Radar voice interaction control method and system based on natural language processing

Abstract

The invention discloses a radar voice interaction control method and system based on natural language processing, relates to the technical field of voice recognition, and solves the problem that it is difficult to perform target state detection by combining radar echo signals with hyperspectral image data. Voice interaction control is difficult to realize by utilizing natural language processing; the priority of the control instruction is difficult to reasonably determine based on the environment comprehensive coefficient and the emergency degree value; a conflict detection mechanism is difficult to establish to ensure the feasibility of the instruction; the instruction is difficult to execute under the condition of multi-target game and priority dynamic change, and the deviation value is monitored and evaluated to establish an error reporting mechanism. According to the method, target state detection is carried out by fusing radar echo signals and hyperspectral image data, voice interaction control is realized by utilizing natural language processing, and hierarchical control signals are dynamically generated by combining an environment comprehensive coefficient and an emergency degree value.

Description

Technical Field

[0001] The present invention belongs to the field of speech recognition technology, and specifically provides a radar speech interaction control method and system based on natural language processing. Background Art

[0002] In natural language processing, speech recognition technology uses pattern recognition and machine learning algorithms to extract acoustic features from speech signals and convert them into text. Natural language understanding performs semantic analysis on text to understand the user's control intent. Radar technology detects object information by emitting radio waves and receiving reflected signals. In greenhouse scenarios, it can be used for contactless crop detection. Combining the advantages of both technologies can meet the needs of intelligent greenhouse control and remote monitoring, enabling interactive voice control.

[0003] The existing radar voice interaction control methods and systems based on natural language processing have the following problems: it is difficult to combine radar echo signals with hyperspectral image data for target status detection; it is difficult to use natural language processing to realize voice interaction control; it is difficult to reasonably determine the priority of control instructions based on the comprehensive environmental coefficient and urgency value; it is difficult to establish a conflict detection mechanism to ensure the feasibility of instructions; it is difficult to execute instructions under the circumstances of multi-target game and dynamic change of priority, and it is difficult to monitor and evaluate deviation values ​​to establish an error reporting mechanism. Summary of the Invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a radar voice interaction control method and system based on natural language processing to solve the above-mentioned technical problems.

[0005] A first aspect of the present invention provides a radar voice interaction control method based on natural language processing, comprising the following steps:

[0006] S1: In the greenhouse, radar echo signals, voice signals and hyperspectral image data are collected synchronously and preprocessed using a multi-threaded parallel preprocessing mechanism;

[0007] S2: Target status detection is performed by combining pre-processed radar echo signals with hyperspectral image data. The pre-processed voice signals are then subjected to voice recognition to analyze the semantics of voice commands and extract control intent and target parameters.

[0008] S3: Calculates the urgency value by calculating the comprehensive environmental coefficient and the target status detection results; dynamically binds the target parameters in the voice command with the radar control logic, and generates radar control commands and hierarchical control signals through natural language processing; establishes a conflict detection mechanism and verifies whether the command is feasible;

[0009] S4: Dynamically execute instructions based on multi-target game and priority; monitor and evaluate deviation values ​​after executing radar control instructions and establish an error reporting mechanism.

[0010] Preferably, the target state detection in step S2 is performed by combining the preprocessed radar echo signal with the hyperspectral image data, comprising the following steps:

[0011] Based on the radar beam width and crop planting density, an adaptive algorithm is used to divide the greenhouse into n rectangular grids;

[0012] Using a sliding window mechanism, the time-series echo intensity and phase difference data of the i-th rectangular grid are continuously collected, where i∈{1,2,...,n}; vegetation index features are extracted from the hyperspectral image data; the time-series echo intensity, phase difference data and vegetation index are fused to construct a joint feature vector;

[0013] A lightweight hybrid network model is constructed by combining convolutional neural networks and long short-term memory networks. A joint feature vector dataset is collected and the target status is labeled, including health status and pest and disease status. The labeled dataset is input into the lightweight hybrid network model for training. The trained lightweight hybrid network model outputs the target status detection results through real-time processing of radar echo signals and hyperspectral image data.

[0014] Preferably, in step S2, the pre-processed voice signal is subjected to voice recognition, and then the semantics of the voice command is analyzed to extract the control intention and target parameters, which includes the following steps:

[0015] Collect historical pre-processed speech signals and input them into a deep learning-based speech recognition model for training. During the training process, the speech recognition model decodes the speech and outputs the corresponding speech commands in text form.

[0016] Input the real-time preprocessed speech signal into the trained speech recognition model and output the recognition result; preprocess the speech command text form and add greenhouse rectangular grid area marks;

[0017] Collect and preprocess a voice command text dataset containing various voice commands and their corresponding control intent labels;

[0018] The preprocessed voice command text dataset is input into the BERT model for training. The real-time voice command text is input into the trained BERT model to output the corresponding control intent label. The token-level classification capability of the BERT model is used to identify the target parameters in the voice command text.

