Agricultural dialect-oriented robust voice control method for water and fertilizer integrated machine

Abstract

A robust voice control method for integrated water and fertilizer machines, oriented towards agricultural dialects, belongs to the field of agricultural automation technology. It solves the problems of low accuracy in dialect command recognition and difficulty in syntactic parsing under complex noise environments in farmland. The method includes: acquiring speech signals and performing mixed feature extraction and embedding to generate speech vectors specific to agricultural scenarios; segmenting dialect commands using a vocabulary of agricultural water and fertilizer terms and a generative adversarial network; constructing a syntactic tree using a dependency parsing model and template library, combined with an attention mechanism, to extract core predicates and parameter slots; performing intent completion through multi-turn context and sentiment analysis, mapping it to device protocol commands; and finally executing the commands and dynamically updating the template library based on environmental feedback. This invention achieves accurate conversion from dialect speech to device commands, significantly improving the robustness, accuracy, and practicality of voice interaction in integrated water and fertilizer machines in agricultural scenarios.

Description

Technical Field

[0001] This invention belongs to the field of agricultural automation technology. Background Technology

[0002] In agricultural scenarios, the logical integration of dependency syntax template library and fast intent completion for voice command recognition of integrated water and fertilizer machines faces a series of interconnected technical challenges.

[0003] First, when dealing with agricultural dialects, dependency syntax template libraries need to dynamically adapt to the differences in syntactic patterns of local language. However, the existing template libraries have limited coverage and are difficult to quickly capture the phrase structure rules of new dialects, resulting in inaccurate dependency relationship labeling during syntax tree construction, which in turn affects the extraction of core control predicates.

[0004] Secondly, the fast intent completion module relies on historical instruction sequence pattern matching. However, in agricultural scenarios, users' voice intonation is greatly affected by emotional fluctuations. The existing emotional tendency judgment logic cannot accurately distinguish the relationship between intonation changes and actual intent, resulting in priority sorting deviations during intent completion. This can easily lead to semantic inconsistencies after mapping to the control protocol format.

[0005] More importantly, the two problems mentioned above are intertwined: the incompleteness of the syntactic template library will directly lead to the lack of reliable syntactic structure support for the intent completion module, while the deviation of intent completion will in turn affect the dynamic update effect of the template library, forming a closed-loop problem.

[0006] Furthermore, the complexity of background noise in agricultural scenarios further exacerbates the difficulty of speech feature extraction and syntactic analysis. If the feedback data from the noise filtering module is not linked with the template library update mechanism in a timely manner, it may lead to failure of real-time adaptive adjustment during instruction execution, resulting in low accuracy of speech recognition and affecting the accuracy of equipment control. Summary of the Invention

[0007] This invention aims to address the problem of low accuracy in recognizing dialect commands and difficulties in syntactic parsing in complex, noisy farmland environments caused by the operation commands of integrated water and fertilizer machines. A robust voice control method for integrated water and fertilizer machines oriented towards agricultural dialects is proposed.

[0008] The robust voice control method for integrated water and fertilizer machines oriented towards agricultural dialects as described in this invention includes:

[0009] Step 1: Collect the speech signal with dialect from the control of the water and fertilizer integrated machine. Use the agricultural scene background noise filtering module to filter the speech signal. Perform hybrid feature calculation on the filtered speech signal to obtain the initial acoustic feature vector. Input the initial acoustic feature vector into the pre-established embedding model for feature space mapping to obtain preliminary general speech features.

[0010] Step 2: The preliminary general speech features are enhanced by using multi-layer transformation mapping combined with residual connection network. At the same time, the farmland operation context is embedded through semantic weight adjustment mechanism to generate a speech vector sequence for agricultural scenarios containing time step markers.

[0011] Step 3: Based on the agricultural scenario-specific speech vector sequence, combine the pre-constructed vocabulary of water, fertilizer, and farmland terms with the vector dimension compression strategy to perform word segmentation processing, and enhance the adaptability to low-quality speech through adversarial generative networks; obtain the preliminary word unit sequence, perform hierarchical parsing, dialect operation feature classification and batch optimization on the preliminary word unit sequence, and then perform deep semantic parsing to obtain the dialect farmland operation instruction word segmentation unit sequence;

[0012] The pre-constructed vocabulary of terms for water and fertilizer farmland includes standard agricultural terms and their corresponding dialect variations, specifically including: verbs for water and fertilizer operations, names of crops and regions, dosage adjectives, time adverbs, and semantic tags for various words in agricultural scenarios;

[0013] The low-quality speech refers to speech signals with a signal-to-noise ratio below a preset threshold (e.g., below 15dB), or speech segments in agricultural scenarios that are severely interfered with by background noise (such as the roar of agricultural machinery, wind noise, etc.) and thus have reduced speech intelligibility.

[0014] Step 4: The sequence of word segments for the dialect farmland operation instructions is parsed using a dependency parsing model. Combined with a dependency parsing template library that stores the rule set of phrase structures for the control context of the integrated water and fertilizer machine, the dependency relationship of the word segments in the agricultural dialect context is calculated using an attention mechanism to obtain a syntactic tree structure with dependency arcs.

[0015] The dependency syntax template library is a pre-stored set of phrase structure rules for the control context of the water and fertilizer integrated machine. The rule set includes: dependency relationship templates between predicates and parameter slots, conversion rules for sentence patterns specific to agricultural dialects, and standardized syntax tree structures for common commands.

