A water-saving intelligent agent interaction control method and system based on semantic intention prediction

Abstract

The application relates to the technical field of intelligent agent interaction, and discloses a water-saving intelligent agent interaction control method and system based on semantic intention prediction, which has the technical scheme as follows: according to semantic text and self-attention weights of a user, a three-level intention set and a deep feature vector are obtained; according to current environment data and the deep feature vector, a coupling time sequence diagram and a dynamic constraint set are obtained; according to the coupling time sequence diagram, the three-level intention set and the dynamic constraint set, a candidate water-saving scheme is obtained; and according to the candidate water-saving scheme, the three-level intention set is updated to obtain a water-saving suggestion. The three-level intention set is extracted from fuzzy semantics layer by layer through the self-attention weights, and the full-dimension demand of the user is covered; then, the dynamic constraint set which is suitable for real-time scenes is generated in combination with the coupling time sequence diagram; finally, the dynamic constraint set is updated through the matching degree of intention and constraint, and the optimal water-saving suggestion which meets the user demand and conforms to the actual scene is formed.

Description

A Water-Saving Intelligent Agent Interactive Control Method and System Based on Semantic Intent Prediction Technical Field

[0001] This invention relates to the field of intelligent agent interaction technology, and more specifically to a water-saving intelligent agent interaction control method and system based on semantic intent prediction. Background Technology

[0002] In intelligent decision-making scenarios, such as industrial process optimization, smart home control, and public service scheduling, user needs are often presented in a fuzzy semantic form, such as equipment energy consumption being too high or a room being too stuffy. At the same time, the scenarios have multi-parameter dynamic coupling, such as the temporal correlation of temperature, pressure, and production capacity in industrial production lines, and the mutual influence of humidity, ventilation, and energy consumption in smart homes. Furthermore, there is an implicit correlation between needs and scenario constraints (such as equipment capabilities, resource limitations, and security rules).

[0003] Current intelligent systems rely heavily on pre-defined rule bases, historical data annotations, or fixed intent templates for decision-making. However, in real-world scenarios, implicit demands (such as simple operation requiring no manual intervention, and time-series delay coupling of parameters, where delayed coupling refers to the need for multiple iterations of related parameters to respond after a parameter is adjusted) and dynamic conflicts between demands and constraints are difficult to cover through pre-defined logic. At the same time, real-time fluctuations in scenario parameters, such as changes in industrial equipment load or sudden changes in urban traffic flow, further increase the complexity of decision-making. Therefore, existing technologies have shortcomings. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a water-saving intelligent agent interactive control method and system based on semantic intent prediction. By deeply mining user intent and matching intent with constraints, the method generates optimal water-saving suggestions that meet user needs and conform to actual scenarios.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention provides a water-saving intelligent agent interaction control method based on semantic intent prediction, comprising:

[0007] Based on the user's semantic text and self-attention weights, a three-level intent set and a deep feature vector are obtained;

[0008] Based on the current environmental data and the deep feature vector, a coupled time series diagram and a dynamic constraint set are obtained. The current environmental data includes soil moisture, irrigation flow rate, and temperature.

[0009] Based on the coupling timing diagram, the three-level intent set, and the dynamic constraint set, candidate water-saving schemes are obtained;

[0010] The third-level intent set is updated based on the candidate water-saving schemes to obtain water-saving suggestions;

[0011] The step of obtaining the coupled time series diagram and dynamic constraint set based on the current environmental data and the deep feature vector includes:

[0012] Based on the current environmental data, a core feature set and time-division data packets are obtained;

[0013] The static constraint logic and logic labels are obtained based on the core feature set.

[0014] Based on the core feature set, the time-division data packet, the static constraint logic and logic label, and the deep feature vector, a coupled timing diagram and a dynamic constraint set are obtained;

[0015] Specifically, based on the core feature set, the time-division data packet, the static constraint logic and logical labels, and the deep feature vector, a coupled timing diagram and a dynamic constraint set are obtained, including:

[0016] Based on the core feature set and the time-division data packets, a time-series feature vector is obtained;

[0017] Based on the value of each dimension in the time-series feature vector, the marginal probability corresponding to each value and the joint probability corresponding to each pair of values ​​are obtained;

[0018] Based on the marginal probabilities and the joint probabilities, the mutual information and entropy of each pair of features under different delays are calculated;

[0019] Based on the mutual information and entropy, the coupling strength of each pair of features under different delays is obtained;

[0020] Feature pairs are selected based on the coupling strength to obtain a coupling time sequence diagram;

[0021] The dynamic constraint set is obtained based on the coupling timing diagram, the static constraint logic and logic labels, and the deep feature vector.

[0022] As a further improvement of the present invention, the step of obtaining a three-level intent set and a deep feature vector based on the user's semantic text and self-attention weights includes:

[0023] Based on the semantic text and self-attention weights, explicit intent labels are obtained;

[0024] Based on the explicit intent labels and the generative adversarial network, a set of implicit intent candidates is obtained;

[0025] Based on the implicit intent candidate set, the explicit intent label, and the semantic text, a three-level intent set and a deep feature vector are obtained.

[0026] As a further improvement of the present invention, a three-level intent set and a deep feature vector are obtained based on the implicit intent candidate set, the explicit intent tags, and the semantic text, including:

[0027] Based on the implicit intent candidate set, the explicit intent labels, and the semantic text, a fused text sequence is obtained;

[0028] Based on the fused text sequence and variational autoencoder, the deep feature vector and text ambiguity are obtained;

[0029] The three-level intent set is obtained based on the implicit intent candidate set, the explicit intent label, the deep feature vector, and the text ambiguity.

[0030] As a further improvement of the present invention, the dynamic constraint set is obtained based on the coupling timing diagram, the static constraint logic and logic labels, and the deep feature vector, including:

[0031] Based on the coupling timing diagram and the static constraint logic, the feature dependency chain is obtained;

[0032] Based on the feature dependency chain and the deep feature vector, a first dynamic constraint set is obtained;

[0033] Based on the first dynamic constraint set and the logical label, a second dynamic constraint set is obtained;

[0034] The dynamic constraint set is obtained based on the coupling timing diagram and the second dynamic constraint set.

[0035] As a further improvement of the present invention, candidate water-saving schemes are obtained based on the coupled timing diagram, the three-level intent set, and the dynamic constraint set, including:

[0036] Based on the coupling sequence diagram, the three-level intent set, and the dynamic constraint set, the matching degree and conflict points between intents and constraints are obtained;

[0037] The implicit intent is updated based on the conflict points to obtain candidate water-saving solutions.

[0038] As a further improvement of the present invention, the three-level intent set is updated according to the candidate water-saving schemes to obtain water-saving suggestions, including:

[0039] Based on the candidate water-saving schemes and the updated implicit intent, the updated dynamic constraint set is obtained;

[0040] Update the third-level intent set according to the updated dynamic constraint set;

[0041] Based on the updated dynamic constraint set and the updated three-level intent set, water-saving suggestions are obtained.