[0019] Preferably, the step S3 includes the following steps: calculating the comprehensive environmental coefficient and calculating the urgency value according to the target state detection result:

[0020] The formula for calculating the comprehensive environmental coefficient is:

[0021]

[0022] Among them, H is the environmental factor, H T and H D Represents the ambient temperature and humidity respectively, R Ti 、R Di and R Fi represent the soil temperature, moisture and fertility of the i-th rectangular grid respectively; and Respectively represent the optimal ambient temperature and ambient humidity; R Tmax 、R Dmax and R Fmax Represent the optimal soil temperature, moisture and fertility respectively; ε1 and ε2 are the influence weights;

[0023] The formula for calculating urgency is:

[0024]

[0025] Among them, E i is the urgency value in the i-th rectangular grid, S i is the planting area of ​​pest-infested crops in the i-th rectangular grid, A is the total area, ΔV i is the vegetation index deviation value, V max is the maximum value of vegetation index deviation, B i is the pest and disease diffusion rate, H is the comprehensive environmental coefficient, σ1 is the confidence level output by the BERT model, σ2 is the confidence level output by the lightweight hybrid network model, and w1, w2, and w3 represent weight coefficients, respectively.

[0026] Preferably, in step S3, the target parameters in the voice command are dynamically bound to the radar control logic, and the radar control command and hierarchical control signal are generated through natural language processing, including the following steps:

[0027] Bind the target state detection results and the target parameters in the voice command to build a mapping table, store the mapping table in the control logic database, and generate radar control operation instructions;

[0028] Based on the information in the mapping table, the corresponding radar control instruction text is generated using natural language processing technology according to the target parameters and the bound control logic, and converted into a coded format;

[0029] Analyze control intent labels and target parameters, and match preset operation templates from the control logic database; divide control intent into three response priorities: high, medium, and low, based on the urgency value;

[0030] According to the divided priorities, hierarchical control signals are generated, including different priority identifiers and corresponding control operation instructions; the generated control signals are used to control the working state of the radar.

[0031] Preferably, establishing a conflict detection mechanism in step S3 and verifying whether the instruction is feasible includes the following steps:

[0032] Create a three-dimensional conflict detection space, including: radar beam coverage conflicts, instruction execution timing conflicts, and radar power allocation conflicts;

[0033] By constructing the conflict matrix, the matrix elements are: C_xy=α1*|m_x∩m_y|+α2*|t_x∩t_y|+α3*(P_x+P_y-P_max),

[0034] Where m_x∩m_y represents the overlapping area of ​​radar beam coverage under the xth and yth instructions; t_x∩t_y represents the overlapping duration of instruction execution under the xth and yth instructions; P_x, P_y, and P_max are the radar power and maximum radar power under the xth and yth instructions, respectively; α1, α2, and α3 are the weight coefficients of radar beam coverage conflict, instruction execution timing conflict, and radar power allocation conflict, respectively.

[0035] By traversing all elements in the conflict matrix, when C_xy>0.35, it is determined that there is a conflict between the xth instruction and the yth instruction;

[0036] By calculating the radar's available power margin in real time and predicting the resource utilization curve in a future time period through Monte Carlo simulation, the confidence interval of the resource utilization rate is obtained;

[0037] When the resources required by an instruction exceed the confidence interval, verifying the instruction is not feasible.

[0038] Preferably, the step S4 dynamically executes instructions based on multi-objective game and priority, including the following steps: when multiple voice instructions are triggered simultaneously, the Nash equilibrium solution is calculated according to the urgency value, target response priority and radar load, and time division multiplexing technology is used to divide the radar beam into independent time slots at the microsecond level to execute different instructions respectively.

[0039] Preferably, monitoring and evaluating the deviation value after executing the radar control instruction in step S4 and establishing an error reporting mechanism include the following steps:

[0040] By real-time monitoring of radar operating status including radar transmission frequency, scanning angle, power and echo signal-to-noise ratio, the target response including target response time is monitored and normalized;

[0041] By weighted evaluation of the deviation from the actual target, when the deviation exceeds the preset threshold, an error message is given to the user.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] The present invention fuses radar echo signals with hyperspectral image data, uses adaptive grid division and sliding window mechanism to collect time series data, constructs a joint feature vector based on vegetation index characteristics, and then processes it through a lightweight hybrid model consisting of a convolutional neural network and a long short-term memory network. It achieves high-precision detection of the health of greenhouse crops and the status of pests and diseases, and significantly improves environmental perception capabilities.

[0044] This invention combines a deep learning speech recognition model with a BERT model to first convert the voice signal into text, and then accurately analyze the semantics to extract control intent and parameters, so that the radar control has a natural and smooth voice interaction function, greatly improving the convenience and flexibility of operation and reducing the complexity of manual operation.

[0045] The present invention determines the command priority by calculating the comprehensive environmental coefficient and the urgency value, dynamically binds voice command parameters with radar control logic to generate hierarchical control signals, establishes a conflict detection mechanism to verify the feasibility of commands, and uses time-division multiplexing technology to allocate radar resources in combination with multi-objective game and priority during execution, thereby enhancing the resource utilization efficiency of radar control. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0047] Figure 1 is a flow chart of the method of the present invention;

[0048] Figure 2 It is a system module diagram of the present invention. DETAILED DESCRIPTION

[0049] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0050] See also Figure 1 As shown, the first embodiment of the present invention provides a radar voice interaction control method based on natural language processing, comprising the following steps:

[0051] S1: In the greenhouse, radar echo signals, voice signals and hyperspectral image data are collected synchronously and preprocessed using a multi-threaded parallel preprocessing mechanism;

[0052] S2: Target status detection is performed by combining pre-processed radar echo signals with hyperspectral image data. The pre-processed voice signals are then subjected to voice recognition to analyze the semantics of voice commands and extract control intent and target parameters.