[0016] Step 5: Based on the root node and dependency path information in the syntax tree structure, and combined with the dependency path constraint rules in the dependency syntax template library, extract the predicates of the control speech and the related water and fertilizer integrated machine parameter slot filling data, and obtain the set of structured operation instruction triples by using the predicate completion prediction algorithm, combined with the context of the control speech and the preset rules of the dependency syntax template library.

[0017] Step Six: Use the fast intent completion module to map the structured set of operation instruction triples to the predefined control protocol format of the water and fertilizer integrated machine to obtain the protocol instruction sequence; perform semantic consistency verification on the protocol instruction sequence. If the verification passes, obtain the final executable device instruction sequence and execute Step Seven; if it fails, discard the protocol instructions that fail the verification and return to Step Five.

[0018] The fast intent completion module matches historical instruction sequence patterns and priority sorting, and performs intent completion by combining the judgment logic of historical speech intonation and emotional tendency.

[0019] Step 7: Send the final executable device instruction sequence to the agricultural irrigation execution module of the water and fertilizer integrated machine, use the agricultural scene background noise filtering processing module to obtain real-time environmental interference data, obtain real-time adaptive adjustment data during instruction execution, and realize voice control of the water and fertilizer integrated machine.

[0020] Furthermore, in this invention, the specific method for obtaining preliminary general speech features in step one is as follows:

[0021] Step 11: The collected control voice signal is filtered using an agricultural scene background noise filtering module to remove non-voice segments and background interference, resulting in a clean voice data stream.

[0022] Steps 1 and 2: Extract features from the clean speech data stream and calculate the Mel frequency cepstral coefficients and log energy spectrum of the extracted features to obtain a mixed feature set;

[0023] Step 13: Using the hybrid feature set, construct an initial acoustic feature vector to obtain structured features; Step 14: Input the structured features into a pre-established embedding model, which is used to map the structured features into an intermediate feature vector adapted to the agricultural noise environment;

[0024] Step 15: Using the intermediate feature vector, a feature space mapping is employed to obtain general speech features. When the discriminative power of the general speech features in different agricultural noise environments falls below a preset threshold, the dimensions of the general speech features are adjusted until the discriminative power of the general speech features in different agricultural noise environments reaches the preset threshold, resulting in optimized features. These optimized features are then used as the final general speech features. Further, in this invention, step 2, the method for generating a speech vector sequence specifically for agricultural scenarios, including time step markers, is as follows:

[0025] Step 2: First, a network-based processing framework is constructed using a combination of multi-layer transformation and mapping. This framework is then used to enhance the detail information of the initial general speech features to obtain the initial speech signal.

[0026] Step 22: Optimize the initial speech signal using a residual connection mechanism, retaining the detailed information in the original signal to obtain the optimized speech signal;

[0027] Steps 2 and 3: Combine the farmland operation context to calculate the semantic attention weight of the optimized speech signal, and use the semantic attention weight to perform weighted adjustment on the optimized speech signal to obtain a set of speech signals with contextual information.

[0028] Step 24: Perform time-step segmentation and labeling on the speech signal set with contextual information to generate a structured intermediate vector sequence;

[0029] Step 25: Perform detail feature enhancement processing on the structured intermediate vector sequence to highlight the speech parts related to farmland operations, and obtain the enhanced speech signal sequence;

[0030] Step 26: Perform temporal correlation coding and feature aggregation on the enhanced speech signal sequence to generate the final speech vector sequence for agricultural scenarios, which includes time step markers.

[0031] Furthermore, in this invention, the method for obtaining the sequence of dialect farmland operation instruction word segments in step three is as follows:

[0032] Step 3: 1. Use a vocabulary list of terms for water and fertilizer farmland, and combine principal component analysis or sparse coding to perform dimensionality compression and initial word segmentation on the speech vector sequence for the agricultural scenario. 2. Use an adversarial generative network to adaptively enhance low-quality speech and obtain a preliminary sequence of dialect farmland operation instruction word units.

[0033] Step 3.2: Using a hierarchical parsing method, match the instruction units in the preliminary dialect farmland operation instruction unit sequence with the water and fertilizer farmland terminology vocabulary to obtain the semantic unit set corresponding to the water and fertilizer farmland terminology vocabulary.

[0034] Step 3: Based on the dialect operation features in the speech vector sequence, classify the semantic unit set using a preset classification rule to obtain the classified instruction unit set;

[0035] The preset classification rules include, but are not limited to: intention classification rules based on operational intent, object classification rules based on the operation object, magnitude classification rules based on the operation magnitude, and dialect mapping rules based on agricultural dialect idioms.

[0036] Steps 3 and 4: Combining the dialect operation requirements in agricultural scenarios, a batch processing method is used to associate each unit in the instruction unit set with the low-quality speech adaptation capability enhanced by adversarial generative network, so as to obtain an optimized instruction unit combination.

[0037] Step 35: Perform deep analysis on the optimized instruction unit combination, extract potential semantic information, and obtain the dialect farmland operation instruction word segmentation unit sequence.

[0038] Furthermore, in this invention, step three-three also includes a process of determining whether there is a deviation between the classified instruction unit set and the semantic units in the water and fertilizer farmland terminology list. If there is a deviation, the semantic boundaries of the instruction units are adjusted by comparing the contextual information of water and fertilizer related terms in the speech vector sequence, and the adjusted instruction unit set is used as the classified instruction unit set.