[0042] As a further improvement of the present invention, water-saving suggestions are obtained based on the updated dynamic constraint set and the updated three-level intent set, including:

[0043] Based on the updated dynamic constraint set and the updated three-level intent set, the updated matching degree between intent and constraint is obtained;

[0044] The water-saving suggestion is obtained based on the updated matching degree between the intent and the constraints, the updated dynamic constraint set, and the candidate water-saving schemes.

[0045] This invention provides a water-saving intelligent agent interaction system based on semantic intent prediction, comprising:

[0046] The intent agent is used to obtain a three-level intent set and a deep feature vector based on the user's semantic text and self-attention weights.

[0047] A constrained agent is used to obtain a coupled temporal graph and a dynamic constraint set based on the current environmental data and the deep feature vector. The current environmental data includes soil moisture, irrigation flow rate, and temperature.

[0048] A decision-making agent is used to obtain candidate water-saving schemes by interacting with an intention agent and a constraint agent, combining the coupled timing diagram, the three-level intention set, and the dynamic constraint set, and to update the three-level intention set according to the candidate water-saving schemes to obtain water-saving suggestions.

[0049] The step of obtaining the coupled time series diagram and dynamic constraint set based on the current environmental data and the deep feature vector includes:

[0050] Based on the current environmental data, a core feature set and time-division data packets are obtained;

[0051] The static constraint logic and logic labels are obtained based on the core feature set.

[0052] Based on the core feature set, the time-division data packet, the static constraint logic and logic label, and the deep feature vector, a coupled timing diagram and a dynamic constraint set are obtained;

[0053] Specifically, based on the core feature set, the time-division data packet, the static constraint logic and logical labels, and the deep feature vector, a coupled timing diagram and a dynamic constraint set are obtained, including:

[0054] Based on the core feature set and the time-division data packets, a time-series feature vector is obtained;

[0055] Based on the value of each dimension in the time-series feature vector, the marginal probability corresponding to each value and the joint probability corresponding to each pair of values ​​are obtained;

[0056] Based on the marginal probabilities and the joint probabilities, the mutual information and entropy of each pair of features under different delays are calculated;

[0057] Based on the mutual information and entropy, the coupling strength of each pair of features under different delays is obtained;

[0058] Feature pairs are selected based on the coupling strength to obtain a coupling time sequence diagram;

[0059] The dynamic constraint set is obtained based on the coupling timing diagram, the static constraint logic and logic labels, and the deep feature vector.

[0060] This invention first mines user intent layer by layer through self-attention weights to form a three-level intent set. Then, it combines the current environmental data to generate a dynamic constraint set to adapt to the rules of the real-time scenario. Next, it detects the matching degree between user intent and constraints to trigger adversarial correction between intent and dynamic constraint set. Finally, it obtains the optimal water-saving suggestion that meets both user needs and actual scenario, solving the problems of existing technologies that are difficult to accurately identify user needs and generate effective suggestions when parameters change in real time. Attached Figure Description

[0061] Figure 1 is a schematic diagram of the method steps of the present invention;

[0062] Figure 2 is a schematic diagram of the steps for generating a three-level intent set;

[0063] Figure 3 is a schematic diagram of the steps to generate the coupling timing diagram and dynamic constraint set;

[0064] Figure 4 is a schematic diagram of the coupling timing diagram;

[0065] Figure 5 is a schematic diagram of the steps to generate candidate water-saving schemes. Detailed Implementation

[0066] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof.

[0067] The term "and / or" in the following text is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0068] This application provides a water-saving intelligent agent interaction system based on semantic intent prediction, including:

[0069] The intent agent is used to obtain a three-level intent set and a deep feature vector based on the user's semantic text and self-attention weights.

[0070] A constrained agent is used to obtain a coupled temporal graph and a dynamic constraint set based on current environmental data and deep feature vectors. The current environmental data includes soil moisture, irrigation flow rate, and temperature.

[0071] The decision-making agent is used to obtain candidate water-saving schemes by interacting with the intention agent and the constraint agent, combining the coupled timing graph, the three-level intention set and the dynamic constraint set, and to update the three-level intention set according to the candidate water-saving schemes to obtain water-saving suggestions.

[0072] The step of obtaining the coupled time series diagram and dynamic constraint set based on the current environmental data and the deep feature vector includes:

[0073] Based on the current environmental data, a core feature set and time-division data packets are obtained;

[0074] The static constraint logic and logic labels are obtained based on the core feature set.

[0075] Based on the core feature set, the time-division data packet, the static constraint logic and logic label, and the deep feature vector, a coupled timing diagram and a dynamic constraint set are obtained;

[0076] Specifically, based on the core feature set, the time-division data packet, the static constraint logic and logical labels, and the deep feature vector, a coupled timing diagram and a dynamic constraint set are obtained, including:

[0077] Based on the core feature set and the time-division data packets, a time-series feature vector is obtained;

[0078] Based on the value of each dimension in the time-series feature vector, the marginal probability corresponding to each value and the joint probability corresponding to each pair of values ​​are obtained;

[0079] Based on the marginal probabilities and the joint probabilities, the mutual information and entropy of each pair of features under different delays are calculated;

[0080] Based on the mutual information and entropy, the coupling strength of each pair of features under different delays is obtained;

[0081] Feature pairs are selected based on the coupling strength to obtain a coupling time sequence diagram;

[0082] The dynamic constraint set is obtained based on the coupling timing diagram, the static constraint logic and logic labels, and the deep feature vector.

[0083] Based on the above system, as shown in Figure 1, this application embodiment provides a water-saving intelligent agent interaction control method based on semantic intent prediction, including:

[0084] Based on the user's semantic text and self-attention weights, a three-level intent set and a deep feature vector are obtained;

[0085] Based on the current environmental data and deep feature vectors, a coupled time series diagram and a dynamic constraint set are obtained. The current environmental data includes soil moisture, irrigation flow rate, and temperature.

[0086] Based on the coupling timing diagram, the three-level intent set, and the dynamic constraint set, candidate water-saving schemes are obtained;

[0087] The three-level intent set is updated based on the candidate water-saving solutions to obtain water-saving recommendations.

[0088] Specifically, the method provided in this embodiment is applied to the above-mentioned water-saving intelligent agent interaction system. The steps in the method are executed by each intelligent agent or by interaction between multiple intelligent agents. Each intelligent agent includes multiple units, wherein the intention intelligent agent includes explicit intention unit, implicit intention unit and deep intention unit, the constraint intelligent agent includes feature extraction unit, static constraint unit and dynamic constraint unit, and the decision intelligent agent includes matching degree calculation unit and scheme generation unit.