[0053] S3: Calculates the urgency value by calculating the comprehensive environmental coefficient and the target status detection results; dynamically binds the target parameters in the voice command with the radar control logic, and generates radar control commands and hierarchical control signals through natural language processing; establishes a conflict detection mechanism and verifies whether the command is feasible;

[0054] S4: Dynamically execute instructions based on multi-target game and priority; monitor and evaluate deviation values ​​after executing radar control instructions and establish an error reporting mechanism.

[0055] Specifically, a radar, microphone array, and hyperspectral camera are installed within the greenhouse to simultaneously collect radar echo signals, voice signals, and hyperspectral image data. A multi-threaded parallel preprocessing mechanism is employed to perform noise filtering and signal normalization on the collected radar echo signals to remove interference and enhance the useful signal. Voice activity detection and endpoint recognition are performed on the voice signals to identify the valid portion of the voice signal and remove silent segments. Radiometric and atmospheric corrections are performed on the hyperspectral image data to improve the accuracy and reliability of the image data. Wavelet filtering, voice endpoint detection, and radiometric correction techniques are employed to optimize data quality. In conjunction with an adaptive gridding strategy, the greenhouse is dynamically segmented into rectangular monitoring units. A sliding window mechanism is used to extract time-series echo intensity, phase difference, and vegetation index features. A lightweight hybrid model is constructed using joint feature vector inputs to enable real-time monitoring of crop health and pest and disease status. A deep learning-based speech recognition model works in conjunction with the BERT semantic parsing model to convert voice commands into text, identify intent, and extract parameters such as spraying area and dosage. Spatial positioning is achieved using the greenhouse grid coordinate system. By integrating ambient temperature and humidity, soil parameters, and crop status data, the system calculates a comprehensive environmental coefficient and urgency value, dynamically binding control logic to generate hierarchical radar control commands, such as adjusting the scanning angle and activating spraying equipment. Prioritization is based on a Nash equilibrium game theory model, and execution time slots are allocated in real time based on radar load. Microsecond-level time-division multiplexing technology is used to achieve multi-command coordination. A three-dimensional conflict detection space is constructed, and a conflict matrix is ​​used to quantify beam coverage, timing overlap, and power allocation risks. Monte Carlo simulation is used to predict resource utilization and automatically avoid conflicting commands. During the execution phase, parameters such as radar transmission frequency, scanning angle, and echo signal-to-noise ratio are monitored in real time. A weighted deviation value evaluation model is used to establish an error reporting mechanism.

[0056] In this embodiment, the target state detection in step S2 is performed by combining the preprocessed radar echo signal with the hyperspectral image data, including the following steps:

[0057] Based on the radar beam width and crop planting density, an adaptive algorithm is used to divide the greenhouse into n rectangular grids;

[0058] Using a sliding window mechanism, the time-series echo intensity and phase difference data of the i-th rectangular grid are continuously collected, where i∈{1,2,...,n}; vegetation index features are extracted from the hyperspectral image data; the time-series echo intensity, phase difference data and vegetation index are fused to construct a joint feature vector;

[0059] A lightweight hybrid network model is constructed by combining convolutional neural networks and long short-term memory networks. A joint feature vector dataset is collected and the target status is labeled, including health status and pest and disease status. The labeled dataset is input into the lightweight hybrid network model for training. The trained lightweight hybrid network model outputs the target status detection results through real-time processing of radar echo signals and hyperspectral image data.

[0060] Specifically, first, the radar beam width and the planting density of crops in the greenhouse are obtained. Based on these parameters, an adaptive algorithm is used to divide the greenhouse monitoring area into n uniformly sized rectangular grids. The size and position of each grid are optimized by the adaptive algorithm based on the radar beam coverage and crop distribution. Specifically, the K-means clustering algorithm can be used to dynamically divide the greenhouse into n rectangular grids. The grid side length calculation formula is: Where L is the grid side length, θ is the radar beamwidth, and ρ is the crop density within the greenhouse. For the i-th rectangular grid, a sliding window mechanism is used to continuously collect time-series echo intensity and phase difference data within it. The sliding window size and step size are set based on the characteristics of the radar signal and the dynamic changes of the crop, ensuring that subtle changes in the target state can be captured. Simultaneously, vegetation index features corresponding to each rectangular grid, such as the Normalized Difference Vegetation Index (NDVI) and the Ratio Vegetation Index (RVI), are extracted from the hyperspectral image data. These indices can effectively reflect the health and growth status of the crop. The collected time-series echo intensity and phase difference data and the extracted vegetation index are fused to construct a joint feature vector. Specifically, these different types of data are normalized, arranged in a specific order, and combined into a vector that serves as a comprehensive feature describing the target state within each rectangular grid. A lightweight hybrid network model is constructed by combining a convolutional neural network (CNN) and a long short-term memory (LSTM) network. CNN is used to extract spatial features from the data, while LSTM is used to capture temporal features. The combination of the two fully exploits the spatiotemporal characteristics of radar echo signals and hyperspectral image data. A large dataset of joint feature vectors representing different target states, including health and pest and disease status, is collected and annotated. This annotated dataset is then fed into a lightweight hybrid network model for training. During training, a backpropagation algorithm is used to adjust model parameters to minimize the error between the predicted results and the actual annotations, enabling the model to accurately learn the feature representations of the target states. The trained lightweight hybrid network model receives real-time processed radar echo signals and hyperspectral image data and outputs detection results for the health and pest status of targets, such as crops, within each rectangular grid, providing a basis for subsequent radar voice interaction control.