[0039] Furthermore, in this invention, after step three and five, there is also a step of determining whether the sequence of word segments for the dialect farmland operation instructions is consistent with the actual needs of the dialect operation;

[0040] Specifically, the sequence of dialect farmland operation instructions is compared with historical operation instruction data and standard operation intent templates in the water and fertilizer farmland terminology table, and the semantic similarity between the two is calculated. If the semantic similarity is lower than a preset confidence threshold, or if the extracted operation action conflicts with the target object in the pre-built farmland logic rule base, it is determined to be inconsistent. If inconsistent, semantic correction is performed through a pre-built semantic correction mechanism to obtain the final set of farmland operation instruction sequences. The final set of farmland operation instruction sequences is used as the sequence of dialect farmland operation instruction words.

[0041] Furthermore, in this invention, the method for obtaining the syntax tree structure with dependency arcs in step four is as follows:

[0042] Step 41: Combining the phrase structure rules under the pre-constructed control context of the integrated water and fertilizer machine, the dependency parsing model is used to perform preliminary analysis on the word segmentation unit sequence of dialect farmland operation instructions to obtain preliminary dependency association information between each word unit;

[0043] Step 4.2: Based on the preliminary dependency association information, apply an attention mechanism to weight the importance of the syntactic role of each word unit in the word unit sequence in the agricultural dialect context, obtain the dependency weight distribution of all word units, and determine the hierarchical relationship of all word units in the syntactic tree structure based on the dependency weight distribution.

[0044] Step 4.3: Use the dependency syntax template library of the control context phrase structure rule set of the integrated water and fertilizer machine to refine the hierarchical relationship in the syntax tree structure and generate a syntax tree structure with dependency arcs.

[0045] Furthermore, in this invention, step four-three also includes a step of judging whether there is a deviation between the dependency arc of the syntactic tree structure and the syntactic annotation information of the agricultural dialect data. If there is a deviation between the dependency arc of the syntactic tree structure and the syntactic annotation information of the agricultural dialect data, the hierarchical relationship between word units or the dependency arc direction is adjusted by the phrase structure rules stored in the dependency syntax template library to obtain the corrected syntactic tree structure; the corrected syntactic tree structure is used as the syntactic tree structure with dependency arcs.

[0046] Furthermore, in this invention, step five involves obtaining a structured set of operation instruction triples:

[0047] Step 51: Based on the root node information in the syntax tree structure and the parameter slot dependency path constraint rules in the dependency syntax template library, obtain the preliminary division basis of the operation parameter slots and determine the priority order of control predicate extraction.

[0048] Step 52: According to the priority order of the control predicate extraction, the operation parameter slots are filled with content using the preset rules in the dependency syntax template library to obtain the preliminary parameter slot content.

[0049] Step 53: Based on the statistical correlation between the control predicates and the context in the instruction sequence during multi-round historical voice control, a predicate completion prediction algorithm is used to fill in the missing control predicates in the preliminary parameter slot content to obtain the completed parameter slot content.

[0050] The predicate completion prediction algorithm is based on the statistical patterns of historical instruction sequences and uses a sequence prediction model to predict and complete missing control predicates.

[0051] Step 54: Based on the completed parameter slot content, combined with the dependency path rules and root node information in the syntax tree structure, construct a complete set of operation instructions;

[0052] Step 55: Structure and store the root node information, dependency path information, and weight information of the syntax tree in the set of operation instructions with the corresponding parameter slot content to obtain instruction mapping data;

[0053] Steps 5 and 6: Use the preceding instruction information from the multi-turn historical dialogue of voice control to perform semantic integration and optimization on the instruction mapping data, and generate the final set of operation instruction triples.

[0054] The dependency syntax template library stores a database of rule sets for the control context phrase structure of the integrated water and fertilizer machine. The rule set includes dependency relationship templates between core predicates and parameter slots, conversion rules for sentence patterns specific to agricultural dialects, and standardized syntax tree structures for common instructions.

[0055] Furthermore, in this invention, step five-three also includes determining whether the supplemented result meets the structural requirements of the operation instruction set; if the supplemented control predicate is inconsistent with the dependency relationship completion information in the voice control multi-turn historical dialogue relationship, then it is corrected by the preset rules in the dependency syntax template library to obtain the adjusted parameter slot content; the adjusted parameter slot content is used as the completed parameter slot content.

[0056] Furthermore, in this invention, the method for obtaining the final executable device instruction sequence in step six is ​​as follows:

[0057] Step 61: Extract the instruction triplet set from the historical rounds of dialogue, and combine it with the operation instruction triplet set from the current round to extract the main operation intent of the current round of dialogue;

[0058] Step 62: Match the current round's set of operation instruction triples with the instruction patterns in the historical instruction triples set, and correct the matching results according to the main body of the current round's dialogue operation intent to determine candidate intent templates that meet the preset matching requirements;

[0059] Step 63: Analyze the pronunciation features of the current controlled speech using speech intonation parameters to obtain the emotional tendency judgment result in the current speech information;

[0060] Step 64: Based on the sentiment tendency judgment result, perform intent completion on the candidate intent template to generate a set of candidate operation instructions;

[0061] Step 65: Map each candidate instruction in the candidate operation instruction set to the predefined control protocol format of the water and fertilizer integrated machine to obtain the corresponding protocol instruction sequence;

[0062] Step 66: Perform semantic consistency verification on the protocol instruction sequence to determine whether it meets the semantic constraints of the integrated water and fertilizer machine control scenario. If the verification passes, output the current protocol instruction sequence as the final executable device instruction sequence. If the verification fails, discard the current candidate instruction and return to Step 64 to perform intent completion again until an instruction sequence that meets the verification requirements is generated or all candidate instructions are traversed.