[0089] This embodiment mines user intent layer by layer from attention weights to form a three-level intent set. Then, it combines the current environmental data to generate a dynamic constraint set to adapt to the rules of the real-time scenario. Next, it detects the matching degree between user intent and constraints to trigger adversarial correction between intent and dynamic constraint set. Finally, it obtains the optimal water-saving suggestion that meets both user needs and actual scenario, which solves the problem that the existing technology is difficult to accurately identify user needs and difficult to generate effective suggestions when parameters change in real time.

[0090] Furthermore, this embodiment provides a step for obtaining a coupled time series diagram and a dynamic constraint set based on current environmental data and deep feature vectors, including:

[0091] Based on the current environmental data, obtain the core feature set and time-division data packets;

[0092] The static constraint logic and logic labels are obtained based on the core feature set;

[0093] Based on the core feature set, time-division data packets, static constraint logic and logical labels, and deep feature vectors, a coupled timing diagram and a dynamic constraint set are obtained.

[0094] The environmental data is determined based on actual conditions. For example, in the agricultural water use scenario of this embodiment, the environmental data includes multiple features such as soil moisture, irrigation flow, wind speed, and temperature corresponding to each sampling point within a preset time period. For each feature in each environmental data, the data is sorted according to the sampling time corresponding to the sampling point to obtain the time series data corresponding to each feature. Then, the core features are determined based on the current scenario. For example, in this embodiment, water-saving suggestions need to be made for water use, so the core feature is the irrigation flow. Then, the correlation coefficient between the irrigation flow sequence and each other sequence is determined, and features with correlation coefficients greater than the preset coefficient are obtained. Finally, the sequence corresponding to the feature, the irrigation flow sequence, and the corresponding correlation coefficient are packaged as the core feature set. This embodiment does not impose any restrictions on the preset coefficient.

[0095] Furthermore, since real-time data is a continuous long-term time-series stream, directly using the full data for modeling would obscure the differences in patterns across different time periods. In this embodiment, each sequence is divided into multiple data packets with smaller time windows. The set of data packets with smaller time windows is called time-division data packets.

[0096] As shown in Figure 3, the steps of generating the core feature set and the time-division data packet are performed by the feature extraction unit. The feature extraction unit needs to send the core feature set to the static constraint unit and the dynamic constraint unit respectively, and send the time-division data packet to the dynamic constraint unit.

[0097] After the static constraint unit receives the core feature set, it needs to perform Granger causality tests on any two sequences in the core feature set. Compared with the correlation coefficient, the Granger causality test can further determine whether there is a causal relationship between the two features, thereby obtaining the causal chain between the core features, and generating static constraint logic based on the causal chain. For example, a static constraint logic is that soil moisture needs to be maintained within a certain range to ensure crop growth. The static constraint logic is directional (the range needs to be maintained), but the specific threshold is not yet clear. Only the core logic of the constraint is determined. Its corresponding logic label includes specific prohibited values. For example, if the lower limit of soil moisture for crop growth in farmland is 40% (below which will cause drought) and the upper limit is 60% (above which will cause root waterlogging), then its corresponding logic label is S<40% and S>60%, where S represents soil moisture. The values ​​in this embodiment are only examples. Those skilled in the art can set them according to the actual situation. This embodiment does not limit them.

[0098] Furthermore, this embodiment provides a step for obtaining a coupled timing diagram and a dynamic constraint set based on a core feature set, time-division data packets, static constraint logic and logical labels, and deep feature vectors, including:

[0099] Based on the core feature set and time-division data packets, a time-series feature vector is obtained;

[0100] Based on the value of each dimension in the temporal feature vector, the marginal probability corresponding to each value and the joint probability corresponding to each pair of values ​​are obtained;

[0101] Based on the marginal probability and joint probability, the mutual information and entropy of each pair of features under different delays are calculated;

[0102] Based on mutual information and entropy, the coupling strength of each pair of features under different delays is obtained;

[0103] Feature pairs are selected based on coupling strength to obtain the coupling time sequence diagram;

[0104] Based on the coupling timing diagram, static constraint logic and logic labels, and deep feature vectors, the dynamic constraint set is obtained.

[0105] Specifically, the steps of obtaining the coupling time sequence diagram and the dynamic constraint set are performed by the dynamic constraint unit. After the dynamic constraint unit receives the core feature set and the time-division data packet, it standardizes and concatenates the sequence corresponding to each feature in the core feature set to obtain the time sequence feature vector corresponding to each time step. Then, based on the time sequence feature vector corresponding to each time step, it calculates the coupling strength of each pair of features under different delays.

[0106] For example, assuming the core feature set includes three features: soil moisture, irrigation flow rate, and temperature, and multiple time steps are obtained according to the time-division data packet segmentation rules, the temporal feature vector corresponding to each time step is obtained from the normalized soil moisture, irrigation flow rate, and temperature corresponding to the current time step. Any two features can form a feature pair. For example, for the three features of soil moisture, irrigation flow rate, and temperature, three feature pairs can be formed. Taking the feature pair of soil moisture S and irrigation flow rate F as an example, different delays can be calculated. The coupling strength is determined by first calculating the value of the irrigation flow rate F at time t. With soil moisture S Value at time The degree of correlation, and The dimensions are the same, and the degree of correlation can be specifically measured through mutual information. Mutual information can be measured using joint probability and marginal probability, for example, when hour, The moment corresponds to the first time step. The time corresponding to the first At this time step, This represents the normalized irrigation flow rate corresponding to the temporal feature vector at the first time step. For the first The normalized soil moisture content corresponding to the temporal feature vector at each time step has a marginal probability equal to the frequency of a particular value of a single feature. There are multiple marginal probabilities, each marginal probability corresponds to In the context of a dimension, the marginal probability for each dimension can be obtained by counting the number of times the corresponding value for that dimension appears in historical data, and then using the ratio of that number of occurrences to the total number of occurrences in historical data as the marginal probability for that dimension. Similarly, we can obtain... and The marginal probability corresponding to each dimension; the joint probability is the probability of each pair of values ​​occurring. For example, assume... and The dimensions are all 3, respectively from and We select one value from the given range, resulting in 3 × 3 = 9 possible values. Each value pair corresponds to a joint probability. The joint probability for each value pair can be obtained by counting the number of times the corresponding value appears in historical data, and using the ratio of this count to the total number of occurrences in historical data as the joint probability. Then, through each marginal probability and each joint probability, we can obtain the mutual information. .