[0061] In this embodiment, step S2 performs speech recognition on the preprocessed speech signal, analyzes the semantics of the speech instruction, and extracts the control intent and target parameters, including the following steps:

[0062] Collect historical pre-processed speech signals and input them into a deep learning-based speech recognition model for training. During the training process, the speech recognition model decodes the speech and outputs the corresponding speech commands in text form.

[0063] Input the real-time preprocessed speech signal into the trained speech recognition model and output the recognition result; preprocess the speech command text form and add greenhouse rectangular grid area marks;

[0064] Collect and preprocess a voice command text dataset containing various voice commands and their corresponding control intent labels;

[0065] The preprocessed voice command text dataset is input into the BERT model for training. The real-time voice command text is input into the trained BERT model to output the corresponding control intent label. The token-level classification capability of the BERT model is used to identify the target parameters in the voice command text.

[0066] Specifically, a large amount of historical preprocessed speech signals covering a wide range of possible commands and sentences are collected to construct a training dataset. These collected speech signals are then fed into a deep learning-based speech recognition model for training. During training, the model uses a speech decoding algorithm to convert the speech signals into corresponding textual voice commands. Common deep learning models include architectures such as recurrent neural networks (RNNs), convolutional neural networks (CNNs), or Transformers. After training on large amounts of data, the models can learn the mapping between speech signals and text. The real-time collected and preprocessed speech signals are fed into the trained speech recognition model, which then outputs the corresponding textual voice commands. Preprocessing may include voice enhancement, background noise removal, and speech signal normalization to improve recognition accuracy. The outputted textual voice commands are then preprocessed, including text cleaning (removing irrelevant characters and punctuation), word segmentation (segmenting the text into words or phrases), and part-of-speech tagging to facilitate subsequent semantic parsing. Greenhouse rectangular grid area markers are added to associate voice commands with specific greenhouse areas. For example, a user might mention "crops in row 3, column 5" in a voice command. The system can then identify and label the corresponding rectangular grid area. A voice command text dataset containing various voice commands and their corresponding control intent labels is collected. Control intent labels can include specific operational instructions such as turning on irrigation, closing ventilation, and spraying pesticides. The collected voice command text dataset is preprocessed, including formatting, labeling intent, and performing data augmentation (such as synonym replacement and sentence reorganization) to enrich the training data. The preprocessed voice command text dataset is then fed into the BERT model for training. BERT is a pretrained language model with powerful semantic understanding capabilities. By pretraining on large amounts of text data, BERT learns the deep semantics of language. Using the trained BERT model, real-time voice command text is fed into the model, which then outputs the corresponding control intent labels. Leveraging the BERT model's token-level classification capabilities, each word or phrase in the voice command text is classified to identify the target parameter. For example, in the instruction "Spray pesticide on the crops in the 3rd row and 5th column with a dosage of 10 ml", the BERT model can recognize the operation intention represented by "grid position, spray pesticide" represented by "3rd row and 5th column" and the dosage parameter represented by "10 ml".

[0067] In this embodiment, the step S3 includes the following steps: calculating the comprehensive environmental coefficient and calculating the urgency value according to the target state detection result:

[0068] The formula for calculating the comprehensive environmental coefficient is:

[0069]

[0070] Among them, H is the environmental factor, H T and H D Represents the ambient temperature and humidity respectively, R Ti 、R Di and R Fi represent the soil temperature, moisture and fertility of the i-th rectangular grid respectively; and Respectively represent the optimal ambient temperature and ambient humidity; R Tmax 、R Dmax and R Fmax Represent the optimal soil temperature, moisture and fertility respectively; ε1 and ε2 are the influence weights;

[0071] The formula for calculating urgency is:

[0072]

[0073] Among them, E i is the urgency value in the i-th rectangular grid, S i is the planting area of ​​pest-infested crops in the i-th rectangular grid, A is the total area, ΔV i is the vegetation index deviation value, V max is the maximum value of vegetation index deviation, B i is the pest and disease diffusion rate, H is the comprehensive environmental coefficient, σ1 is the confidence level output by the BERT model, σ2 is the confidence level output by the lightweight hybrid network model, and w1, w2, and w3 represent weight coefficients, respectively.