[0063] The verification includes: checking whether the execution order of the instructions is consistent with the historical intent backbone, checking whether the adjustment range exceeds the preset safety threshold of the water and fertilizer integrated machine, and checking whether there are conflicting control actions; if all verification items pass, it is determined that the semantic constraint requirements are met; if any item fails, it is determined that it does not meet the requirements.

[0064] Furthermore, in this invention, step seven, the process of realizing voice control of the integrated water and fertilizer machine, is as follows:

[0065] Step 71: Send the final executable device instruction sequence to the agricultural irrigation control unit, and obtain the initial status information of the instruction issuance from the agricultural irrigation control unit;

[0066] Step 72: Based on the initial state information, analyze the response data after the agricultural irrigation control unit sends the command using the pre-established command parsing rules, and determine whether the initial completeness of the command execution reaches a preset threshold; if yes, proceed to step 76; otherwise, proceed to step 73.

[0067] The instruction parsing rules include: comparing the status code returned by the device with the preset successful execution status code, and calculating the deviation between the actual execution parameters and the instruction issuance parameters;

[0068] Step 73: Obtain real-time environmental interference data from the agricultural scene background noise filtering module, analyze the impact of the current noise environment on command execution, and determine the scope of interference.

[0069] Step 74: Based on the noise type and intensity corresponding to the affected area, retrieve historical dialect expression data associated with the noise environment from the dependency syntax template library, and obtain the dialect instruction patterns successfully executed in the noise scenario;

[0070] Step 75: Calculate the semantic similarity between the initial state information of the issued command and the dialect command pattern successfully executed in the noise scenario. If the semantic similarity is lower than a preset threshold, it is determined that the command expression method needs to be adjusted; otherwise, it is determined that no adjustment is needed. If the command expression method needs to be adjusted, the dependency parsing model is used to perform syntactic structure analysis on the dialect command pattern successfully executed in the similar noise scenario to generate a new command format adapted to the dialect expression habits, and obtain the optimized command content. The optimized command content is retransmitted to the agricultural irrigation control unit for execution, and the updated command state information is obtained. The updated command state information replaces the initial state information of the issued command, and step 76 is executed.

[0071] Step 76: Determine whether the initial state information of the issued instruction meets the expected requirements of the agricultural scenario. If it does, generate instruction execution success status information. If it does not, determine whether the current number of adjustments has reached the preset maximum adjustment threshold. If it has not reached the threshold, return to step 73 and readjust the instruction expression method based on the current noise environment. If it has reached the threshold, generate instruction execution failure status information.

[0072] Furthermore, in this invention, step seven also includes step seven seven, which involves recording the final state information of instruction execution to the system log database, and simultaneously updating the dialect expression data involved in this execution process and its corresponding noise environment features and syntactic patterns to the dependency syntax template library.

[0073] This invention effectively solves the key technical problems of low accuracy in dialect command recognition and difficulty in syntactic structure parsing in complex noisy farmland environments by constructing a robust voice control method for agricultural dialects. Starting from underlying acoustic features, this method achieves accurate conversion from dialect speech to equipment control commands through hybrid feature extraction and noise robustness optimization, speech vector generation incorporating farmland context, low-quality speech segmentation enhanced by adversarial networks, dialect dependency parsing based on attention mechanisms, and intent completion combining multi-turn dialogue context and sentiment. Simultaneously, through a dynamic update mechanism of the dependency parsing template library, the system possesses adaptive capabilities to new dialect patterns and complex environments, thus significantly improving the robustness, accuracy, and practicality of voice interaction for integrated water and fertilizer machines in agricultural scenarios. Attached Figure Description

[0074] Figure 1 This is a flowchart of the method described in this invention. Detailed Implementation

[0075] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0076] Specific implementation method one: Refer to Figure 1 This embodiment specifically describes a robust voice control method for integrated water and fertilizer machines based on agricultural dialects, including:

[0077] Step 1: Collect the voice signal controlling the integrated water and fertilizer machine, filter the voice signal using the agricultural scene background noise filtering module, perform hybrid feature calculation on the filtered voice signal to obtain the initial acoustic feature vector, and input the initial acoustic feature vector into the pre-established embedding model for feature space mapping to obtain preliminary general voice features.

[0078] Step 2: The preliminary general speech features are enhanced by using multi-layer transformation mapping combined with residual connection network. At the same time, the farmland operation context is embedded through semantic weight adjustment mechanism to generate a speech vector sequence for agricultural scenarios containing time step markers.

[0079] Step 3: Based on the agricultural scenario-specific speech vector sequence, combine the pre-constructed vocabulary of water, fertilizer, and farmland terms with the vector dimension compression strategy to perform word segmentation processing, and enhance the adaptability to low-quality speech through adversarial generative networks; obtain the preliminary word unit sequence, perform hierarchical parsing, dialect operation feature classification and batch optimization on the preliminary word unit sequence, and then perform deep semantic parsing to obtain the dialect farmland operation instruction word segmentation unit sequence;

[0080] The pre-constructed vocabulary of terms for water and fertilizer farmland includes standard agricultural terms and their corresponding dialect variations, specifically including: verbs for water and fertilizer operations, names of crops and regions, dosage adjectives, time adverbs, and semantic tags for various words in agricultural scenarios;

[0081] The term "low-quality speech" refers to speech signals with a signal-to-noise ratio below a preset threshold (e.g., below 15dB), or speech segments in agricultural settings that are severely interfered with by background noise (such as the roar of agricultural machinery, wind noise, etc.), resulting in decreased speech intelligibility.

