Crop cloud supervision method and system based on Internet of Things

Abstract

The invention relates to a crop cloud supervision method and system based on the Internet of Things, and the method comprises the steps: collecting the environment data of a crop planting region through a sensor, determining a self-adaptive noise weight for the time sequence sensing data collected by each sensor through a stage residual signal entropy value, and carrying out the recognition of the self-adaptive noise weight; performing ensemble empirical mode decomposition on the time sequence sensing data through candidate component weighting to obtain a multi-order intrinsic mode component; acquiring crop growth environment context information including time, growth stages and external weather; based on the context information, using an attention network to calculate an attention score and normalize the attention score to obtain a context sensing fusion weight of each order of intrinsic mode component; performing weighted fusion on the data feature vector of each order of intrinsic mode component by using a context sensing fusion weight to obtain a comprehensive fusion feature; and inputting the comprehensive fusion features into a pre-trained crop state analysis model to obtain crop state information, and generating crop supervision information based on the crop state information and a preset rule.

Description

Technical Field

[0001] The present invention relates to the field of agriculture, and particularly to a method and system for cloud supervision of crops based on the Internet of Things. Background Art

[0002] By deploying various sensor nodes in farmland, key parameters of the crop growth environment can be collected in real time and continuously, such as soil temperature and humidity, air temperature and humidity, light intensity, carbon dioxide concentration, etc. These data are transmitted to the cloud platform through a wireless network for storage and analysis, providing a data basis for realizing refined management of crops, optimizing resource utilization, and improving yield and quality. However, the data collected by Internet of Things sensors often have characteristics such as large data volume, non-linearity, non-stationarity, and severe noise interference. This brings noise to the collected data. For the processing of noise, generally, denoising methods such as empirical mode decomposition are used. Traditional EMD-based methods still have problems such as mode mixing, end effect, and sensitivity to noise when processing complex signals. In ICEEMDAN, adding auxiliary noise can reduce mode mixing, but the amplitude of the auxiliary noise, that is, the weight, is difficult to determine, and a fixed noise weight is difficult to adapt to all signals. For the multi-source data of the cloud supervision of crops, existing methods often ignore the dynamic changes in the importance of different data sources for judging the crop state at different times, different growth stages, or different environmental conditions, and fail to make full use of the available context information to guide the fusion process, resulting in limited representation ability of the fused features and being difficult to comprehensively and accurately reflect the true situation of the crops, thus affecting the accuracy of subsequent state analysis and decision-making. Therefore, reducing the noise of sensors located in farmland and improving the expression ability of fused features are the keys to the precise supervision of crops. Summary of the Invention

[0003] In view of the above problems, in the first aspect of the present invention, a method for cloud supervision of crops based on the Internet of Things is provided. The method includes the following steps: Collect environmental data of the crop planting area through sensors. For the time-series sensing data collected by each sensor, determine the adaptive noise weight using the stage residual signal entropy value, and perform ensemble empirical mode decomposition on the time-series sensing data through weighted candidate components to obtain multi-order intrinsic mode components; Obtain the context information of the crop growth environment including time, growth stage, and external meteorology; based on the context information, calculate and normalize the attention scores using an attention network to obtain the context-aware fusion weights of each order of intrinsic mode components; Perform weighted fusion on the data feature vectors of the eigenmode components of each order by using the context-aware fusion weight to obtain a comprehensive fusion feature; input the comprehensive fusion feature into a pre-trained crop status analysis model to obtain crop status information, and generate crop supervision information based on the crop status information and preset rules.

[0004] Preferably, determining the adaptive noise weight by using the stage residual signal entropy value, and performing ensemble empirical mode decomposition on the time series sensing data through weighted candidate components, specifically: For the time series sensing data from one sensor, initialize it as the 0th order residual signal , set the decomposition order k = 1; loop to execute the following steps until the residual signal meets the preset decomposition termination condition: a) Calculate the sample entropy or permutation entropy value of the time segment of the (k - 1)th order residual signal ; b) Map the entropy value obtained in step a) through a predefined function to obtain the adaptive noise weight required for the kth order decomposition ; c) Generate N groups of independent Gaussian white noise sequences with a mean of 0 and a standard deviation of ; Add each group of noise sequences to the (k - 1)th order residual signal to form N groups of noise-assisted signals; d) Apply the empirical mode decomposition algorithm to the N groups of noise-assisted signals respectively, and extract N candidate kth order eigenmode components , where i is a positive integer, and ; e) Calculate the orthogonality index between each candidate eigenmode component and the (k - 1)th order residual signal ; f) According to the orthogonality index calculated in step e), perform weighted averaging on the N candidate eigenmode components to calculate the final kth order eigenmode component ; The weight of the weighted average is proportional to the orthogonality index; g) Calculate the kth order residual signal ; h) Increment the decomposition order k by 1; After the loop ends, obtain all eigenmode components as the decomposition result of the sensor time series data.