[0074] Specifically, temperature, humidity, and soil sensors are installed in the greenhouse to collect environmental and soil data in real time. The target state detection results are used to determine the planting area and vegetation index deviation of pest-infested crops. The confidence values ​​output by the BERT model and the lightweight hybrid network model are collected. The optimal ambient temperature and humidity, as well as the optimal soil temperature, humidity, and fertility, are determined based on the type of crop being planted. In this embodiment, they can be set to an optimal ambient temperature of 25 degrees Celsius, an ambient humidity of 65%, and an optimal soil temperature of 20 degrees Celsius, a humidity of 70%, and a fertility value, i.e., an EC value of 2.5 mS / cm. ε1 and ε2 are influence weights of 0.4 and 0.6, respectively. w1, w2, and w3 represent weight coefficients of 0.4, 0.4, and 0.2, respectively. Based on crop growth requirements, the influence weights and weight coefficients of each parameter are adjusted through experiments or expert experience. For each rectangular grid, the above parameters are substituted and the above formula is applied to calculate the comprehensive environmental coefficient and urgency value.

[0075] In this embodiment, step S3 dynamically binds the target parameters in the voice command to the radar control logic, and generates radar control commands and hierarchical control signals through natural language processing, including the following steps:

[0076] Bind the target state detection results and the target parameters in the voice command to build a mapping table, store the mapping table in the control logic database, and generate radar control operation instructions;

[0077] Based on the information in the mapping table, the corresponding radar control instruction text is generated using natural language processing technology according to the target parameters and the bound control logic, and converted into a coded format;

[0078] Analyze control intent labels and target parameters, and match preset operation templates from the control logic database; divide control intent into three response priorities: high, medium, and low, based on the urgency value;

[0079] According to the divided priorities, hierarchical control signals are generated, including different priority identifiers and corresponding control operation instructions; the generated control signals are used to control the working state of the radar.

[0080] Specifically, the target status detection results, such as pest and disease status, crop health status, etc., are bound to the target parameters in the voice command.

[0081] For example, if the voice command is "Spray pesticide on the crops in the 3rd row and 5th column with a dosage of 10 ml", then "3rd row and 5th column" and "Spray pesticide at a dosage of 10 ml" are bound to the target state detection result such as "The crops in the 3rd row and 5th column are in a disease and insect pest state".

[0082] A mapping table is constructed, recording the status information, voice commands, and corresponding target parameters for each rectangular grid. This table is stored in the control logic database to facilitate subsequent control command generation and execution. Based on the information in the mapping table, combined with the target parameters and bound control logic, specific radar control operation commands are generated.

[0083] For example, if the target state detection result shows that "the crops in the 3rd row and 5th column are in a disease and insect pest state" and the voice command is "spray pesticides", the generated operation instruction may be "spray 10 ml of pesticide on the crops in the 3rd row and 5th column".

[0084] Based on the information in the mapping table, combined with the target parameters and bound control logic, natural language processing technology such as text generation models is used to generate detailed radar control instruction text. For example, the generated instruction text may be "Radar system, please spray pesticide on the crop in row 3, column 5, with a dosage of 10 ml and an operation time of immediate execution." The generated radar control instruction text is converted into a coding format that can be recognized by the radar system. For example, the instruction text can be converted into binary code or other machine-readable formats for execution by the radar system. The control intent label and target parameters are parsed, and the preset operation template is matched from the control logic database. For example, if the control intent is "spraying pesticides", the corresponding "spraying pesticides" operation template is found from the database. The template may contain specific spraying steps, dosage, time, and other information.

[0085] Based on the urgency level, control intentions are classified into three response priorities: high, medium, and low. When the urgency level is greater than a threshold of 0.8, the control intention is marked as high priority and requires immediate execution. When the urgency level is greater than or equal to 0.4 and less than or equal to the threshold of 0.8, it is marked as medium priority. When the urgency level is less than the threshold of 0.4, it is marked as low priority and can be deferred. Based on the priority level, a hierarchical control signal is generated, including different priority identifiers and corresponding control operation instructions. The thresholds are adjusted based on actual conditions or historical data.

[0086] For example, a high-priority control signal may include a tag “high priority” and a corresponding pesticide spraying operation instruction; a low-priority control signal may include a tag “low priority” and a corresponding irrigation operation instruction.

[0087] The generated control signal is sent to the radar system to control the radar's operating status. The radar system performs corresponding operations based on the received control signal, such as adjusting the radar beam direction and controlling spraying equipment.

[0088] In this embodiment, the conflict detection mechanism is established in step S3, and the verification of whether the instruction is feasible includes the following steps:

[0089] Create a three-dimensional conflict detection space, including: radar beam coverage conflicts, instruction execution timing conflicts, and radar power allocation conflicts;

[0090] By constructing the conflict matrix, the matrix elements are: C_xy=α1*|m_x∩m_y|+α2*|t_x∩t_y|+α3*(P_x+P_y-P_max),

[0091] Where m_x∩m_y represents the overlapping area of ​​radar beam coverage under the xth and yth instructions; t_x∩t_y represents the overlapping duration of instruction execution under the xth and yth instructions; P_x, P_y, and P_max are the radar power and maximum radar power under the xth and yth instructions, respectively; α1, α2, and α3 are the weight coefficients of radar beam coverage conflict, instruction execution timing conflict, and radar power allocation conflict, respectively.