[0107] Then, based on the above marginal probabilities, we can calculate them separately. and corresponding entropy and Therefore, the coupling strength is obtained as follows:

[0108]

[0109] Then it kept changing The value of , The range of values ​​is determined by the number of time steps, yielding soil moisture S and irrigation flow rate F at different delays. The coupling strength under different delays. Similarly, for other feature pairs, the coupling strength can also be obtained under different delays. The coupling strength is determined by first setting a preset coupling strength, then iterating through all coupling strengths to identify feature pairs with coupling strengths greater than a preset strength. These are then considered strong coupling pairs. Based on the delays corresponding to these strong coupling pairs, a coupling timing diagram is obtained. This embodiment does not impose any restrictions on the value of the preset coupling strength. The final coupling timing diagram is a schematic diagram that visually illustrates the feature pairs and delay durations. It is assumed that in... At that time, there is a strong coupling with irrigation flow. and soil moisture ,exist At that time, there is a strong coupling with irrigation flow. and temperature The resulting coupling timing sequence is shown in Figure 4.

[0110] This embodiment measures the influence relationship between feature pairs by coupling strength, for example... At that time, there is a strong coupling with irrigation flow. and soil moisture This indicates that changes in irrigation flow require a two-time-step delay before significantly affecting soil moisture. This embodiment quantifies the delay duration of feature pairs, transforming the ambiguous relationship between parameters into a time-series dependency rule with clear numerical values. This provides a benchmark for the order of feature influence in subsequent dynamic constraint adjustments, ensuring the accuracy of water-saving recommendations.

[0111] Furthermore, this embodiment provides a step for obtaining a three-level intent set and a deep feature vector based on the user's semantic text and self-attention weights, including:

[0112] Based on semantic text and self-attention weights, explicit intent labels are obtained;

[0113] Based on the explicit intent labels and the generative adversarial network, a candidate set of implicit intents is obtained;

[0114] Based on the implicit intent candidate set, explicit intent labels, and semantic text, a three-level intent set and a deep feature vector are obtained.

[0115] For example, assuming a user inputs the semantic text "farmland irrigation uses a lot of water" in an agricultural water use scenario, the first step is to obtain explicit intent labels based on the semantic text. Specifically, after preprocessing the semantic text, multiple augmented samples with semantic consistency with the semantic text can be generated through contrastive learning. For example, one augmented sample might include ["farmland", "irrigation", "high water consumption"]. Then, the BERT model is used to generate multiple word vectors corresponding to each sample, with each word vector corresponding to a core word in the text. Finally, all word vectors are clustered. This embodiment does not limit the number of clusters. After clustering, each cluster corresponds to a semantic topic, ensuring that the intent generated based on these clusters is accurate. The graph covers all the core semantics of the text. For example, the first cluster corresponds to the scene theme, and its core text words are "farmland." The second cluster corresponds to the behavior theme, and its core text words are "watering" and "irrigation." The third cluster corresponds to the theme of "requests," and its core text words are "high water consumption." Then, based on the combination patterns of each cluster, the high-frequency co-occurrence relationships of each cluster under historical water-saving scenarios are statistically analyzed. Multiple candidate intents are generated by combining these with interaction intents, including suggestions, queries, and complaints. For example, based on the first and third clusters and the suggestion interaction intent, a candidate intent of "farmland water-saving suggestion" can be obtained; based on the first and second clusters and the query interaction intent, a candidate intent of "querying farmland water consumption" can be obtained. Next, each candidate intent needs to be encoded to obtain a feature vector. Then, the similarity between each candidate intent feature vector and the semantic vector corresponding to the semantic text is calculated. The candidate intent with the highest similarity is determined, and an explicit intent label is generated. The explicit intent label includes the content of the candidate intent and its corresponding similarity. This candidate intent is denoted as the explicit intent.

[0116] After obtaining the explicit intent labels, a set of implicit intent candidates needs to be obtained through a generative adversarial network. Specifically, the vector corresponding to the explicit intent is first concatenated with the vector of each intent seed and input into the generator. The intent seeds are generated based on high-frequency appeal words in historical water-saving scenarios. For example, intent seeds can be "do not affect", "operation", "cost", "efficiency", "experience", etc. The generator will expand each intent seed into a complete implicit intent based on the autoregression of the Transformer. For example, the implicit intent generated by the intent seed "do not affect" can be "do not affect crop growth", "do not affect irrigation efficiency", "do not affect farmland yield", and the implicit intent generated by the intent seed "operation" can be "simple operation", "few operation steps", "no human intervention required". For each generated implicit intent, the logical matching probability between it and the explicit intent is calculated using the loss function in the discriminator. This embodiment does not restrict the calculation method of the logical matching probability. For example, labels can be added to implicit and explicit intents, and then the loss function value is calculated using the binary cross-entropy loss function. The loss function value is then recorded as the logical matching probability. Finally, implicit intents with logical matching probabilities less than a preset probability are selected, and the set of these implicit intents is used as the implicit intent candidate set.

[0117] As shown in Figure 2, the step of obtaining the explicit intent label is executed by the explicit intent unit, and the step of obtaining the implicit intent candidate set is executed by the implicit intent unit. The explicit intent unit needs to send the explicit intent label to the implicit intent unit and send instructions to the constrained agent to enable it to obtain the current environmental data. The implicit intent unit needs to send the implicit intent candidate set to the deep intent unit to generate the three-level intent set, and send the implicit intent candidate set to the decision agent to generate candidate water-saving schemes.

[0118] This embodiment generates explicit and implicit intents based on user semantic text. The explicit intent anchors the user's surface-level needs, while the candidate intents uncover implicit needs that cover the user's unexpressed requirements, thus solving the problem of information loss due to ambiguous semantics and ensuring the accuracy of water-saving suggestions.

[0119] Furthermore, this embodiment provides a step for obtaining a three-level intent set and a deep feature vector based on a latent intent candidate set, explicit intent labels, and semantic text, including:

[0120] Based on the implicit intent candidate set, explicit intent labels, and semantic text, a fused text sequence is obtained;

[0121] Based on the fused text sequence and variational autoencoder, deep feature vectors and text ambiguity are obtained;

[0122] A three-level intent set is obtained based on the implicit intent candidate set, explicit intent labels, deep feature vectors, and the text ambiguity.

[0123] The step of obtaining the three-level intent set and deep feature vector based on the implicit intent candidate set, explicit intent label and semantic text is performed by the deep intent unit. Then the deep intent unit sends the three-level intent set to the decision agent and sends the deep feature vector to the dynamic constraint unit.

[0124] Specifically, the process begins by concatenating the text corresponding to the implicit intent in the implicit intent candidate set, the candidate intent in the explicit intent label, and the semantic text into a single sentence according to semantic logic. This sentence serves as the fused text sequence. The fused text sequence is then broken down into lexical units, resulting in a lexical unit sequence. This sequence is input into the encoder, which first identifies fuzzy words within the lexical unit sequence. For example, fuzzy words can be determined by comparing a pre-defined fuzzy word list with the similarity of each lexical unit in the sequence. For each lexical unit, a weight is assigned using an attention formula. This weighting is adjusted based on the similarity to the fuzzy word, increasing its weight percentage. Finally, after obtaining the weight for each lexical unit, each lexical unit is multiplied by its weight to obtain a weighted lexical unit vector sequence. Furthermore, the ratio of the sum of the weights corresponding to each fuzzy word to the sum of the weights corresponding to each lexical unit is recorded as the text fuzziness.