[0005] Preferably, based on the context information, calculate the attention score by using the attention network and normalize it to obtain the context-aware fusion weight of each order eigenmode component, specifically: Input the context information into the embedding layer or the encoder network to generate the context feature vector Q; Input each component in the set of final intrinsic mode components of all orders obtained by decomposing all sensors into its corresponding embedding layer or encoder network to generate a set of intrinsic mode component feature vectors and , where M is the total number of all intrinsic mode components generated by all sensors, and are the key vector and value vector of the j-th intrinsic mode component respectively; Use the context feature vector Q as the query, and calculate the attention score between the query Q and each key ; Apply the Softmax function to normalize all the attention scores to obtain the final context-aware fusion weight , where is the fusion weight corresponding to the j-th intrinsic mode component.

[0006] Preferably, the weighted fusion is performed on the data feature vectors of each order of intrinsic mode components by using the context-aware fusion weight to obtain the comprehensive fusion feature, specifically: Multiply the obtained fusion weight corresponding to each intrinsic mode component by its corresponding value vector to perform a scalar multiplication operation to obtain the weighted value vector ; Perform a vector summation or vector concatenation operation on all the weighted value vectors to generate the comprehensive fusion feature vector.

[0007] Preferably, the obtaining of the context information of the crop growth environment including time, growth stage, and external meteorology is specifically: Obtain the current date and time information; based on the crop planting date and the preset crop growth cycle model, judge and record the current growth stage of the crop; Request the current and short-term future meteorological data of the specified geographical location from the third-party meteorological service platform through the network interface, including but not limited to temperature, humidity, precipitation, wind speed, and light intensity; Integrate the obtained time information, growth stage identifier, and structured meteorological data into a context information.

[0008] Preferably, the inputting of the comprehensive fusion feature into the pre-trained crop status analysis model to obtain the crop status information, and generating the crop supervision information based on the crop status information and the preset rules is specifically: Use the comprehensive fusion feature vector as the input data of the crop status analysis model, and the crop status analysis model outputs an evaluation result; According to a pre-set rule base, match the evaluation result output by the model with the rule conditions; when the match is successful, trigger the generation of the corresponding crop supervision information text.

[0009] In the second aspect of the present invention, a crop cloud supervision system based on the Internet of Things is provided, and the system includes the following modules: A data acquisition module, configured to collect environmental data of the crop planting area through sensors. For the time-series sensing data collected by each sensor, determine the adaptive noise weight using the phase residual signal entropy value, and perform ensemble empirical mode decomposition on the time-series sensing data through candidate component weighting to obtain multi-order intrinsic mode components; A feature fusion module, configured to obtain crop growth environment context information including time, growth stage, and external meteorology; based on the context information, calculate and normalize the attention score using an attention network to obtain the context-aware fusion weight of each order of intrinsic mode components; An information output module, configured to perform weighted fusion on the data feature vectors of each order of intrinsic mode components using the context-aware fusion weight to obtain a comprehensive fusion feature; input the comprehensive fusion feature into a pre-trained crop status analysis model to obtain crop status information, and generate crop supervision information based on the crop status information and preset rules.

[0010] Preferably, the method of determining the adaptive noise weight using the phase residual signal entropy value and performing ensemble empirical mode decomposition on the time-series sensing data through candidate component weighting to obtain multi-order intrinsic mode components is specifically as follows: For the time-series sensing data from one sensor, initialize it as the 0th order residual signal , set the decomposition order k = 1; loop and execute the following steps until the residual signal meets the preset decomposition termination condition: a) Calculate the sample entropy or permutation entropy value of the time segment of the (k - 1)th order residual signal ; b) Map the entropy value obtained in step a) through a predefined function to obtain the adaptive noise weight required for the kth order decomposition ; c) Generate N groups of independent Gaussian white noise sequences with a mean of 0 and a standard deviation of ; add each group of noise sequences to the (k - 1)th order residual signal to form N groups of noise-assisted signals; d) Apply the empirical mode decomposition algorithm to each of the N groups of noise-assisted signals to extract N candidate kth order intrinsic mode components , where i is a positive integer, and ; e) Calculate the orthogonality index between each candidate intrinsic mode component and the (k - 1)-th order residual signal ; f) According to the orthogonality index calculated in step e), perform weighted averaging on the N candidate intrinsic mode components to calculate the final k-th order intrinsic mode component ; the weight of the weighted averaging is proportional to the orthogonality index; g) Calculate the k-th order residual signal ; h) Increment the decomposition order k by 1; After the loop ends, all intrinsic mode components are obtained as the decomposition result of the sensor time series data.

