the irrigation method using the method of the present invention;

[0107] Figure 18 It is the result of using the traditional irrigation method; Detailed implementation manners

[0108] The present invention will be further described in detail below in conjunction with specific embodiments.

[0109] As Figure 1 shown, a data-driven prediction irrigation method for the growth state of Panax notoginseng includes the following steps:

[0110] S1. Collect Panax notoginseng environmental data and Panax notoginseng growth data;

[0111] As Figure 2 shown, the Panax notoginseng environmental data is collected once an hour through a data remote collection system based on the Internet of Things technology in the Panax notoginseng planting base, including: soil temperature, soil humidity, soil conductivity, air temperature and air humidity;

[0112] The Panax notoginseng growth data includes: Panax notoginseng growth data and water demand data at different growth stages; among them, the Panax notoginseng growth data includes: recording the plant height and stem diameter growth data of Panax notoginseng;

[0113] The collection steps of the Panax notoginseng environmental data are as follows:

[0114] S1.1. Install the soil temperature, humidity and conductivity sensors in two layers. The first layer is installed 10 cm underground, as Figure 3 shown in part (a) of Figure 3 ; the second layer is installed 20 cm underground, as Figure 4 shown in part (b) of

[0115] S1.2. Connect each sensor to the communication daughter board and number them. The communication daughter board transmits the data to the communication main board via Bluetooth. The communication main board collects the information collected by no more than 16 communication daughter boards, as Figure 5 shown; finally, the communication main board packs the total information collected and transmits it to the cloud via 4G, as Figure 6 shown;

[0116] S1.3. As Figure 7 shown, after the on-site layout is completed, the data can be viewed in real time at the terminal;

[0117] As Figure 8 shown, in this embodiment, the terminal is the Xshell software. The user opens the Xshell software, logs in to the account, selects the data to be viewed, and then views it in real time. Finally, the data is downloaded to the local file for subsequent data processing and analysis;

[0118] S2. Denoise the data collected in S1;

[0119] As Figure 9 and Figure 10 shown, since the data is collected in the real environment, there is noise in the collected data. In order to reduce the impact of noise on subsequent data analysis, the data is denoised;

[0120] Use Gaussian denoising to denoise the data. The denoising model is as follows:

[0121]

[0122] In the formula, G(x) represents the output signal after Gaussian denoising; x represents the value of the input signal, and the input signal includes: soil temperature, soil humidity, soil conductivity, air temperature, and air humidity; σ represents the standard deviation of the Gaussian distribution, and the standard deviation is set to 10 in this embodiment; the denoising result is as Figure 11 shown in parts (a) and (b);

[0123] S3. As Figure 12 shown, according to the deep learning algorithm Informer-LSTM-EWMA, construct a prediction model for the growth state of Panax notoginseng, and predict the irrigation demand of Panax notoginseng by inputting the result of S2 into the prediction model for the growth state of Panax notoginseng;

[0124] Specifically, adopt the Informer-LSTM-EWMA model, combine historical data and expert knowledge to construct a prediction model for the growth state of Panax notoginseng. The input of the model is the historical environmental data and the morphological data of Panax notoginseng, including: historical data of soil temperature, soil humidity, soil conductivity, air temperature, air humidity, and the plant height and stem diameter of Panax notoginseng, and the output is the prediction of the future growth state (plant height) and irrigation timing of Panax notoginseng;

[0125] The construction steps are as follows:

[0126] S3.1. Divide the denoised data into a training set, a test set, and a validation set according to 7:2:1;

[0127] S3.2. Construct an Informer-LSTM-EWMA model; the model is divided into three layers, and the construction steps are as follows:

[0128] S3.2.1. Construct the first-layer Informer model;

[0129] The self-attention mechanism of the Informer model is based on tuple inputs, namely query Q, key K, and value V, and then the scaled dot product, and the expression is as follows:

[0130]