[0125] The weighted word vector sequence is then input into the encoder, which consists of two LSTMs and two fully connected layers. LSTM (Long Short-Term Memory) is a model for processing sequence data and can capture the temporal relationships between words. The first LSTM needs to perform preliminary encoding on the weighted word vector sequence and output the hidden state at each time step. The second LSTM needs to further extract the global semantic hidden state based on the hidden state in the first LSTM. The global semantic hidden state output by the LSTM is a high-dimensional vector, which then needs to be passed through two parallel fully connected layers. One fully connected layer maps the global semantic hidden state to the mean vector of the latent variables, and the other fully connected layer maps the global hidden state to the variance vector of the latent variables.

[0126] This embodiment takes into account that semantics in reality are continuously changing. For example, "wasteful of water" and "consuming a lot of water" are similar semantics, and "simple to operate" and "no human intervention required" are related semantics. Probability distribution is a continuous numerical space, and different points in the distribution can represent similar but different semantics. Therefore, this embodiment assumes that text semantics can be represented by a continuous probability distribution. Moreover, using probability distribution can naturally adapt to the continuity and ambiguity of semantics, avoiding the problem of solidifying semantics into a single vector, which would fail to reflect the fluctuation and association of semantics. Furthermore, the distribution of semantics is clustered, that is, similar semantics will be concentrated around a certain center, and the bell curve of the normal distribution conforms to this rule. Therefore, this embodiment sets a decoder to generate a mean vector. The center and variance vectors corresponding to the core semantics The corresponding semantic fluctuation range can then be represented by a normal distribution to obtain latent variables. , This refers to random noise that follows a standard normal distribution.

[0127] The latent variables are then input into the decoder to calculate the cosine similarity between the latent variables and each dimension of the environmental anchors. The environmental anchors are vectors representing environmental limits, which need to be determined based on the specific scenario. For example, in the farmland irrigation scenario of this embodiment, environmental limits include the upper limit of crop water requirements and the equipment capacity boundary. The obtained environmental limit representations are then encoded into vectors and aggregated. The aggregated vectors are recorded as environmental anchors. Next, the value corresponding to each dimension of the latent variables is multiplied by its corresponding weight to obtain weighted latent variables. The weights are determined based on the aforementioned cosine similarity; the higher the cosine similarity, the larger the weight. Then, an activation function maps the weighted latent variables into deep feature vectors. Finally, the latent intent candidate set, explicit intent labels, deep feature vectors, and text ambiguity are integrated to obtain a structured dictionary, serving as a three-level intent set.

[0128] This embodiment uses explicit intent labels and implicit intent candidate sets as a basis, combined with the user's original semantic text, to finally generate a three-level intent set, achieving full-dimensional coverage of user intent. This solves the problem of information loss caused by intent only remaining at the surface level in the prior art. At the same time, the output deep feature vector provides information for the generation of subsequent dynamic constraints, enabling the constraint generation process to adapt to environmental limits, avoiding the disconnect between subsequent constraints and the actual bearing capacity of real scenarios, and ensuring the feasibility of water-saving suggestions.

[0129] Furthermore, this embodiment provides a step for obtaining a dynamic constraint set based on a coupling timing diagram, static constraint logic and logic labels, and deep feature vectors, including:

[0130] Based on the coupling sequence diagram and static constraint logic, the feature dependency chain is obtained;

[0131] Based on the feature dependency chain and deep feature vectors, the first dynamic constraint set is obtained;

[0132] Based on the first dynamic constraint set and the logical labels, the second dynamic constraint set is obtained;

[0133] The dynamic constraint set is obtained based on the coupling timing diagram and the second dynamic constraint set.

[0134] Specifically, based on the delay corresponding to each strongly coupled pair in the coupling time sequence diagram, each feature in the core feature set is arranged according to its delay. Arrange them from largest to smallest, and add their corresponding logical labels to each feature to obtain the feature dependency chain.

[0135] Then, each feature is optimized sequentially according to the feature dependency chain. For example, assuming the feature dependency chain is irrigation flow, temperature, and soil moisture, taking irrigation flow as an example, first, the logical labels corresponding to irrigation flow and soil moisture need to be obtained. Then, based on the current environmental data, the current soil moisture state and irrigation flow state are obtained, and the irrigation flow state is used as the initial value. Then, it is determined whether maintaining this initial value can make the soil moisture closer to its corresponding reasonable range. For example, if the current soil moisture state is 45%, its corresponding reasonable range is [40%, 60%], and the irrigation flow state is 8m 3 / hour, at this point we need to determine if it should remain at 8m 3 The direction of soil moisture change per hour is analyzed. If it deviates from the median of the reasonable range, it indicates that the soil is moving away from the corresponding reasonable range. If it approaches the median of the reasonable range, it indicates that the soil is moving closer to the corresponding reasonable range. This embodiment does not limit the specific judgment method; for example, it can be judged by constructing a fitting model of irrigation flow and soil moisture. If the soil moisture can be brought closer to the corresponding reasonable range, the current matching degree is positive; otherwise, the current matching degree is negative. However, this embodiment does not limit the specific values ​​of positive and negative numbers; for example, it can be determined based on the degree of approach or departure. Next, the similarity between the vector corresponding to the logical label of the feature and the deep feature vector is calculated. The similarity is used as the constraint matching degree. Then, the current matching degree and the constraint matching degree are weighted to obtain the reward function corresponding to soil moisture. This embodiment does not limit the specific values ​​of the weights.

[0136] Then, based on the reward function, the constraint range corresponding to the update adjuster can be adjusted using a PPO policy network. For example, it is first necessary to determine the current irrigation flow state, the logical label corresponding to the irrigation flow, and the current state value function. In the PPO policy network, the state value function is obtained by training a value network to predict the expected reward corresponding to the current state. The above reward function represents the actual reward that can be obtained after executing the current state (i.e., the actual reward corresponding to the current state). For example, it can be obtained using a regression neural network. First, the logical labels corresponding to the irrigation flow and soil moisture, as well as the current soil moisture state and irrigation flow state, are organized into a feature vector. This vector is then input into the trained regression neural network to obtain the current state value function. This embodiment does not elaborate on the training process of the neural network model. The core of the training is to minimize the difference between the actual reward and the expected reward.