[0011] Preferably, based on the context information, an attention network is used to calculate the attention score and normalize it to obtain the context-aware fusion weight of each order of intrinsic mode component, specifically: Input the context information into the embedding layer or encoder network to generate the context feature vector Q; Input each component in the set of final intrinsic mode components of all orders decomposed from all sensors into its corresponding embedding layer or encoder network to generate a set of intrinsic mode component feature vectors and , where M is the total number of all intrinsic mode components generated by all sensors, and are the key vector and value vector of the j-th intrinsic mode component respectively; Use the context feature vector Q as the query to calculate the attention score between the query Q and each key ; Apply the Softmax function to normalize all attention scores to obtain the final context-aware fusion weight , where is the fusion weight corresponding to the j-th intrinsic mode component.

[0012] Preferably, the weighted fusion is performed on the data feature vectors of each order of intrinsic mode components by using the context-aware fusion weight to obtain the comprehensive fusion feature, specifically: Perform a scalar multiplication operation on the obtained fusion weight corresponding to each intrinsic mode component and its corresponding value vector to obtain the weighted value vector ; Sum or splice all weighted value vectors to generate a comprehensive fusion feature vector through vector summation or vector splicing operation.

[0013] Preferably, the obtaining of the context information of the crop growth environment including time, growth stage, and external meteorology is specifically as follows: Obtain the current date and time information; based on the crop planting date and a preset crop growth cycle model, judge and record the current growth stage of the crop; Request the current and short-term future meteorological data of a specified geographical location from a third-party meteorological service platform through a network interface, including but not limited to temperature, humidity, precipitation, wind speed, and light intensity; Integrate the obtained time information, growth stage identifier, and structured meteorological data into a context information.

[0014] Preferably, the inputting of the comprehensive fusion feature into a pre-trained crop status analysis model to obtain crop status information and generating crop supervision information based on the crop status information and preset rules is specifically as follows: Use the comprehensive fusion feature vector as the input data of the crop status analysis model, and the crop status analysis model outputs an evaluation result; According to a preset rule library, match the evaluation result output by the model with the rule conditions; when the match is successful, trigger the generation of the corresponding crop supervision information text.

[0015] In the present invention, at each order of the ensemble empirical mode decomposition, the adaptive noise weight is dynamically determined by using the entropy value of the current residual signal, and based on the orthogonality evaluation and weighted average of the candidate components generated by multiple noise experiments, it can more finely adapt to the local time-varying characteristics of the signal, effectively suppress the mode mixing phenomenon, and improve the accuracy and stability of the intrinsic mode component decomposition; by introducing an attention mechanism based on the context information of the crop growth environment to calculate the fusion weights of each intrinsic mode component, the model can focus on the information components that are most important for judging the crop status, assign higher fusion weights to them, thereby generating a more discriminative and representative comprehensive fusion feature, and further improving the accuracy of crop supervision. Description of the Drawings

[0016] Figure 1 is the flowchart of Embodiment 1; Figure 2 is a schematic diagram of the data collected by a humidity sensor; Figure 3 is a schematic diagram of the modal decomposition of the data collected by the humidity sensor; Figure 4 is the prompt information received in the industrial brain system. Detailed implementation manners

[0017] In this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the element.

[0018] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0019] Embodiment 1 provides a cloud supervision method for crops based on the Internet of Things, as Figure 1 shown, the method includes the following steps: S1. Collect environmental data of the crop planting area through sensors. For the time-series sensing data collected by each sensor, determine the adaptive noise weight using the stage residual signal entropy value, and perform ensemble empirical mode decomposition on the time-series sensing data through weighted candidate components to obtain multi-order intrinsic mode components; Use sensors such as temperature, humidity, and light sensors to collect environmental information of the farmland. In another embodiment, the sensors further include image sensors, soil pH sensors, etc. The image sensors include, but are not limited to, multi-spectral or hyperspectral sensors. Since this information changes over time, the data collected by each sensor constitutes the time-series data of this sensor. Figure 2 Shows a curve graph of the data collected by the humidity sensor. For the data sequence collected by each sensor, a method including, but not limited to, ensemble empirical mode decomposition is used for decomposition. Before decomposition, first calculate the degree of chaos of the signal at different stages, that is, the residual signal entropy value, to determine how much interference noise to add during decomposition, and decompose the original complex signal into a series of relatively simple basic waveforms from high frequency to low frequency, that is, intrinsic mode components (IMFs). By weighted averaging multiple candidate IMFs generated during the decomposition process, the final IMFs of each order are obtained.