[0082] Step 4: The sequence of word segments for the dialect farmland operation instructions is parsed using a dependency parsing model. Combined with a dependency parsing template library that stores the rule set of phrase structures for the control context of the integrated water and fertilizer machine, the dependency relationship of the word segments in the agricultural dialect context is calculated using an attention mechanism to obtain a syntactic tree structure with dependency arcs.

[0083] The dependency syntax template library is a pre-stored set of phrase structure rules for the control context of the water and fertilizer integrated machine. The rule set includes: dependency relationship templates between predicates and parameter slots, conversion rules for sentence patterns specific to agricultural dialects, and standardized syntax tree structures for common instructions.

[0084] Step 5: Based on the root node and dependency path information in the syntax tree structure, and combined with the dependency path constraint rules in the dependency syntax template library, extract the predicates of the control speech and the related water and fertilizer integrated machine slot filling data. Then, through the predicate completion prediction algorithm, combined with the context of the control speech and the preset rules of the dependency syntax template library, obtain the set of structured operation instruction triples.

[0085] The triplet described in this embodiment includes a core control predicate (a predicate closely related to the physical actions of the device, such as switching or adjusting), an operation object / execution logic (target or sequence), and specific parameter constraints (degree, magnitude, or time).

[0086] Step Six: Use the fast intent completion module to map the structured set of operation instruction triples to the predefined control protocol format of the water and fertilizer integrated machine to obtain the protocol instruction sequence; perform semantic consistency verification on the protocol instruction sequence. If the verification passes, obtain the final executable device instruction sequence and execute Step Seven; if it fails, discard the protocol instructions that fail the verification and return to Step Five.

[0087] The fast intent completion module matches historical instruction sequence patterns and priority sorting, and performs intent completion by combining the judgment logic of historical speech intonation and emotional tendency.

[0088] Step 7: Send the final executable device instruction sequence to the agricultural irrigation execution module of the water and fertilizer integrated machine, use the agricultural scene background noise filtering processing module to obtain real-time environmental interference data, obtain real-time adaptive adjustment data during instruction execution, and realize voice control of the water and fertilizer integrated machine.

[0089] Furthermore, in this invention, the specific method for obtaining preliminary general speech features in step one is as follows:

[0090] Step 11: The collected control voice signal is filtered using an agricultural scene background noise filtering module to remove non-voice segments and background interference, resulting in a clean voice data stream.

[0091] Steps 1 and 2: Extract features from the clean speech data stream and calculate the Mel frequency cepstral coefficients and log energy spectrum of the extracted features to obtain a mixed feature set;

[0092] Step 13: Using the aforementioned hybrid feature set, construct an initial acoustic feature vector to obtain structured features;

[0093] Step 14: Input the structured features into a pre-established embedding model, which is used to map the structured features into intermediate feature vectors adapted to agricultural noise environments;

[0094] Step 15: Using the intermediate feature vector, a feature space mapping is used to obtain general speech features. When the discriminative power of the general speech features in different agricultural noise environments is lower than a preset threshold, the dimensions of the general speech features are adjusted until the discriminative power of the general speech features in different agricultural noise environments reaches the preset threshold, thus obtaining optimized features. The optimized features are then used as the final general speech features.

[0095] Furthermore, in this invention, the method for generating a speech vector sequence specifically for agricultural scenarios, including time step markers, in step two is as follows:

[0096] Step 2: First, a network-based processing framework is constructed using a combination of multi-layer transformation and mapping. This framework is then used to enhance the detail information of the initial general speech features to obtain the initial speech signal.

[0097] Step 22: Optimize the initial speech signal using a residual connection mechanism, retaining the detailed information in the original signal to obtain the optimized speech signal;

[0098] Steps 2 and 3: Combine the farmland operation context to calculate the semantic attention weight of the optimized speech signal, and use the semantic attention weight to perform weighted adjustment on the optimized speech signal to obtain a set of speech signals with contextual information.

[0099] Step 24: Perform time-step segmentation and labeling on the speech signal set with contextual information to generate a structured intermediate vector sequence;

[0100] Step 25: Perform detail feature enhancement processing on the structured intermediate vector sequence to highlight the speech parts related to farmland operations, and obtain the enhanced speech signal sequence;

[0101] Step 26: Perform temporal correlation coding and feature aggregation on the enhanced speech signal sequence to generate the final speech vector sequence for agricultural scenarios, which includes time step markers.

[0102] Furthermore, in step three, the method for obtaining the sequence of segmented units for dialect farmland operation instructions is as follows:

[0103] Step 3: 1. Use a vocabulary list of terms for water and fertilizer farmland, and combine principal component analysis or sparse coding to perform dimensionality compression and initial word segmentation on the speech vector sequence for the agricultural scenario. 2. Use an adversarial generative network to adaptively enhance low-quality speech and obtain a preliminary sequence of dialect farmland operation instruction word units.

[0104] Step 3.2: Using a hierarchical parsing method, match the instruction units in the preliminary dialect farmland operation instruction unit sequence with the water and fertilizer farmland terminology vocabulary to obtain the semantic unit set corresponding to the water and fertilizer farmland terminology vocabulary.

[0105] Step 3: Based on the dialect operation features in the speech vector sequence, classify the semantic unit set using a preset classification rule to obtain the classified instruction unit set;

[0106] The preset classification rules include, but are not limited to: intention classification rules based on operational intent, object classification rules based on the operation object, magnitude classification rules based on the operation magnitude, and dialect mapping rules based on agricultural dialect idioms.

[0107] Steps 3 and 4: Combining the dialect operation requirements in agricultural scenarios, a batch processing method is used to associate each unit in the instruction unit set with the low-quality speech adaptation capability enhanced by adversarial generative network, so as to obtain an optimized instruction unit combination.