[0137] After obtaining the reward function corresponding to the current irrigation flow rate according to the above steps, the difference between the reward function value and the expected reward is used as the dominance function value. If the dominance function value is greater than zero, it means that the reward corresponding to the current irrigation flow rate is higher than expected, and the probability of this value needs to be strengthened. For example, the initial reasonable range for the irrigation flow rate obtained from the logical label is 5m. 3 / hour-10m 3 / hour, after increasing the probability of the current value, the updated value range is 5m. 3 / hour-9m 3 / hour, thereby reducing irrigation flow by more than 8m³ / h by narrowing the value range. 3 The probability per hour is provided in this embodiment only as an exemplary way to enhance the probability of the value. Those skilled in the art can choose a reasonable way, and this embodiment does not limit it. Furthermore, when using the method of narrowing the value range, this embodiment does not limit the extent of narrowing the value range.

[0138] After obtaining the updated reasonable range, a second iteration is performed. An irrigation flow rate needs to be randomly generated from this range as the current irrigation flow rate state. Then, the above steps of judging whether maintaining the current irrigation flow rate state can make the soil moisture closer to its corresponding reasonable range, calculating the reward function, and updating the reasonable range are repeated until the value of the reward function converges. The iteration stops, and the value range of the irrigation flow rate after the last update is obtained. The reward function convergence means that the difference between two adjacent reward functions is less than a preset difference. In this embodiment, the value of the preset difference is not limited.

[0139] Next, following the order of the feature dependency chain, the reasonable ranges corresponding to temperature and soil moisture need to be optimized sequentially. The specific steps are the same as those described above, and will not be repeated here. It should be noted that, since the three features are dependent on each other, the initial values ​​of temperature and soil moisture need to be determined based on the irrigation flow rate during the first iteration. For example, when optimizing temperature, the initial value of temperature needs to be predicted based on the current irrigation flow rate and the delay between temperature and irrigation flow rate. Similarly, when optimizing soil moisture, the initial value of soil moisture needs to be predicted based on the delay, and then the subsequent steps are executed. This embodiment does not limit the method of prediction based on delay.

[0140] After obtaining the updated reasonable range for each feature in the feature dependency chain, the logical label corresponding to each feature in the feature dependency chain is modified based on this reasonable range to obtain the first dynamic constraint set. Then, the logical labels in the first dynamic constraint set are verified based on the logical labels in the feature dependency chain before the update. It must be ensured that the logical labels in the feature dependency chain before the update are a subset of the logical labels in the first dynamic constraint set. If the boundary of the logical label in the feature dependency chain before the update exceeds the boundary of the logical label in the first dynamic constraint set, no update is performed, resulting in the second dynamic constraint set. Finally, the coupling timing diagram and the second dynamic constraint set are packaged to obtain the dynamic constraint set. Furthermore, for each feature in the second dynamic constraint set, the relative deviation before and after the update needs to be calculated based on the corresponding values ​​in its logical labels before and after the update. This can be determined by the ratio of the interval widths corresponding to the logical labels before and after the update. The relative deviation and the dynamic constraint set are then sent to the matching degree calculation unit.

[0141] This embodiment automatically identifies strong coupling relationships by quantifying the correlation and delay patterns between features, replacing the traditional method of relying on human experience to judge feature associations. It achieves accurate quantitative identification of coupling relationships between features and, based on the PPO policy network, ensures that constraints simultaneously match the current scenario requirements and environmental carrying capacity. The final integrated output dynamic constraint set clearly presents the feature association logic and provides accurate constraints that can be directly executed, ensuring the accuracy of water-saving recommendations.

[0142] As shown in Figure 5, this embodiment further provides a step for obtaining candidate water-saving schemes based on a coupled timing diagram, a three-level intent set, and a dynamic constraint set, including:

[0143] Based on the coupling sequence diagram, the three-level intent set, and the dynamic constraint set, the matching degree and conflict points between intents and constraints are obtained;

[0144] The implicit intent is updated based on the conflict points to obtain candidate water-saving solutions.

[0145] Specifically, after receiving the coupled timing graph, the three-level intent set, and the dynamic constraint set, the matching degree calculation unit in the decision-making agent first needs to obtain the feature vector corresponding to each explicit intent in the three-level intent set. Then, it needs to obtain the feature vector corresponding to the logical label of each operational feature in the dynamic constraint set. For the dynamic constraint set, the features it contains can be divided into operational features and environmental features. Following the example above, when the dynamic constraint set includes irrigation flow, temperature, and soil moisture, irrigation flow is a feature directly controlled by humans and belongs to operational features, while temperature and soil moisture belong to environmental features. Feature vectors can be generated through kernel function mapping to eliminate dimensionality differences; this embodiment does not impose restrictions on this. Then, the Euclidean distance between each explicit intent and the logical label corresponding to each operational feature is calculated based on the feature vectors. The similarity between each explicit intent and the logical label corresponding to each operational feature is obtained based on the Euclidean distance and the Gaussian kernel formula, serving as the matching degree. For deep feature vectors, it is necessary to obtain the feature vector corresponding to the logical label of each environmental feature in the dynamic constraint set to obtain the matching degree between the deep feature vector and the logical label corresponding to each operational feature. The specific process is the same as the above steps, and this embodiment will not elaborate further.

[0146] Similarly, the feature vector corresponding to each implicit intent in the three-level intent set is obtained, and the feature vector corresponding to the coupled temporal graph is also obtained. The feature vector of the coupled temporal graph is determined according to the number of features and the degree of delay. For example, different quantization values ​​can be set for different numbers of features. If there are 3 features, the quantization value is 0.5; if there are 4 features, the quantization value is 0.6. At the same time, different quantization values ​​are set for different delays. For example, if the total delay is less than 3, the quantization value is 0.3; if the total delay is between 3 and 5, the quantization value is 0.5. Therefore, if the coupled temporal graph includes 4 features and the total delay is 4, its corresponding feature vector is [0.6, 0.5]. In this embodiment, non-numerical temporal flow attributes are transformed into numerical features that the model can calculate through feature vectors, so that subsequent calculations fit the actual temporal flow. The numerical values ​​in this embodiment are only examples. This embodiment does not limit the specific numerical values ​​of the quantization values, and this embodiment only provides a way to generate feature vectors. Other feature vectors used in the process of calculating the matching degree provided in this embodiment can also be determined by selecting appropriate quantization values, or those skilled in the art can choose other methods. This embodiment does not limit this. Then, the Euclidean distance between each latent intention and the coupled timing graph is calculated using the feature vectors to obtain the matching degree between each latent intention and the coupled timing graph. The specific principle is the same as the above steps, and this embodiment will not elaborate on it.

[0147] This embodiment selects matching objects for each intent based on its functional positioning. Specifically, explicit intents are direct manifestations of water conservation goals, and are matched with features that can be directly controlled. Implicit intents are related to operation steps, so they are matched with a coupling sequence diagram that reflects the process logic. Deep intents are obtained based on environmental anchor points, so they are matched with environmental features, so that the function of each intent can find its corresponding execution carrier, which is convenient for accurately locating the specific location when there is a conflict.