[0131] In the formula, d represents the input dimension. In this embodiment, a single soil sensor can collect three characteristic data of soil temperature, soil humidity, and soil conductivity. Since 4 sensors are installed in the test field, the soil can collect 4×3 characteristics. Plus 4 characteristic data of air temperature, air humidity, plant height, and stem diameter of Panax notoginseng, there are a total of 16 characteristic data. Therefore, d is set to 16 in this embodiment; represents the set of natural numbers; L Q , L K , L V are the lengths (number of time steps) of the input sequences of Q, K, and V, and can also be understood as the amount of data required for a single model training. In this embodiment, it is set to 136;

[0132] Let qi, ki, and vi represent the i-th rows of Q, K, and V respectively; the attention of the m-th query is defined as kernel smoothing in the form of probability; the expression is as follows:

[0133]

[0134] In the formula, represents the probability distribution of attention; k(q i , k j ) represents an asymmetric exponential kernel j represents the row of the matrix, l represents the row of the matrix, and j ∈ l;

[0135] In the Informer model, attention values are used to select important q-k pairs to describe the similarity between the n-th query and the key. q-k represents the query pairs formed by the corresponding rows in matrices Q and K; the attention value of the n-th query to the key can be expressed in probability as q(k j |q i); Consider q(k j |q i ) = 1 / L K is a uniform distribution. The closer p is to the uniform distribution q, the less important it is;

[0136] The expression for the sparsity measurement value M of the nth query is as follows:

[0137]

[0138] where L k represents the rows of the K matrix;

[0139] The above formula can be simplified by scaling to:

[0140]

[0141] Based on the proposed metric Define u = c·lnL Q dominant query number, where L Q represents the rows of the Q matrix; thus, probabilistic sparse self-attention can be defined as:

[0142]

[0143] where represents a sparse matrix of the same size as Q and containing only the largest u queries in M, where u is controlled by the sampling factor c, and the sampling factor c should be set differently according to the actual data in different environments. In this embodiment, the sampling factor c is set to 5;

[0144] To extract dominant features, a distillation operation is added to the Informer model. Through the distillation operation, the input dimension can be significantly reduced; the distillation operation is a serial combination of 1D convolution, ELU activation function, and max pooling;

[0145] Advancing from the ith layer to the (j + 1)th layer, the formula is as follows:

[0146]

[0147] where [·] AB represents the attention block, which contains multi-head ProbSparse self-attention. In Informer, j is the same as the number of encoder and decoder layers of the model; in this embodiment, the model has 2 encoder layers and 2 decoder layers, and the range of j is from 1 to 2;

[0148] S3.2.2. Input the result of S3.2.1 into the LSTM model to construct the second-layer LSTM model;

[0149] LSTM is a deep learning model for processing sequential data; it contains a storage unit and three gates: the forget gate, the input gate, and the output gate; the role of these gates is to selectively control the flow of information, allowing the model to more effectively capture long-term dependencies in long sequences;

[0150] At each time step t, each gate generates a state variable f t , i t , o t , the output unit h t and the cell state C t , and the expressions are as follows:

[0151] f t = σ(w f [h t-1 , x t-1 + b f )

[0152] i t = σ(w i [h t-1 , x t-1 + b i )

[0153]

[0154] o t = σ(w o [h t-1 , x t-1 + b o )

[0155]

[0156] h t = o t * tanh(C t )

[0157] In the formula, [w f , b f , [w i , b i , [w c , b c , [w o , b o correspond to the forget gate f t , the input gate i t , the cell state C t , and the output gate o tThe weight matrix and bias term. At the beginning of training, the weight matrix and bias term are randomly given. In a Python program, the weight matrix and bias term can be randomly generated through the random() function. Although the weight matrix and bias term are random at the beginning, as the model is continuously trained, the weight matrix and bias term are iteratively updated with the training of the model. At the end of the model training, the weight matrix and bias term finally reach the ideal values; [h t-1 ,x t-1 are the output in the hidden layer at time t-1 and the input at time t-1 respectively; is the input node state;

[0158] Through this mechanism, LSTM can retain important information in sequential data and make predictions or classification tasks based on this;