[0020] S2, obtain the crop growth environment context information including time, growth stage, and external meteorology; based on the context information, use the attention network to calculate the attention scores and normalize them to obtain the context-aware fusion weights of each order of intrinsic mode components. Obtain information related to the current crop growth environment, including but not limited to the current time point, which growth and development stage the crop is in, such as the seedling stage, flowering stage, fruiting stage, etc., and the external weather conditions, such as temperature, humidity, light, rainfall forecast, etc. In one example, the context information is: "Time: 10:00 am, March 20, 2024, Crop stage: Tea tree budding stage, Weather: Sunny, Temperature 22°C, Humidity 50%, No rainfall in the next 3 hours".

[0021] Use the context information to determine which of the IMFs decomposed in S1 is more important in the current environment through the attention network. The attention network will calculate an attention score for each IMF, and the higher the score, the more worthy of attention in the current context. Then normalize these scores, preferably with the sum of the scores being 1, to obtain the final context-aware fusion weights. For example, according to the context of "sunny, high temperature, flowering stage", the attention network will consider the "IMF related to temperature" and the "IMF related to light" to be more important and give them higher weights; while the "IMF related to humidity" is relatively less important and is given a lower weight.

[0022] S3, perform weighted fusion on the data feature vectors of each order of intrinsic mode components using the context-aware fusion weights to obtain a comprehensive fusion feature; input the comprehensive fusion feature into a pre-trained crop status analysis model to obtain crop status information, and generate crop supervision information based on the crop status information and preset rules.

[0023] Multiply the weight of each IMF calculated in S2 by the data or feature vector corresponding to that IMF. Then combine all these weighted vectors, preferably by adding or concatenating, to form a comprehensive feature vector. This vector represents the set of the most important environmental features adjusted by the context information in the current environment.

[0024] The obtained comprehensive feature vector is input into a pre-trained crop status analysis model, and the analysis model will output the current crop status information, such as drought, etc. According to these status information and a preset rule base, specific crop supervision suggestions or alarm information are generated. For example, when the comprehensive feature vector is input into the model, the output after model analysis is "Status: Mild water shortage". Search the rule base and find the rule: "IF Status = Mild water shortage AND Crop stage = Germination stage THEN Generate information: 'The crop is in the germination stage, mild water shortage is detected, it is recommended to supplement irrigation in a timely manner'", and send this supervision information to the manager.

[0025] In denoising such as EEMD, white noise will be added to avoid mode mixing. However, the added noise is too small and mode mixing still exists. If the noise is too large, it will cover up the true signal, and since the noise is randomly generated, different noises have an impact on the final separation. In an optional embodiment, the method for determining the adaptive noise weight by using the entropy value of the stage residual signal and performing ensemble empirical mode decomposition on the time series sensing data through candidate component weighting to obtain multi-order intrinsic mode components is specifically as follows: For the time series sensing data from a sensor, initialize it as the 0th order residual signal , set the decomposition order k = 1; loop and execute the following steps until the residual signal meets the preset decomposition termination condition: a) Calculate the sample entropy or permutation entropy value of the time segment of the (k - 1)th order residual signal ; b) Map the entropy value obtained in step a) through a predefined function to obtain the adaptive noise weight required for the kth order decomposition ; c) Generate N groups of independent Gaussian white noise sequences with a mean of 0 and a standard deviation of ; Add each group of noise sequences to the (k - 1)th order residual signal to form N groups of noise-assisted signals; d) Apply the empirical mode decomposition algorithm to the N groups of noise-assisted signals respectively, and extract N candidate kth order intrinsic mode components , where i is a positive integer, and ; e) Calculate the orthogonality index between each candidate intrinsic mode component and the (k - 1)th order residual signal ; f) According to the orthogonality index calculated in step e), perform weighted averaging on the N candidate intrinsic mode components to calculate the final kth order intrinsic mode component ; The weight of the weighted average is proportional to the orthogonality index. g) Calculate the k-th order residual signal ; h) Increment the decomposition order k by 1; After the loop ends, all the intrinsic mode components are obtained as the decomposition result of the sensor time series data.

[0026] For example, when k = 4, the 3rd residual signal is , calculate of the sample entropy or permutation entropy. Assuming the calculated entropy value is 0.5, input the calculated entropy value into a predefined mapping function, and this mapping function will output the noise standard deviation applicable to the 4th order decomposition, denoted as . Since is usually simpler than , or , and is also smaller than the previous value. If is obtained through the mapping function calculation.

[0027] In a more specific example, the mapping function is a piecewise function. Preferably, if the entropy is less than 0.5, the adaptive noise weight is 0.1; if the entropy is greater than or equal to 0.5 and less than 1.2, the adaptive noise weight is 0.2; if the entropy is greater than or equal to 1.2, the adaptive noise weight is 0.3. Those skilled in the art should note that the piecewise manner and the corresponding values of the above piecewise function can be adjusted, not limited to the above data, and not limited to the above piecewise function.