[0148] The relative deviations of each received feature are then weighted to obtain the constraint evolution consistency index. Next, each matching degree calculated above is weighted to obtain the weighted matching degree. Finally, the constraint evolution consistency index and the weighted matching degree are weighted to obtain the total matching degree. In this embodiment, the total matching degree takes into account the adaptability of the intent through the weighted matching degree, and integrates the constraint evolution consistency index to avoid the situation where the intent and constraint have a high matching degree but the constraint itself changes frequently.

[0149] Finally, the total matching degree is compared with the first preset matching degree. If the total matching degree is less than the first preset matching degree, it indicates that there is a conflict point. Then, each matching degree is compared with the second preset matching degree. The intentions that are less than the preset matching degree and their corresponding logical tags are regarded as conflict points. In this embodiment, there are no restrictions on the first and second preset matching degrees.

[0150] Then, the intent is updated based on the conflict point. The purpose of the update is to solve the problem of low matching degree between the intent and the logical constraint. For example, if the conflict point is the explicit intent and its corresponding logical label, several candidate extended intents can be generated based on the original explicit intent through the adversarial training logic of the generative adversarial network. Then, the matching degree of each candidate extended intent is calculated by the discriminator. The method of calculating the matching degree is the same as described above, and will not be repeated here. The loss function is calculated by the matching degree to update the parameters in the generator, so that the generator generates several candidate extended intents again until a preset termination condition is reached to obtain the final candidate extended intent. The preset termination condition is the convergence of the loss function or the matching degree. Finally, the original explicit intent is replaced with the final candidate extended intent. This embodiment does not limit the specific form of the loss function. It can be calculated by the matching degree. The purpose of the loss function is to measure the difference between the currently obtained candidate extended intent and the logical label. Similarly, if the conflict point includes implicit intent or deep intent vectors, the above method can be used to replace the original implicit intent or deep intent vectors. It should be noted that since explicit intents were used when initially generating implicit labels and deep intent vectors, when the conflict point includes implicit intents or deep intent vectors, a verification step needs to be added after obtaining several candidate extended intents before calculating the matching degree. For example, for implicit intents, it is necessary to verify whether the candidate extended intents match explicit labels and intent seeds; for deep intent vectors, it is necessary to verify whether the candidate extended intents match explicit labels, coupling sequence graphs, and environmental anchors. This embodiment does not limit the matching method; for example, it can be judged by similarity.

[0151] Following the updated intent, the solution generation unit generates candidate water-saving solutions. Specifically, the generator randomly generates several water-saving solutions based on the input intent and historical water-saving solutions. These solutions can include specific values ​​corresponding to operational features and triggering logic, such as initiating irrigation when soil moisture S ≤ 50% with an irrigation flow rate of F = 7m³. 3 / hour, this scheme is only one example, and this embodiment does not limit it. Based on the discriminator, the multi-objective score of each scheme is calculated. The multi-objective score is obtained by weighting the intent satisfaction and constraint fit of each scheme. Specifically, based on the above method, the matching degree corresponding to each intent and logical label is generated, and the matching degree of each intent is weighted to obtain the intent satisfaction. Then, the overlap between the updated logical label and the original logical label is calculated and weighted as the constraint fit. Then, the loss function is calculated based on the intent satisfaction and constraint fit, and the water-saving scheme output by the generator is updated. This embodiment does not limit the specific form of the loss function. The purpose of the loss function is to make the water-saving scheme satisfy the intent and logical constraints until a preset termination condition is reached, and a candidate water-saving scheme is output. The preset termination condition can be the convergence of the loss function or the reaching of a preset number of iterations. This embodiment does not limit it. Finally, only one candidate water-saving scheme is output.

[0152] This embodiment uses the strong coupling relationship of features quantized by coupled timing graph as the logical framework, anchors the constraint transmission rhythm between parameters, avoids the disorder of parameter adjustment order, and ensures the logical self-consistency of the constraint optimization process from the bottom layer. It also solves the problem of insufficient constraint adaptation dimensions under the guidance of a single requirement by using the multi-dimensional requirement decomposition and matching quantification method of three-level intent set. The dynamic constraint set serves as the execution boundary, ensuring that the generated water-saving scheme complies with the basic safety and operational rules.

[0153] Furthermore, this embodiment provides a step for updating the three-level intent set based on candidate water-saving schemes to obtain water-saving recommendations, including:

[0154] Based on the candidate water-saving schemes and the updated implicit intent, the updated dynamic constraint set is obtained;

[0155] Update the level 3 intent set based on the updated dynamic constraint set;

[0156] Based on the updated dynamic constraint set and the updated three-level intent set, water-saving suggestions are obtained.

[0157] Specifically, after obtaining candidate water-saving schemes, it is necessary to determine the cumulative deviation of each feature when executing the water-saving scheme in historical situations based on historical data, and then adjust the logical labels corresponding to the features based on the cumulative deviations so that the logical labels can better match the actual execution capabilities of the features. Finally, an updated dynamic constraint set is obtained, where the cumulative deviation is the average difference between the value executed by the feature and the value actually achieved in each historical situation.

[0158] Next, for the updated dynamic constraint set, the feature vector corresponding to the logical label of each operation feature is obtained. The steps for obtaining the feature vector are the same as those described above, and will not be repeated in this embodiment. Then, in order to avoid the incompatibility problem between features and intents, the VAE encoder can compress them into latent space vectors. Then, the VAE decoder decodes the corresponding deep intent from the latent space vectors and converts the deep intent into a feature vector. The newly decoded deep intent is integrated into the original three-level intent set to obtain the updated three-level intent set.

[0159] Furthermore, this embodiment provides a step for obtaining water-saving suggestions based on the updated dynamic constraint set and the updated three-level intent set, including:

[0160] Based on the updated dynamic constraint set and the updated three-level intent set, the matching degree between the updated intent and constraints is obtained;

[0161] Based on the updated match between intent and constraints, the updated set of dynamic constraints, and candidate water-saving schemes, water-saving recommendations are obtained.

[0162] Specifically, based on the updated dynamic constraint set and the updated level-three intent set, the matching degree between the updated intent and constraints is obtained in the same way as the steps described above for calculating the matching degree. The constraints include logical labels and coupling sequence diagrams, which will not be elaborated here in this embodiment. When the matching degree verification is successful, it means that the current constraints completely match the level-three intent. Then, the candidate water-saving scheme can be modified according to the updated dynamic constraint set to generate water-saving suggestions that conform to the current environment.