[0092] By traversing all elements in the conflict matrix, when C_xy>0.35, it is determined that there is a conflict between the xth instruction and the yth instruction;

[0093] By calculating the radar's available power margin in real time and predicting the resource utilization curve in a future time period through Monte Carlo simulation, the confidence interval of the resource utilization rate is obtained;

[0094] When the resources required by an instruction exceed the confidence interval, verifying the instruction is not feasible.

[0095] Specifically, a three-dimensional conflict detection model is constructed, corresponding to radar beam coverage conflicts, instruction execution timing conflicts, and radar power allocation conflicts. Corresponding conflict judgment criteria are defined for each dimension to quantify the potential conflict level between instructions at different levels. A conflict matrix is ​​constructed, in which each element corresponds to the degree of conflict between two instructions. α1, α2, and α3 are weight coefficients for radar beam coverage conflicts, instruction execution timing conflicts, and radar power allocation conflicts, respectively, specifically 0.5, 0.3, and 0.2. Their values ​​are determined based on the relative importance of different conflict types in actual application scenarios, and the sum of α1, α2, and α3 is 1. All elements in the conflict matrix are traversed. If the matrix element value is greater than 0.35, a conflict is determined between instructions x and y. The threshold of 0.35 can be dynamically adjusted based on the instruction response priority. This indicates that there is an irreconcilable conflict between the two instructions in their use of radar resources, requiring further processing to avoid resource conflict. The radar's available power margin is calculated in real time, which involves monitoring the radar's current power consumption and total power capacity. Using Monte Carlo simulation, we predict resource utilization over a future timeframe based on current instruction execution and historical data, and calculate a confidence interval for resource utilization. If the resources required by an instruction exceed the predicted confidence interval, the system deems the instruction infeasible. This means that, given the current resource allocation, executing the instruction could overload system resources or impact the execution of other critical instructions, and therefore requires rejection or adjustment.

[0096] In this embodiment, step S4 dynamically executes instructions based on multi-objective game and priority, including the following steps: when multiple voice commands are triggered simultaneously, the Nash equilibrium solution is calculated based on the urgency value, target response priority and radar load, and time division multiplexing technology is used to divide the radar beam into independent time slots at the microsecond level to execute different instructions respectively.

[0097] Specifically, relevant information about all simultaneously triggered voice commands is collected, including urgency, target response priority, and radar load. The collected data is preprocessed, such as through normalization. A game model is constructed, treating multiple voice commands as multiple game participants. The goal of each command is to prioritize radar resources to execute its mission. A strategy set is defined for each command, including different combinations of execution time and resource requirements. A payoff function is established, taking into account factors such as urgency, target response priority, and radar load, to quantify the payoff of each command under different strategy combinations. Game theory algorithms, such as iterative elimination of inferior strategies, best response dynamics, or simulated annealing, are used to find a Nash equilibrium solution. A Nash equilibrium solution corresponds to a strategy combination in which no single command can increase its payoff by unilaterally changing its strategy. Based on the execution order and time allocation determined by the Nash equilibrium solution, time division multiplexing technology is used to divide the radar beam into multiple independent time slots at the microsecond level. Each voice command is assigned a corresponding time slot, ensuring that the corresponding radar operation is independently executed within the allocated time. Radar control operations corresponding to different commands are executed sequentially according to the assigned time slots. In another embodiment, an optimization algorithm, such as the Hungarian algorithm, can be used to solve the conflict matrix, with the goal of minimizing the total amount of conflicts, i.e., sumC_xy, where the sum is calculated across all conflicting instruction pairs. The Hungarian algorithm finds the optimal solution by finding the minimum weight matching in the conflict matrix. This involves adjusting parameters such as the instruction execution order, beam pointing, or power allocation to minimize the total amount of conflicts. Based on the results of the Hungarian algorithm, a set of conflict-free or conflict-minimizing instruction execution plans can be obtained, which serve as the final execution sequence for the radar control instructions.

[0098] In this embodiment, monitoring and evaluating the deviation value after executing the radar control instruction in step S4 and establishing an error reporting mechanism include the following steps:

[0099] By real-time monitoring of radar operating status including radar transmission frequency, scanning angle, power and echo signal-to-noise ratio, the target response including target response time is monitored and normalized;

[0100] By weighted evaluation of the deviation from the actual target, when the deviation exceeds the preset threshold, an error message is given to the user.