[0028] Generate N = 50 independent Gaussian white noise sequences with a mean of 0 and a standard deviation of . Add these N groups of noise sequences independently to the 3rd residual signal to generate N different . Apply standard empirical mode decomposition to each of the N noise-assisted signals. From each EMD decomposition, extract the first IMF decomposed from it, and these are the N candidate 4th order . For each candidate , calculate its orthogonality index with respect to the signal from which it is extracted. This measures how well each candidate component is separated from the remaining signal . In one embodiment, the orthogonality index is and the (k - 1)-th order residual signal The reciprocal of the absolute value of the dot product or inner product, or the negative exponent of the absolute value of the dot product or inner product. More specifically, when taking the reciprocal, to prevent the inner product from being 0, after calculating the absolute value of the dot product or inner product, a very small positive number is added.

[0029] By performing a weighted average on all candidates to calculate the final . Each has a weight proportional to the calculated orthogonality index. The more cleanly separated, i.e., the higher the orthogonality, the greater the weight of the candidate component. Using the weights, the 50 are averaged to obtain the final . By subtracting from the residual signal from the previous step, the next residual signal is calculated. Increase the decomposition order by 1, i.e., k = 5, and perform the next decomposition. Figure 3 shows the obtained after decomposing the humidity data sequence.

[0030] In this embodiment, the adaptive noise weight enables the EEMD process to be adjusted according to the continuously changing complexity of the signal at each decomposition step, better balancing the need to effectively reduce mode mixing and prevent noise from introducing errors, thereby obtaining a more accurate decomposition result. Moreover, using the orthogonality index, the final IMFs are more concentrated on the components with good decomposition effects.

[0031] The state of the crop is not only determined by the sensor readings themselves but also related to the current environment. Under different environmental conditions, the importance of different sensors or different frequency components (IMFs) of the same sensor for judging the crop state is different. In an alternative embodiment, based on the context information, an attention network is used to calculate attention scores and normalize them to obtain context-aware fusion weights for each order of intrinsic mode components, specifically: Input the context information into an embedding layer or an encoder network to generate a context feature vector Q; Input each component in the set of final intrinsic mode components of all orders obtained by decomposing all sensors into its corresponding embedding layer or encoder network to generate a set of intrinsic mode component feature vectors and , where M is the total number of all intrinsic mode components generated by all sensors, and are the key vector and value vector of the jth intrinsic mode component respectively; Use the context feature vector Q as a query to calculate the attention score between the query Q and each key ; For all attention scores The Softmax function is applied for normalization to obtain the final context-aware fusion weights , where is the fusion weight corresponding to the j-th eigenmode component.

[0032] The context information is input as a whole into the embedding layer or the encoder network, where the encoder network includes, but is not limited to, the encoders of RNN, LSTM or Transformer, to obtain a vector with a fixed dimension, and the vector represents the current environmental state information, which is the query vector Q. For each IMF component in the set, the j-th IMF component is input into its corresponding embedding layer or encoder network. In one embodiment, different types of sensors or different orders of IMF use different encoders. In another embodiment, a shared encoder that can process all IMFs is used. For the j-th IMF, the network will generate two vectors, namely the key vector and the value vector , and then obtain M groups of key-value pairs: and . Calculate the similarity or correlation between Q and the key vector of each IMF as the attention score , and the calculation method includes, but is not limited to, dot product. Then M attention scores are obtained. The higher the score, the more important the IMF is in the current context. The Softmax function is used to normalize all the calculated M raw attention scores. The Softmax function converts any real-valued score into a probability distribution, that is, all output weights are positive and their sum is equal to 1, and the output result is the final context-aware fusion weight, and each weight is the final importance weight assigned to the j-th IMF component in the current specific context.

[0033] After obtaining the context-aware fusion weights, multiply the fusion weight corresponding to each obtained eigenmode component by its corresponding value vector to perform a scalar multiplication operation to obtain the weighted value vector ; Perform a vector summation or vector concatenation operation on all the weighted value vectors to generate a comprehensive fusion feature vector.

[0034] For example, for the of the main period IMF feature of the temperature sensor, and a certain fluctuating IMF feature of the humidity sensor , , then the comprehensive fusion feature vector calculated by the summation method is .

[0035] A simple sensor reading itself is an isolated value. Whether it is abnormal depends to a large extent on the specific environment. Moreover, the requirements of crops for environmental factors such as water, nutrients, temperature, and light vary dynamically during different growth stages. Only by knowing the current growth stage can it be determined whether the environmental conditions meet the specific requirements of that stage. In an alternative embodiment, the obtaining of context information of the crop growth environment including time, growth stage, and external meteorology is specifically as follows: Obtain the current date and time information; based on the crop planting date and a preset crop growth cycle model, judge and record the current growth stage of the crop; Request current and short-term future meteorological data of a specified geographical location from a third-party meteorological service platform through a network interface, including but not limited to temperature, humidity, precipitation, wind speed, and light intensity; Integrate the obtained time information, growth stage identifier, and structured meteorological data into a context information.