[0163] This application provides a water-saving intelligent agent interaction control method and system based on semantic intent prediction. First, it mines user intent layer by layer through self-attention weights to form a three-level intent set. Then, it generates a dynamic constraint set by combining the current environmental data to adapt to the rules of the real-time scenario. Next, it detects the matching degree between user intent and constraints to trigger adversarial correction between intent and dynamic constraint set. Finally, it obtains the optimal water-saving suggestion that meets both user needs and actual scenario, solving the problems of existing technologies that are difficult to accurately identify user needs and generate effective suggestions when parameters change in real time.

[0164] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0165] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0166] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more block diagrams.

[0167] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A water-saving intelligent agent interaction control method based on semantic intent prediction, characterized in that, include: Based on the user's semantic text and self-attention weights, a three-level intent set and a deep feature vector are obtained; Based on the current environmental data and the deep feature vector, a coupled time series diagram and a dynamic constraint set are obtained. The current environmental data includes soil moisture, irrigation flow rate, and temperature. Based on the coupled timing diagram, the three-level intent set, and the dynamic constraint set, candidate water-saving schemes are obtained; the three-level intent set is updated based on the candidate water-saving schemes to obtain water-saving suggestions; wherein, obtaining the coupled timing diagram and dynamic constraint set based on the current environmental data and the deep feature vector includes: obtaining a core feature set and time-sharing data packets based on the current environmental data; obtaining static constraint logic and logic labels based on the core feature set; obtaining the coupled timing diagram and dynamic constraint set based on the core feature set, the time-sharing data packets, the static constraint logic and logic labels, and the deep feature vector; wherein, based on the core feature set, the time-sharing data packets, and the static constraint logic... The process involves obtaining a coupled temporal graph and a dynamic constraint set based on the logical labels and the deep feature vectors, including: obtaining a temporal feature vector based on the core feature set and the time-division data packets; obtaining the marginal probability corresponding to each value and the joint probability corresponding to each pair of values ​​based on the values ​​of each dimension in the temporal feature vectors; calculating the mutual information and entropy of each pair of features under different delays based on the marginal probabilities and the joint probabilities; obtaining the coupling strength of each pair of features under different delays based on the mutual information and entropy; filtering feature pairs based on the coupling strength to obtain a coupled temporal graph; and obtaining the dynamic constraint set based on the coupled temporal graph, the static constraint logic and logical labels, and the deep feature vectors.

2. The water-saving intelligent agent interaction control method based on semantic intent prediction according to claim 1, characterized in that, The process of obtaining a three-level intent set and a deep feature vector based on the user's semantic text and self-attention weights includes: obtaining explicit intent labels based on the semantic text and self-attention weights; obtaining a latent intent candidate set based on the explicit intent labels and a generative adversarial network; and obtaining a three-level intent set and a deep feature vector based on the latent intent candidate set, the explicit intent labels, and the semantic text.

3. The water-saving intelligent agent interaction control method based on semantic intent prediction according to claim 2, characterized in that, Based on the implicit intent candidate set, the explicit intent label, and the semantic text, a three-level intent set and a deep feature vector are obtained, including: obtaining a fused text sequence based on the implicit intent candidate set, the explicit intent label, and the semantic text; obtaining the deep feature vector and text ambiguity based on the fused text sequence and a variational autoencoder; and obtaining the three-level intent set based on the implicit intent candidate set, the explicit intent label, the deep feature vector, and the text ambiguity.

4. The water-saving intelligent agent interaction control method based on semantic intent prediction according to claim 1, characterized in that, The dynamic constraint set is obtained based on the coupling timing diagram, the static constraint logic and logic labels, and the deep feature vectors, including: obtaining feature dependency chains based on the coupling timing diagram and the static constraint logic; obtaining a first dynamic constraint set based on the feature dependency chains and the deep feature vectors; obtaining a second dynamic constraint set based on the first dynamic constraint set and the logic labels; and obtaining the dynamic constraint set based on the coupling timing diagram and the second dynamic constraint set.

5. The water-saving intelligent agent interaction control method based on semantic intent prediction according to claim 1, characterized in that, Based on the coupled timing diagram, the three-level intent set, and the dynamic constraint set, candidate water-saving schemes are obtained, including: obtaining the matching degree and conflict points between intents and constraints based on the coupled timing diagram, the three-level intent set, and the dynamic constraint set; updating the implicit intent based on the conflict points to obtain candidate water-saving schemes.

6. The water-saving intelligent agent interaction control method based on semantic intent prediction according to claim 5, characterized in that, The process of updating the third-level intent set based on the candidate water-saving schemes to obtain water-saving suggestions includes: obtaining an updated dynamic constraint set based on the candidate water-saving schemes and the updated implicit intents; updating the third-level intent set based on the updated dynamic constraint set; and obtaining water-saving suggestions based on the updated dynamic constraint set and the updated third-level intent set.

7. The water-saving intelligent agent interaction control method based on semantic intent prediction according to claim 6, characterized in that, Based on the updated dynamic constraint set and the updated three-level intent set, water-saving suggestions are obtained, including: obtaining the updated matching degree between intent and constraint based on the updated dynamic constraint set and the updated three-level intent set; and obtaining the water-saving suggestions based on the updated matching degree between intent and constraint, the updated dynamic constraint set, and the candidate water-saving schemes.

8. A water-saving intelligent agent interaction system based on semantic intent prediction, characterized in that, include: The intent agent is used to obtain a three-level intent set and a deep feature vector based on the user's semantic text and self-attention weights. A constraint agent is used to obtain a coupled time sequence diagram and a dynamic constraint set based on current environmental data and the deep feature vector. The current environmental data includes soil moisture, irrigation flow rate, and temperature. A decision agent is used to obtain candidate water-saving schemes by interacting with the intention agent and the constraint agent, combining the coupled time sequence diagram, the three-level intention set, and the dynamic constraint set, and to update the three-level intention set based on the candidate water-saving schemes to obtain water-saving suggestions. The step of obtaining the coupled time sequence diagram and the dynamic constraint set based on the current environmental data and the deep feature vector includes: obtaining a core feature set and a time-sharing data package based on the current environmental data; obtaining static constraint logic and logical labels based on the core feature set; and obtaining the core feature set, the time-sharing data package, the static constraint logic and logical labels, and the deep feature vector. The process involves obtaining a coupled temporal graph and a dynamic constraint set based on the core feature set, the time-division data packets, the static constraint logic and logical labels, and the deep feature vector. This includes: obtaining a temporal feature vector based on the core feature set and the time-division data packets; obtaining the marginal probability corresponding to each value and the joint probability corresponding to each pair of values ​​based on the values ​​of each dimension in the temporal feature vector; calculating the mutual information and entropy of each pair of features under different delays based on the marginal probabilities and the joint probabilities; obtaining the coupling strength of each pair of features under different delays based on the mutual information and entropy; filtering feature pairs based on the coupling strength to obtain the coupled temporal graph; and obtaining the dynamic constraint set based on the coupled temporal graph, the static constraint logic and logical labels, and the deep feature vector.

Images