[0159] S3.2.3. Input the result of S3.2.2 into the EWMA model to construct the third-layer EWMA model;

[0160] The EWMA model is a method used to estimate trends and periodic changes in time series data. It performs a weighted average of historical data to better capture the change trend of the most recent observations and is widely used in the smoothing processing and trend prediction of time series data;

[0161] In the EWMA model, newer observations are given larger weights, while older observations are given smaller weights. This allows the model to more flexibly adapt to data changes and also reduces the impact of random noise on the prediction results. The expression of the EWMA model is as follows:

[0162] EWMA T =α*X T +(1-α)EWMA T

[0163] In the formula, EWMA t is the EWMA value at time t; X t is the observation value at time t; α is the smoothing parameter, (0 < α < 1), which determines the relative weight of the observation value and the previous EWMA t-1 value;

[0164] S3.2.4. Normalize the output of S3.2.3 so that the output result is within a preset range; the preset range is: [0, 1]

[0165] Specifically, when making predictions, the historical data of soil temperature, soil humidity, soil conductivity, air temperature, air humidity, as well as the plant height and stem diameter data of Panax notoginseng are used as the input of the model, and the future growth state (plant height) data of Panax notoginseng are used as the output; during the experiment, only the growth state of Panax notoginseng needs to be known, and the accurate value of the plant height morphological data of Panax notoginseng does not need to be known; therefore, when making predictions, the output data is normalized so that the output result is within the range of [0, 1], and the expression is as follows:

[0166]

[0167] In the formula, max is the maximum value in the feature data; min is the minimum value in the feature data; x′ is the data before normalization, and x* is the data after normalization;

[0168] S3.2.5. Optimize the model parameters through the cross-validation method to obtain the optimal values, and substitute the optimal values into the model to obtain the optimal prediction model;

[0169] Specifically, the software automatically optimizes the optimal values of the encoder layer number, decoder layer number, training batch size, and training times parameters in the model through the cross-validation method, and imports the optimal value parameters into the verified Informer-LSTM-EWMA algorithm model to obtain the optimal prediction model;

[0170] S3.3. Use the training dataset to train the Informer-LSTM-EWMA algorithm model to obtain a training model, use the test dataset to test the trained model, and finally use the validation set to validate the model;

[0171] S3.4. Take the collected data as the input and input it into the verified model, and then irrigate according to the predicted growth state of Panax notoginseng; at the same time, determine the irrigation amount according to the humidity data collected by the current sensor;

[0172] Take the newly collected data as the input and input it into the verified model, and then according to the predicted growth state of Panax notoginseng, as Figure 13 shown, determine when to irrigate according to the time indicated by the dotted line in the figure. In this embodiment, according to Figure 13 shown, the time indicated by the dotted line is 72, which means that starting from the current time reference point, an irrigation is carried out approximately 72 hours later; at the same time, determine the irrigation amount according to the humidity data collected by the current sensor;

[0173] S4. Make an irrigation decision according to the prediction result;

[0174] Combined with the growth requirements and irrigation principles of Panax notoginseng, make a scientific and reasonable irrigation decision;

[0175] Irrigation decision-making includes: linearly fitting the intervention timing, the irrigation amount of the soil, and the soil humidity; aiming to ensure that Panax notoginseng grows under the best growth conditions while maximizing water resource conservation;

[0176] Specifically, after the model is established, according to the current air temperature, air humidity, soil temperature, soil humidity, and soil conductivity, the trained model is used to predict the irrigation demand of Panax notoginseng;

[0177] The steps for formulating the irrigation decision are as follows:

[0178] S4.1, as shown in Figure 13 , according to the results predicted by the model, the intervention timing T can be found through maximum value retrieval itt , and the intervention timing T itt The expression is as follows:

[0179] T itt = index(max([y 1 , y 2 , …y k ))

[0180] In the formula, [y 1 , y 2 ,..., y k represents the model prediction result, and k is a preset value;

[0181] S4.2, linearly fit the irrigation amount of the soil and the soil humidity to obtain the relationship between the irrigation amount of the soil and the soil humidity;