[0036] The comprehensive fusion feature not only covers sensor data but also information such as weather and crop growth stage. In an alternative embodiment, inputting the comprehensive fusion feature into a pre-trained crop status analysis model to obtain crop status information, and generating crop supervision information based on the crop status information and preset rules is specifically as follows: Use the comprehensive fusion feature vector as the input data of the crop status analysis model, and the crop status analysis model outputs an evaluation result; According to a pre-set rule library, match the evaluation result output by the model with the rule conditions; when the match is successful, trigger the generation of the corresponding crop supervision information text.

[0037] Use the calculated comprehensive fusion feature vector as input data and input it into a pre-trained crop status analysis model. The crop status analysis model outputs an evaluation result, such as a water shortage probability of 0.85 and a health probability of 0.9, etc. Compare and match the evaluation result output by the model with the rule conditions in the preset rule library. An exemplary rule is that if the model evaluation result = water shortage and the precipitation probability in the next 24 hours < 20%, then generate supervision information such as suggesting irrigation to supplement water, Figure 4 Shows the information received in the industrial brain system. Among them, the crop status analysis model includes but is not limited to support vector machine, multi-layer perceptron, or gradient boosting decision tree, etc. The present invention does not limit the specific structure of the crop status analysis model.

[0038] Embodiment 2. In this embodiment, a cloud supervision system for crops based on the Internet of Things is provided. The system includes the following modules: A data acquisition module, which is used to collect environmental data of the crop planting area through sensors. For the time-series sensing data collected by each sensor, an adaptive noise weight is determined by using the phase residual signal entropy value, and ensemble empirical mode decomposition is performed on the time-series sensing data through candidate component weighting to obtain multi-order intrinsic mode components; A feature fusion module, which is used to obtain crop growth environment context information including time, growth stage, and external meteorology; based on the context information, an attention network is used to calculate attention scores and normalize them to obtain context-aware fusion weights for each order of intrinsic mode components; An information output module, which is used to perform weighted fusion on the data feature vectors of each order of intrinsic mode components by using the context-aware fusion weights to obtain a comprehensive fusion feature; input the comprehensive fusion feature into a pre-trained crop status analysis model to obtain crop status information, and generate crop supervision information based on the crop status information and preset rules.

[0039] Preferably, the method for determining the adaptive noise weight by using the phase residual signal entropy value and performing ensemble empirical mode decomposition on the time-series sensing data through candidate component weighting to obtain multi-order intrinsic mode components is specifically as follows: For the time-series sensing data from one sensor, initialize it as the 0th-order residual signal , and set the decomposition order k = 1; loop and execute the following steps until the residual signal meets the preset decomposition termination condition: a) Calculate the sample entropy or permutation entropy value of the time segment of the (k - 1)th-order residual signal ; b) Map the entropy value obtained in step a) through a predefined function to obtain the adaptive noise weight required for the kth-order decomposition ; c) Generate N groups of independent Gaussian white noise sequences with a mean of 0 and a standard deviation of ; add each group of noise sequences to the (k - 1)th-order residual signal to form N groups of noise-assisted signals; d) Apply the empirical mode decomposition algorithm to each of the N groups of noise-assisted signals to extract N candidate kth-order intrinsic mode components , where i is a positive integer, and ; e) Calculate the orthogonality index between each candidate intrinsic mode component and the (k - 1)th-order residual signal ; f) According to the orthogonality index calculated in step e), perform weighted averaging on the N candidate intrinsic mode components to calculate the final kth-order intrinsic mode component ; The weight of the weighted average is proportional to the orthogonality index; g) Calculate the k-th order residual signal ; h) Increment the decomposition order k by 1; After the loop ends, all the intrinsic mode components are obtained as the decomposition result of the sensor time series data.

[0040] Preferably, based on the context information, an attention network is used to calculate the attention score and normalize it to obtain the context-aware fusion weight of each order of the intrinsic mode component. Specifically: Input the context information into the embedding layer or the encoder network to generate a context feature vector Q; Input each component in the set of final intrinsic mode components of all orders decomposed from all sensors into its corresponding embedding layer or encoder network to generate a set of intrinsic mode component feature vectors and , where M is the total number of all intrinsic mode components generated by all sensors, and are the key vector and value vector of the j-th intrinsic mode component respectively; Use the context feature vector Q as the query, and calculate the attention score between the query Q and each key ; Apply the Softmax function to normalize all the attention scores to obtain the final context-aware fusion weight , where is the fusion weight corresponding to the j-th intrinsic mode component.