[0101] Specifically, real-time collection of key operating parameters such as radar transmit frequency, scan angle, power, and echo signal-to-noise ratio (SNR) can reflect the radar's operating status and performance. For example, a frequency counter measures the radar transmit frequency, an angle encoder acquires scan angle information, a power sensor detects the radar transmit power, and a SNR estimation algorithm calculates the echo SNR. Simultaneously, the target's response to radar control commands is monitored by installing response monitoring sensors on the target device or system to record the time delay from receiving the radar command to executing the corresponding action. The collected radar operating status parameters and target response time are normalized and mapped to the same numerical range, such as between 0 and 1. Weighting coefficients are set based on the impact of each monitored parameter on the radar's control effectiveness. The weighting coefficients for radar transmit frequency, scan angle, power, echo SNR, and target response time are 0.25, 0.2, 0.1, 0.25, and 0.2, respectively. These weighting coefficients can be determined based on expert experience, statistical analysis of historical data, or optimization using machine learning algorithms. Using preset ideal target values ​​such as ideal transmission frequency, scanning angle, power, echo signal-to-noise ratio, and target response time as a benchmark, the degree of deviation between the actual monitored value and the ideal target value is calculated. For each parameter, the deviation value can be expressed as the difference between the actual value and the ideal value or as a relative error. The deviation values ​​of each parameter are then weighted and summed according to the weight coefficient to obtain a comprehensive deviation value. The deviation value threshold is pre-set at 0.75, which can be determined and adjusted based on factors such as the design requirements of the radar system, the mission type, and historical operating data. During actual operation, if the calculated comprehensive deviation value exceeds this preset threshold, it indicates that there is a significant deviation between the actual operating status of the radar and the expected target, which may affect the normal execution of the command, and an error message will be immediately fed back to the user.

[0102] See also Figure 2 As shown, the present invention is a radar voice interactive control system based on natural language processing, which includes the following modules:

[0103] Data acquisition and preprocessing module: In the greenhouse, radar echo signals, voice signals and hyperspectral image data are collected synchronously and preprocessed using a multi-threaded parallel preprocessing mechanism;

[0104] State detection and semantic analysis module: Target state detection is performed by combining pre-processed radar echo signals with hyperspectral image data. The pre-processed voice signals are then subjected to voice recognition to analyze the semantics of voice commands and extract control intent and target parameters.

[0105] Dynamic Control Logic Management Module: This module calculates the urgency level based on the target status detection results by calculating the comprehensive environmental coefficient; dynamically binds the target parameters in the voice command to the radar control logic; generates radar control commands and hierarchical control signals through natural language processing; establishes a conflict detection mechanism, and verifies whether the command is feasible;

[0106] Control execution and status monitoring module: dynamically executes instructions based on multi-objective game and priority; monitors and evaluates deviation values ​​after executing radar control instructions and establishes an error reporting mechanism.

[0107] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A radar voice interactive control method based on natural language processing, characterized in that: The following steps are involved: S1: In the greenhouse, radar echo signals, voice signals and hyperspectral image data are collected synchronously and preprocessed using a multi-threaded parallel preprocessing mechanism; S2: Target status detection is performed by combining pre-processed radar echo signals with hyperspectral image data. The pre-processed voice signals are then subjected to voice recognition to analyze the semantics of voice commands and extract control intent and target parameters. S3: Calculates the urgency value by calculating the comprehensive environmental coefficient and the target status detection results; dynamically binds the target parameters in the voice command with the radar control logic, and generates radar control commands and hierarchical control signals through natural language processing; establishes a conflict detection mechanism and verifies whether the command is feasible; S4: Dynamically execute instructions based on multi-target game and priority; monitor and evaluate deviation values ​​after executing radar control instructions and establish an error reporting mechanism.

2. The radar voice interactive control method based on natural language processing according to claim 1, characterized in that: In step S2, target state detection is performed by combining the pre-processed radar echo signal with the hyperspectral image data, including the following steps: Based on the radar beam width and crop planting density, an adaptive algorithm is used to divide the greenhouse into n rectangular grids; Using a sliding window mechanism, the time-series echo intensity and phase difference data of the i-th rectangular grid are continuously collected, where i∈{1, 2, ..., n}. Vegetation index features are extracted from the hyperspectral image data. The time-series echo intensity, phase difference data, and vegetation index are fused to construct a joint feature vector. A lightweight hybrid network model is constructed by combining convolutional neural networks and long short-term memory networks. A joint feature vector dataset is collected and the target status is labeled, including health status and pest and disease status. The labeled dataset is input into the lightweight hybrid network model for training. The trained lightweight hybrid network model outputs the target status detection results through real-time processing of radar echo signals and hyperspectral image data.

3. The radar voice interactive control method based on natural language processing according to claim 2, characterized in that: In step S2, the pre-processed voice signal is subjected to voice recognition, and the semantics of the voice command is analyzed to extract the control intention and target parameters, including the following steps: Collect historical pre-processed speech signals and input them into a deep learning-based speech recognition model for training. During the training process, the speech recognition model decodes the speech and outputs the corresponding speech commands in text form. Input the real-time preprocessed speech signal into the trained speech recognition model and output the recognition result; preprocess the speech command text form and add greenhouse rectangular grid area marks; Collect and preprocess a voice command text dataset containing various voice commands and their corresponding control intent labels; The preprocessed voice command text dataset is input into the BERT model for training. The real-time voice command text is input into the trained BERT model to output the corresponding control intent label. The token-level classification capability of the BERT model is used to identify the target parameters in the voice command text.