[0182] Specifically, let the humidity of the sensor before irrigation be x1, and irrigate the soil with x 2 liters of volume water in the normal irrigation method. Wait for a period of time, and record the soil humidity x 3 after the water penetrates downward; the expression is as follows:

[0183] x 3 = 8x 2 + x 1

[0184] In the formula, s represents the current soil water irrigation coefficient. In this case, the value of s is measured to be 0.048;

[0185] In addition, it should be noted that due to the differences in soil texture, altitude, and other various environmental factors in the planting area, the s values under various environmental conditions will also change accordingly;

[0186] Through this linear fitting method, the impact of the irrigation amount on the soil humidity can be roughly estimated, which also provides guidance for subsequent irrigation;

[0187] S4.3. Irrigate by combining the prediction results of the model, the intervention timing, and the soil humidity;

[0188] In this embodiment, the irrigation records are shown in Table 1:

[0189] Table 1: Irrigation records of different test plots

[0190]

[0191]

[0192] In the table, \ means no intervention measures are taken;

[0193] To achieve visual operation, the present invention develops a simple visual interface using the flet interface library of python, as Figure 14 shown. In the Figure 14 interface, the user can select the data of the corresponding test plot for viewing. There are three characteristic data of temperature, humidity, and conductivity for each test plot that can be selected for viewing. After clicking on the corresponding characteristic data for viewing, it will jump to the visual interface as Figure 15 . In this interface, the viewed characteristic data can be displayed. Figure 15 . In Figure 14 , parts (a), (b), (c), and (d) are respectively the schematic diagrams of the results of sensors 1 to 4 in this embodiment. There are two buttons, "Return to Main Menu" and "Start Prediction", at the top. Clicking the "Return to Main Menu" button will return to the

[0194] S5. Execute the irrigation and monitoring strategy;

[0195] Irrigation execution: Execute the irrigation operation according to the irrigation plan to ensure that the pseudo-ginseng obtains the required moisture. Since it takes time for the moisture to penetrate, the humidity after irrigation is based on the sensor readings of the next day. Table 2 shows the humidity of the test plot after irrigation in this case;

[0196] Table 2 Soil humidity after irrigation

[0197]

[0198] Irrigation effect monitoring: Evaluate the irrigation effect by monitoring the soil humidity and the growth data of pseudo-ginseng, and adjust the irrigation plan according to the actual situation.

[0199] Figure 16 and Figure 18 are for the comparison before and after the traditional irrigation decision experiment.Figure 16 and Figure 17 is a comparison before and after the experiment of adopting a data-driven irrigation method. From Figure 17 and Figure 18 the results, it can be clearly seen that the solution proposed by the present invention is superior to the traditional irrigation decision. And Figure 18 the density of Panax notoginseng is much lower than that before the experiment Figure 16 ; the reason for this phenomenon is that unreasonable irrigation leads to the death of Panax notoginseng.

[0200] It should be noted that the irrigation of Panax notoginseng should avoid waterlogging and the equipment needs to be maintained regularly;

[0201] Avoid waterlogging: Panax notoginseng likes humidity but is afraid of waterlogging. When irrigating, it should be ensured that the soil is moist but not waterlogged, so as not to have an adverse impact on the growth of Panax notoginseng;

[0202] Regular maintenance: Regularly maintain the irrigation equipment and monitoring system to ensure its normal operation and accurate monitoring.

[0203] A data-driven irrigation system for predicting the growth status of Panax notoginseng;

[0204] It includes: a data acquisition module, a data processing module, a model training module, and an irrigation decision-making module;

[0205] The data acquisition module is used to perform the following steps:

[0206] S1. Collect the environmental data and growth data of Panax notoginseng;

[0207] The data processing module is used to perform the following steps:

[0208] S2. Perform noise reduction processing on the data;

[0209] The model training module is used to perform the following steps:

[0210] S3. Construct a prediction model for the growth status of Panax notoginseng according to the deep learning algorithm Informer-LSTM-EWMA, and predict the irrigation demand of Panax notoginseng by inputting the result of S2 into the prediction model for the growth status of Panax notoginseng;