[0041] Preferably, the weighted fusion is performed on the data feature vectors of each order of the intrinsic mode components by using the context-aware fusion weight to obtain the comprehensive fusion feature. Specifically: Multiply the obtained fusion weight corresponding to each intrinsic mode component by its corresponding value vector to perform a scalar multiplication operation to obtain the weighted value vector ; Sum or concatenate all the weighted value vectors to generate a comprehensive fusion feature vector.

[0042] Preferably, the context information of the crop growth environment including time, growth stage, and external meteorology is obtained as follows: Obtain the current date and time information; based on the crop planting date and the preset crop growth cycle model, judge and record the current growth stage of the crop; Request current and short-term future meteorological data of a specified geographical location from a third-party meteorological service platform through a network interface, including but not limited to temperature, humidity, precipitation, wind speed, and light intensity; Integrate the obtained time information, growth stage identifier, and structured meteorological data into a context information.

[0043] Preferably, inputting the comprehensive fusion feature into a pre-trained crop status analysis model to obtain crop status information, and generating crop supervision information based on the crop status information and preset rules, specifically: Use the comprehensive fusion feature vector as the input data of the crop status analysis model, and the crop status analysis model outputs an evaluation result; According to a pre-set rule library, match the evaluation result output by the model with the rule conditions; when the match is successful, trigger the generation of the corresponding crop supervision information text.

[0044] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of adding a necessary general hardware platform, and of course, can also be implemented by a combination of hardware and software. Based on such an understanding, the above technical solutions, in essence, or the part that contributes to the prior art can be embodied in the form of a computer product. The present invention can be implemented in the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes.

[0045] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Other embodiments can also be adopted; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A crop cloud monitoring method based on the Internet of Things, characterized in that: The method comprises the following steps: Collect environmental data of the crop planting area through sensors, determine adaptive noise weights for time series sensing data collected by each sensor using the stage residual signal entropy value, and perform ensemble empirical mode decomposition on the time series sensing data through candidate component weighting to obtain multi-order intrinsic mode components; Acquire the context information of crop growth environment including time, growth stage, and external weather; based on the context information, use the attention network to calculate the attention score and normalize it to obtain the context-aware fusion weights of each order of intrinsic modal components; The context-aware fusion weights are used to perform weighted fusion on the data feature vectors of the intrinsic modal components of each order to obtain comprehensive fusion features; the comprehensive fusion features are input into a pre-trained crop state analysis model to obtain crop state information, and crop supervision information is generated based on the crop state information and preset rules.

2. The method according to claim 1, characterized in that The adaptive noise weight is determined by using the entropy value of the residual signal in the stage, and the set empirical mode decomposition is performed on the time series sensor data by weighting the candidate components to obtain multi-order intrinsic mode components, specifically: For the time series sensor data from a sensor, initialize it as the 0th order residual signal , set the decomposition order k=1; loop through the following steps until the residual signal meets the preset decomposition termination condition: a) Calculate the k-1th order residual signal The sample entropy or permutation entropy value of the time segment; b) Mapping the entropy value obtained in step a) through a predefined function to obtain the adaptive noise weight required for the k-th order decomposition ; c) Generate N independent groups with mean 0 and standard deviation Gaussian white noise sequence; each set of noise sequence is added to the k-1th order residual signal On the other hand, N groups of noise auxiliary signals are formed; d) applying the empirical mode decomposition algorithm to the N groups of noise auxiliary signals respectively to extract N candidate k-th order intrinsic mode components , where i is a positive integer, and ; e) Calculate each candidate eigenmode component and the k-1th order residual signal Orthogonality index between ; f) Based on the orthogonality index calculated in step e), Perform weighted averaging to calculate the final k-th order eigenmode component ; The weight of the weighted average is proportional to the orthogonality index; g) Calculate the kth order residual signal ; h) The decomposition order k increases by 1; After the cycle is completed, all intrinsic mode components are obtained as the decomposition results of the sensor time series data.

3. The method according to claim 1, characterized in that Based on the context information, the attention score is calculated and normalized using the attention network to obtain the context-aware fusion weights of each order of intrinsic modal components, specifically: Input the context information into the embedding layer or encoder network to generate the context feature vector Q; Each component in the final intrinsic mode component set of all orders obtained by decomposing all sensors is input into its corresponding embedding layer or encoder network to generate a set of intrinsic mode component feature vectors and , where M is the total number of all intrinsic mode components generated by all sensors, and are the key vector and value vector of the j-th eigenmode component respectively; Take the context feature vector Q as the query and calculate the correlation between the query Q and each key The attention score between ; For all attention scores Apply the Softmax function for normalization to obtain the final context-aware fusion weights ,in is the fusion weight corresponding to the j-th eigenmode component.