4. The radar voice interactive control method based on natural language processing according to claim 1, characterized in that: The step S3 includes the following steps: calculating the comprehensive environmental coefficient and calculating the urgency value according to the target state detection result: The formula for calculating the comprehensive environmental coefficient is: Among them, H is the environmental factor, H T and H D Represents the ambient temperature and humidity respectively, R Ti 、R Di and R Fi represent the soil temperature, moisture and fertility of the i-th rectangular grid respectively; and Respectively represent the optimal ambient temperature and ambient humidity; and Represent the optimal soil temperature, moisture and fertility respectively; ε1 and ε2 are the influence weights; The formula for calculating urgency is: Among them, E i is the urgency value in the i-th rectangular grid, S i is the planting area of ​​pest-infested crops in the i-th rectangular grid, A is the total area, ΔV i is the vegetation index deviation value, V max is the maximum value of vegetation index deviation, B i is the pest and disease diffusion rate, H is the comprehensive environmental coefficient, σ1 is the confidence level output by the BERT model, σ2 is the confidence level output by the lightweight hybrid network model, and w1, w2, and w3 represent weight coefficients, respectively.

5. The radar voice interactive control method based on natural language processing according to claim 4, characterized in that: In step S3, the target parameters in the voice command are dynamically bound to the radar control logic, and the radar control command and hierarchical control signal are generated through natural language processing, including the following steps: Bind the target state detection results and the target parameters in the voice command to build a mapping table, store the mapping table in the control logic database, and generate radar control operation instructions; Based on the information in the mapping table, the corresponding radar control instruction text is generated using natural language processing technology according to the target parameters and the bound control logic, and converted into a coded format; Analyze control intent labels and target parameters, and match preset operation templates from the control logic database; divide control intent into three response priorities: high, medium, and low, based on the urgency value; According to the divided priorities, hierarchical control signals are generated, including different priority identifiers and corresponding control operation instructions; the generated control signals are used to control the working state of the radar.

6. The radar voice interactive control method based on natural language processing according to claim 5, characterized in that: The step S3 establishes a conflict detection mechanism and verifies whether the instruction is feasible, including the following steps: Create a three-dimensional conflict detection space, including: radar beam coverage conflicts, instruction execution timing conflicts, and radar power allocation conflicts; By constructing the conflict matrix, the matrix elements are: C_xy=α1*|m_x∩m_y|+α2*|t_x∩t_y|+α3*(P_x+P_y-P_max), Where m_x∩m_y represents the overlapping area of ​​radar beam coverage under the xth and yth instructions; t_x∩t_y represents the overlapping duration of instruction execution under the xth and yth instructions; P_x, P_y, and P_max are the radar power and maximum radar power under the xth and yth instructions, respectively; α1, α2, and α3 are the weight coefficients of radar beam coverage conflict, instruction execution timing conflict, and radar power allocation conflict, respectively. By traversing all elements in the conflict matrix, when C_xy>0.35, it is determined that there is a conflict between the xth instruction and the yth instruction; By calculating the radar's available power margin in real time and predicting the resource utilization curve in a future time period through Monte Carlo simulation, the confidence interval of the resource utilization rate is obtained; When the resources required by an instruction exceed the confidence interval, verifying the instruction is not feasible.

7. The radar voice interactive control method based on natural language processing according to claim 1, characterized in that: In step S4, instructions are dynamically executed based on multi-objective game and priority, including the following steps: when multiple voice commands are triggered simultaneously, a Nash equilibrium solution is calculated based on the urgency value, target response priority and radar load, and time division multiplexing technology is used to divide the radar beam into independent time slots at the microsecond level to execute different instructions respectively.

8. The radar voice interactive control method based on natural language processing according to claim 1, characterized in that: After executing the radar control command in step S4, monitoring and evaluating the deviation value and establishing an error reporting mechanism include the following steps: By real-time monitoring of radar operating status including radar transmission frequency, scanning angle, power and echo signal-to-noise ratio, the target response including target response time is monitored and normalized; By weighted evaluation of the deviation from the actual target, when the deviation exceeds the preset threshold, an error message is given to the user.

9. A radar voice interactive control system based on natural language processing, using the radar voice interactive control method based on natural language processing as described in any one of claims 1 to 8, characterized in that: Includes the following modules: Data acquisition and preprocessing module: In the greenhouse, radar echo signals, voice signals and hyperspectral image data are collected synchronously and preprocessed using a multi-threaded parallel preprocessing mechanism; State detection and semantic analysis module: Target state detection is performed by combining pre-processed radar echo signals with hyperspectral image data. The pre-processed voice signals are then subjected to voice recognition to analyze the semantics of voice commands and extract control intent and target parameters. Dynamic Control Logic Management Module: This module calculates the urgency level based on the target status detection results by calculating the comprehensive environmental coefficient; dynamically binds the target parameters in the voice command to the radar control logic; generates radar control commands and hierarchical control signals through natural language processing; establishes a conflict detection mechanism, and verifies whether the command is feasible; Control execution and status monitoring module: dynamically executes instructions based on multi-objective game and priority; monitors and evaluates deviation values ​​after executing radar control instructions and establishes an error reporting mechanism.

Images