[0211] The irrigation decision-making module is used to perform the following steps:

[0212] S4. Make an irrigation decision according to the prediction result;

[0213] S5. Execute the irrigation and monitoring strategy;

[0214] Finally, it should be emphasized that the above embodiments are only intended to illustrate the technical solutions of the present invention in the field of greenhouse planting, rather than limiting its protection scope. Although the present invention has been described in detail through preferred embodiments, those skilled in the art should recognize that necessary modifications or equivalent alternative solutions can be made without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A data-driven irrigation method for predicting the growth status of Panax notoginseng, characterized in that: The following steps are involved: S1. Collecting Panax notoginseng environmental data and Panax notoginseng growth data; The environmental data of Panax notoginseng is collected once an hour through the remote data collection system based on Internet of Things technology at the Panax notoginseng planting base, including: soil temperature, soil moisture, soil conductivity, air temperature and air humidity; The growth data of Panax notoginseng include: the growth data of Panax notoginseng and the water demand data at different growth stages; the growth data of Panax notoginseng include: the growth data of plant height and stem diameter of Panax notoginseng; S2, performing noise reduction processing on the data collected by S1; The noise reduction process uses Gaussian noise reduction; S3, build a Panax notoginseng growth state prediction model based on the deep learning algorithm Informer-LSTM-EWMA, and predict the irrigation demand of Panax notoginseng by inputting the results of S2 into the Panax notoginseng growth state prediction model; S4. Make irrigation decisions based on the prediction results; S5. Execute irrigation decision and monitoring strategies.

2. The data-driven irrigation method for predicting the growth status of Panax notoginseng according to claim 1, characterized in that: The steps for collecting the Panax notoginseng environmental data are as follows: S1.

1. Install the soil temperature, humidity and conductivity sensors in two layers, the first layer is installed 10cm underground, and the second layer is installed 20cm underground; the air temperature and humidity sensors are installed on the ground 50cm above the ground; S1.

2. Connect each sensor to the communication sub-board and number them. The communication sub-board transmits the data to the communication main board via Bluetooth. The communication main board collects the information collected by no more than 16 communication sub-boards. Finally, the communication main board packages the total information collected and transmits it to the cloud via 4G. S1.

3. After the on-site layout is completed, the data can be viewed in real time on the terminal.

3. The data-driven irrigation method for predicting the growth status of Panax notoginseng according to claim 1, characterized in that: The steps of constructing a Panax notoginseng growth state prediction model according to the deep learning algorithm Informer-LSTM-EWMA and predicting the irrigation demand of Panax notoginseng by inputting the result of S2 into the Panax notoginseng growth state prediction model are as follows: S3.1, divide the denoised data into training set, test set and validation set according to 7:2:1; S3.

2. Build the Informer-LSTM-EWMA model. The model is divided into three layers, and the construction steps are as follows: S3.2.