4. The method according to claim 1, characterized in that The weighted fusion of the data feature vectors of each order of intrinsic modal components by using the context-aware fusion weights to obtain a comprehensive fusion feature is specifically: The fusion weight corresponding to each eigenmode component obtained The corresponding value vector Perform scalar multiplication to obtain a weighted value vector ; All weighted value vectors Perform vector summation or vector concatenation operations to generate a comprehensive fusion feature vector.

5. The method according to claim 1, characterized in that The acquisition of context information of the crop growth environment including time, growth stage, and external weather conditions is specifically as follows: Obtain the current date and time information; determine and record the current growth stage of the crop based on the crop planting date and the preset crop growth cycle model; Request current and short-term future meteorological data of a specified geographic location from a third-party meteorological service platform through a network interface, including but not limited to temperature, humidity, precipitation, wind speed, and light intensity; The acquired time information, growth stage identifiers, and structured meteorological data are integrated into a context information.

6. The method according to claim 1, characterized in that The comprehensive fusion feature is input into the pre-trained crop state analysis model to obtain crop state information, and the crop supervision information is generated based on the crop state information and preset rules, specifically: The comprehensive fusion feature vector is used as input data of the crop state analysis model, and the crop state analysis model outputs the evaluation result; According to the pre-set rule base, the evaluation results output by the model are matched with the rule conditions; when the match is successful, the corresponding crop regulatory information text is generated.

7. A crop cloud monitoring system based on the Internet of Things, characterized in that: The system includes the following modules: A data acquisition module is used to collect environmental data of a crop planting area through sensors, and for each time series sensor data collected by the sensor, an adaptive noise weight is determined using a stage residual signal entropy value, and a set empirical mode decomposition is performed on the time series sensor data through candidate component weighting to obtain multi-order intrinsic mode components; The feature fusion module is used to obtain the context information of the crop growth environment including time, growth stage, and external weather; based on the context information, the attention score is calculated and normalized using the attention network to obtain the context-aware fusion weights of each order of intrinsic modal components; An information output module, used for performing weighted fusion on the data feature vectors of the intrinsic modal components of each order using the context-aware fusion weights to obtain a comprehensive fusion feature; The comprehensive fusion features are input into a pre-trained crop state analysis model to obtain crop state information, and crop supervision information is generated based on the crop state information and preset rules.

8. The system according to claim 7, characterized in that The adaptive noise weight is determined by using the entropy value of the residual signal in the stage, and the set empirical mode decomposition is performed on the time series sensor data by weighting the candidate components to obtain multi-order intrinsic mode components, specifically: For the time series sensor data from a sensor, initialize it as the 0th order residual signal , set the decomposition order k=1; loop through the following steps until the residual signal meets the preset decomposition termination condition: a) Calculate the k-1th order residual signal The sample entropy or permutation entropy value of the time segment; b) Mapping the entropy value obtained in step a) through a predefined function to obtain the adaptive noise weight required for the k-th order decomposition ; c) Generate N independent groups with mean 0 and standard deviation Gaussian white noise sequence; each set of noise sequence is added to the k-1th order residual signal On the other hand, N groups of noise auxiliary signals are formed; d) applying the empirical mode decomposition algorithm to the N groups of noise auxiliary signals respectively to extract N candidate k-th order intrinsic mode components , where i is a positive integer, and ; e) Calculate each candidate eigenmode component and the k-1th order residual signal Orthogonality index between ; f) Based on the orthogonality index calculated in step e), Perform weighted averaging to calculate the final k-th order eigenmode component ; The weight of the weighted average is proportional to the orthogonality index; g) Calculate the kth order residual signal ; h) The decomposition order k increases by 1; After the cycle is completed, all intrinsic mode components are obtained as the decomposition results of the sensor time series data.

9. The system according to claim 7, characterized in that Based on the context information, the attention score is calculated and normalized using the attention network to obtain the context-aware fusion weights of each order of intrinsic modal components, specifically: Input the context information into the embedding layer or encoder network to generate the context feature vector Q; Each component in the final intrinsic mode component set of all orders obtained by decomposing all sensors is input into its corresponding embedding layer or encoder network to generate a set of intrinsic mode component feature vectors and , where M is the total number of all intrinsic mode components generated by all sensors, and are the key vector and value vector of the j-th eigenmode component respectively; Take the context feature vector Q as the query and calculate the correlation between the query Q and each key The attention score between ; For all attention scores Apply the Softmax function for normalization to obtain the final context-aware fusion weights ,in is the fusion weight corresponding to the j-th eigenmode component.

10. The system according to claim 7, characterized in that The weighted fusion of the data feature vectors of each order of intrinsic modal components by using the context-aware fusion weights to obtain a comprehensive fusion feature is specifically: The fusion weight corresponding to each eigenmode component obtained , and its corresponding value vector Perform scalar multiplication to obtain a weighted value vector ; All weighted value vectors Perform vector summation or vector concatenation operations to generate a comprehensive fusion feature vector.

Images