[0108] Step 35: Perform deep analysis on the optimized instruction unit combination, extract potential semantic information, and obtain the dialect farmland operation instruction word segmentation unit sequence.

[0109] Furthermore, in this invention, step three-three also includes a process of determining whether there is a deviation between the classified instruction unit set and the semantic units in the water and fertilizer farmland terminology list. If there is a deviation, the semantic boundaries of the instruction units are adjusted by comparing the contextual information of water and fertilizer related terms in the speech vector sequence, and the adjusted instruction unit set is used as the classified instruction unit set.

[0110] Furthermore, in this invention, after step three and five, there is also a step of determining whether the sequence of word segments for the dialect farmland operation instructions is consistent with the actual needs of the dialect operation;

[0111] Specifically, the sequence of dialect farmland operation instructions is compared with historical operation instruction data and standard operation intent templates in the water and fertilizer farmland terminology table to calculate the semantic similarity between the two. If the semantic similarity is lower than a preset confidence threshold, or if the extracted operation action (such as "fertilizing") conflicts with the target object (such as "water") in the pre-built farmland logic rule base, it is determined to be inconsistent. If inconsistent, semantic correction is performed through a pre-built semantic correction mechanism to obtain the final set of farmland operation instruction sequences. The final set of farmland operation instruction sequences is used as the sequence of dialect farmland operation instruction words.

[0112] Furthermore, in this invention, the method for obtaining the syntax tree structure with dependency arcs in step four is as follows:

[0113] Step 41: Combining the phrase structure rules under the pre-constructed control context of the integrated water and fertilizer machine, the dependency parsing model is used to perform preliminary analysis on the word segmentation unit sequence of dialect farmland operation instructions to obtain preliminary dependency association information between each word unit;

[0114] Step 4.2: Based on the preliminary dependency association information, apply an attention mechanism to weight the importance of the syntactic role of each word unit in the word unit sequence in the agricultural dialect context, obtain the dependency weight distribution of all word units, and determine the hierarchical relationship of all word units in the syntactic tree structure based on the dependency weight distribution.

[0115] Step 4.3: Use the dependency syntax template library of the control context phrase structure rule set of the integrated water and fertilizer machine to refine the hierarchical relationship in the syntax tree structure and generate a syntax tree structure with dependency arcs.

[0116] Furthermore, in this invention, step four-three also includes a step of judging whether there is a deviation between the dependency arc of the syntactic tree structure and the syntactic annotation information of the agricultural dialect data. If there is a deviation between the dependency arc of the syntactic tree structure and the syntactic annotation information of the agricultural dialect data, the hierarchical relationship between word units or the dependency arc direction is adjusted by the phrase structure rules stored in the dependency syntax template library to obtain the corrected syntactic tree structure; the corrected syntactic tree structure is used as the syntactic tree structure with dependency arcs.

[0117] Furthermore, in this invention, step five involves obtaining a structured set of operation instruction triples:

[0118] Step 51: Based on the root node information in the syntax tree structure and the parameter slot dependency path constraint rules in the dependency syntax template library, obtain the preliminary division basis of the operation parameter slots and determine the priority order of control predicate extraction.

[0119] Step 52: According to the priority order of the control predicate extraction, the operation parameter slots are filled with content using the preset rules in the dependency syntax template library to obtain the preliminary parameter slot content.

[0120] Step 53: Based on the statistical correlation between the control predicates and the context in the instruction sequence during multi-round historical voice control, a predicate completion prediction algorithm is used to fill in the missing control predicates in the preliminary parameter slot content to obtain the completed parameter slot content.

[0121] The predicate completion prediction algorithm is based on the statistical patterns of historical instruction sequences and uses a sequence prediction model to predict and complete missing control predicates.

[0122] Step 54: Based on the completed parameter slot content, combined with the dependency path rules and root node information in the syntax tree structure, construct a complete set of operation instructions;

[0123] Step 55: Structure and store the root node information, dependency path information, and weight information of the syntax tree in the set of operation instructions with the corresponding parameter slot content to obtain instruction mapping data;

[0124] Steps 5 and 6: Use the preceding instruction information from the multi-turn historical dialogue of voice control to perform semantic integration and optimization on the instruction mapping data, and generate the final set of operation instruction triples.

[0125] The dependency syntax template library stores a database of rule sets for the control context phrase structure of the integrated water and fertilizer machine. The rule set includes dependency relationship templates between core predicates and parameter slots, conversion rules for sentence patterns specific to agricultural dialects, and standardized syntax tree structures for common instructions.

[0126] Furthermore, in this invention, step five-three also includes determining whether the supplemented result meets the structural requirements of the operation instruction set; if the supplemented control predicate is inconsistent with the dependency relationship completion information in the voice control multi-turn historical dialogue relationship, then it is corrected by the preset rules in the dependency syntax template library to obtain the adjusted parameter slot content; the adjusted parameter slot content is used as the completed parameter slot content.