1. Build the first-layer Informer model; The self-attention mechanism of the Informer model is based on tuple input, namely query Q, key K and value V, and then scaled dot product, as shown below: In the formula, Where d represents the input dimension; represents the set of natural numbers; L Q , L K , L V is the length of the input sequence of Q, K and V; The expression of the sparsity measure M of the nth query in the Informer model is as follows: Where, L k represents the row of K matrix; After scaling, it can be simplified to: Definition u=c·lnL Q The number of dominant queries, among which L Q represents the row of the Q matrix; thus, the probabilistic sparse self-attention can be defined as: In the formula, represents a sparse matrix of the same size as Q and containing only the largest u queries in M, where u is controlled by the sampling factor c; In order to extract dominant features, a distillation operation is added to the Informer model; the distillation operation is a series combination of 1D convolution, ELU activation function and maximum pooling; The formula for advancing from the jth layer to the (j+1) layer is as follows: In the formula, [·] AB represents an attention block, including multi-head ProbSparse self-attention. In Informer, j has the same number of encoder and decoder layers as the model; S3.2.2, input the result of S3.2.1 into the LSTM model to build the second layer LSTM model; LSTM consists of a storage unit and three gates: forget gate, input gate and output gate; each gate generates a state variable f at time t t 、i t , o t , output unit h t and cell status C t , the expression is as follows: f t =σ(w f [h t-1 ,x t-1 ]+b f ) i t =σ(w i [h t-1 ,x t-1 ]+b i ) the t =σ(w o [h t-1 ,x t-1 ]+b o ) h t =o t *tanh(C t ) In the formula, [w f , b f ]、[w i , b i ]、[w c , b c ]、[w o , b o ] respectively correspond to the forget gate f at time t t , input gate i t , unit status C t , output gate o t The weight matrix and bias term of [h t-1 ,x t-1 ] are the output of the hidden layer at time t-1 and the input at time t-1 respectively; is the input node status; S3.2.3, input the results of S3.2.2 into the EWMA model to construct the third-level EWMA model; The expression of the EWMA model is as follows: EWMA T =α*X T +(1-a)EWMA T Where, EWMA t is the EWMA value at time t; X t is the observed value at time t; α is the smoothing parameter, (0<α<1), which determines the difference between the observed value and the previous EWMA t-1 The relative weight of the values; S3.2.4, normalize the output of S3.2.3 so that the output result is within a preset range; The default range is: [0, 1] When making predictions, the output data is normalized so that the output result is in the range of [0, 1]. The expression is as follows: In the formula, max is the maximum value of the feature data; min is the minimum value of the feature data; x′ is the data before normalization, and x* is the data after normalization; S3.2.

5. Optimize the model parameters through cross-validation method to obtain the optimal value, and bring the optimal value into the model to obtain the optimal prediction model; S3.3, use the training data set to train the Informer-LSTM-EWMA algorithm model, obtain the training model, use the test data set to test the trained model, and finally use the validation set to validate the model; S3.

4. The collected data is used as input into the verified model, and then irrigation is performed according to the predicted growth status of Panax notoginseng; at the same time, the irrigation amount is determined according to the humidity data collected by the current sensor.

4. The data-driven irrigation method for predicting the growth status of Panax notoginseng according to claim 1, characterized in that: In the irrigation decision making according to the prediction result, the irrigation decision includes: linear fitting of intervention timing, soil irrigation amount and soil moisture; The steps to making irrigation decisions are as follows: S4.

1. Based on the results predicted by the model, the intervention opportunity T can be found through maximum value retrieval. itt , intervention time T itt The expression is as follows: T itt =index(max([y1,y2,…y k ])) In the formula, [y1,y2,...,y k ] represents the model prediction result, and k is the preset value; S4.2, linearly fitting the soil irrigation amount and soil moisture to obtain the relationship between the soil irrigation amount and soil moisture; The humidity of the sensor before irrigation is x1; the soil is irrigated with x2 liters of water in the normal irrigation method; wait for a period of time, and record the soil humidity x3 after irrigation after the water penetrates downward; the linear fitting expression is as follows: x3=sx2+x1 In the formula, s represents the current soil water irrigation coefficient; S4.

3. Irrigate based on the model's predictions, intervention timing, and soil moisture.

5. The data-driven Panax notoginseng growth status prediction irrigation system according to claim 1, characterized in that: The system includes: a data acquisition module, a data processing module, a model training module and an irrigation decision-making module; The data acquisition module is used to perform the following steps: S1. Collecting Panax notoginseng environmental data and Panax notoginseng growth data; The data processing module is used to perform the following steps: S2, perform noise reduction on the data; The model training module performs the following steps: S3, build a Panax notoginseng growth state prediction model based on the deep learning algorithm Informer-LSTM-EWMA, and predict the irrigation demand of Panax notoginseng by inputting the results of S2 into the Panax notoginseng growth state prediction model; The irrigation decision making module is used to perform the following steps: S4. Make irrigation decisions based on the prediction results; S5. Implement irrigation strategies and monitoring strategies.

Images