[0127] Furthermore, in this invention, the method for obtaining the final executable device instruction sequence in step six is ​​as follows:

[0128] Step 61: Extract the instruction triplet set from the historical rounds of dialogue, and combine it with the operation instruction triplet set from the current round to extract the main operation intent of the current round of dialogue;

[0129] Step 62: Match the current round's set of operation instruction triples with the instruction patterns in the historical instruction triples set, and correct the matching results according to the main body of the current round's dialogue operation intent to determine candidate intent templates that meet the preset matching requirements;

[0130] Step 63: Analyze the pronunciation features of the current controlled speech using speech intonation parameters to obtain the emotional tendency judgment result in the current speech information;

[0131] Step 64: Based on the sentiment tendency judgment result, perform intent completion on the candidate intent template to generate a set of candidate operation instructions;

[0132] Step 65: Map each candidate instruction in the candidate operation instruction set to the predefined control protocol format of the water and fertilizer integrated machine to obtain the corresponding protocol instruction sequence;

[0133] Step 66: Perform semantic consistency verification on the protocol instruction sequence to determine whether it meets the semantic constraints of the integrated water and fertilizer machine control scenario. If the verification passes, output the current protocol instruction sequence as the final executable device instruction sequence. If the verification fails, discard the current candidate instruction and return to Step 64 to perform intent completion again until an instruction sequence that meets the verification requirements is generated or all candidate instructions are traversed.

[0134] The verification includes: checking whether the execution order of the instructions is consistent with the historical intent backbone, checking whether the adjustment range exceeds the preset safety threshold of the water and fertilizer integrated machine, and checking whether there are conflicting control actions; if all verification items pass, it is determined that the semantic constraint requirements are met; if any item fails, it is determined that it does not meet the requirements.

[0135] Furthermore, in this invention, step seven, the process of realizing voice control of the integrated water and fertilizer machine, is as follows:

[0136] Step 71: Send the final executable device instruction sequence to the agricultural irrigation control unit, and obtain the initial status information of the instruction issuance from the agricultural irrigation control unit;

[0137] Step 72: Based on the initial state information, analyze the response data after the agricultural irrigation control unit sends the command using the pre-established command parsing rules, and determine whether the initial completeness of the command execution reaches a preset threshold; if yes, proceed to step 76; otherwise, proceed to step 73.

[0138] The instruction parsing rules include: comparing the status code returned by the device with the preset successful execution status code, and calculating the deviation between the actual execution parameters and the instruction issuance parameters;

[0139] Step 73: Obtain real-time environmental interference data from the agricultural scene background noise filtering module, analyze the impact of the current noise environment on command execution, and determine the scope of interference.

[0140] Step 74: Based on the noise type and intensity corresponding to the affected area, retrieve historical dialect expression data associated with the noise environment from the dependency syntax template library, and obtain the dialect instruction patterns successfully executed in the noise scenario;

[0141] Step 75: Calculate the semantic similarity between the initial state information of the issued command and the dialect command pattern successfully executed in the noise scenario. If the semantic similarity is lower than a preset threshold, it is determined that the command expression method needs to be adjusted; otherwise, it is determined that no adjustment is needed. If the command expression method needs to be adjusted, the dependency parsing model is used to perform syntactic structure analysis on the dialect command pattern successfully executed in the similar noise scenario to generate a new command format adapted to the dialect expression habits, and obtain the optimized command content. The optimized command content is retransmitted to the agricultural irrigation control unit for execution, and the updated command state information is obtained. The updated command state information replaces the initial state information of the issued command, and step 76 is executed.

[0142] Step 76: Determine whether the initial state information of the issued instruction meets the expected requirements of the agricultural scenario. If it does, generate instruction execution success status information. If it does not, determine whether the current number of adjustments has reached the preset maximum adjustment threshold. If it has not reached the threshold, return to step 73 and readjust the instruction expression method based on the current noise environment. If it has reached the threshold, generate instruction execution failure status information.

[0143] Furthermore, in this invention, step seven also includes step seven seven, which involves recording the final state information of instruction execution to the system log database, and simultaneously updating the dialect expression data involved in this execution process and its corresponding noise environment features and syntactic patterns to the dependency syntax template library.

[0144] like Figure 1 This embodiment of a robust voice control method for integrated water and fertilizer machines based on agricultural dialects may specifically include: Step S101: By performing a hybrid feature calculation method combining Mel frequency cepstral coefficients and logarithmic energy spectrum on the input speech signal, an initial acoustic feature vector is obtained, and then input into an embedding model that is optimized for robustness in specific noise environments for agricultural dialect datasets, to obtain a preliminary general speech feature representation.

[0146] By performing preprocessing operations on the acquired speech signal, non-speech segments and background interference are removed to obtain a clean speech data stream;

[0147] Feature extraction is performed using a clean speech data stream, and the Mel frequency cepstral coefficients and log energy spectrum are calculated to generate a mixed feature set;

[0148] By utilizing the generated hybrid feature set, an initial acoustic feature vector is constructed to obtain a structured feature representation.

[0149] The structured feature representation is input into a pre-built embedding model, and the optimized parameters of the agricultural dialect dataset are combined to determine the intermediate feature vector adapted to the specific noise environment.

[0150] The intermediate feature vectors are mapped into the feature space to obtain a general speech feature representation.

[0151] If the discriminative power of the general speech feature representation in a specific noise environment is lower than a preset threshold, then its dimensions are adjusted to obtain the optimized final feature representation.

[0152] By optimizing the final feature representation, robust processing of speech signals in agricultural dialect scenarios is completed, and its applicability in different noise environments is determined.

[0153] Specifically, by preprocessing the collected agricultural scene voice signals, non-speech segments such as field wind noise, farm machinery roar, and distant human voices can be effectively removed, resulting in a relatively clean voice data stream.

[0154] For example, when voice recordings are made while a tractor is operating, after endpoint detection and spectral subtraction, the background noise energy decreases by an average of about 18 decibels, and the clarity of the voice segments is significantly improved.

[0155] In one embodiment, a 13-dimensional Mel-frequency cepstral coefficient and a 1-dimensional logarithmic energy spectrum are calculated using a clean speech data stream, forming a 14-dimensional hybrid feature set. This combination method preserves the characteristics of human auditory perception while supplementing energy information, which is beneficial for subsequent models to distinguish voiced sounds from background interference.

[0156] When constructing the initial acoustic feature vector for a mixed feature set, the features of each frame are typically concatenated into a fixed-length vector, for example, 25 consecutive frames forming a 350-dimensional structured representation. This structured approach facilitates the embedding model in capturing temporal speech variation patterns.

[0157] Preferably, by processing the aforementioned structured features using an embedding model fine-tuned on an agricultural dialect dataset, and combining parameters optimized for low-frequency tractor noise and wind noise, an intermediate feature vector adapted to the specific noise environment of the field can be obtained. This intermediate vector improves the discriminative power of noise robustness by approximately 12% compared to a general embedding.

[0158] Specifically, when generating a general speech feature representation through feature space mapping, if the inter-class discrimination is lower than a preset threshold of 0.78 under the high-intensity noise environment of a harvester, a dimensionality adjustment process is triggered. Principal component analysis is typically used to retain the first 90% of the energy information, compressing the dimensionality from 512 to between 280 and 320 dimensions, resulting in the optimized final feature representation. For example, in a scenario of recording voices in a cornfield, the general features without dimensionality adjustment have a recognition rate of only 67% near a harvester, while the final feature representation after the above optimization improves the recognition rate to over 89%, fully demonstrating the balance between dimensionality reduction and noise reduction and information preservation. It should be noted that this complete link from preprocessing to feature optimization to applicability judgment enables the system to maintain high robustness under different agricultural operation noise environments, significantly reducing false recognition and rejection caused by environmental changes. In practical applications, it can reduce the overall word error rate by approximately 21 percentage points.

[0159] Step S102: Multi-layer transform mapping combined with residual connection network is used to retain the detailed information of the original speech signal. At the same time, the contextual features of farmland operation are embedded through semantic weight adjustment mechanism to generate a speech vector sequence for agricultural scenarios containing time step markers, and to determine its applicability to subsequent dialect rhythm feature capture.

[0160] By employing multi-layer transformation and mapping, a network-based processing framework is constructed to process the detailed information in the original signal and obtain a preliminary speech signal.

[0161] A residual connection mechanism is used to optimize the initial speech signal, retaining the detailed information in the original signal and determining the form of the optimized signal.

[0162] By incorporating semantic weights and adjustment mechanisms to adjust the optimized signal representation, and then incorporating relevant characteristics of the farmland context, a set of speech signals with contextual information is obtained.

[0163] The application time step is used to label a set of speech signals with contextual information, generate a structured vector sequence, and determine its applicability in agricultural scenarios;

[0164] The enhanced speech signal sequence is obtained by extracting the parts related to detailed information from the structured vector sequence and combining them with the characteristics of the farmland context.

[0165] By using the enhanced speech signal sequence and employing feature integration, a dedicated speech vector sequence suitable for agricultural scenarios is constructed, and the dedicated speech vector sequence for agricultural scenarios is obtained.

[0166] The time step of the speech vector sequence for agricultural scenarios is organized and its correlation with the vector sequence is analyzed to generate the final speech vector set for agricultural scenarios, and its suitability for subsequent processing is determined.

[0167] For example, when building a network-based processing framework, the original speech signal can be processed in layers using a multi-layer neural network to gradually extract detailed information from the signal.

[0168] The initial layer can focus on the low-frequency characteristics of the signal, while subsequent layers gradually focus on high-frequency details and time-domain changes, forming a feature extraction process that progresses from shallow to deep.

[0169] In one possible implementation, assuming the input signal has a sampling rate of 16kHz, local features can be extracted first through a convolutional layer, and then the dimensionality can be reduced through a pooling layer, ultimately mapping to a preliminary speech signal representation.

[0170] This layered design helps capture the unique patterns of speech signals in farmland environments.

[0171] For example, to optimize the residual connection mechanism, skip connections can be introduced into the network to directly add the preliminary signal representation to a portion of the original input signal, preventing the loss of details.

[0172] Specifically, when dealing with background noise common in agricultural scenarios, residual connections can preserve subtle speech details in the original signal, such as voices being muffled by the wind in a field.

[0173] In one embodiment, residual paths can be added between every two network layers to ensure that the optimized signal representation contains both high-level abstract features and retains the subtle changes in the original signal.

[0174] For example, when incorporating semantic weights and adjustment mechanisms, specific weight allocation strategies can be designed based on the farmland context to highlight semantic information related to agriculture.

[0175] Suppose that when processing speech, the system identifies signal segments related to keywords such as "irrigation" and "sowing". The weight of these segments can be increased through an attention mechanism to generate a set of speech signals with contextual information.

[0176] This approach is particularly suitable for processing command-based speech in agricultural settings.

[0177] For example, when applying time step marking, the speech signal can be segmented into 10ms units, and a label can be assigned to each time step to form a structured vector sequence. It should be noted that this marking method facilitates subsequent assessment of the signal's applicability in agricultural scenarios, such as distinguishing between short, abrupt command speech and continuous environmental noise.

[0178] By using time step marking, the system can more accurately locate key speech segments. For example, when extracting detailed information and combining it with the characteristics of the farmland context, the system can enhance the high-energy parts of the structured vector sequence to generate an enhanced speech signal sequence.

[0179] Specifically, the noise of tractors commonly heard in farmland can be suppressed while amplifying the clearer parts of the speech signal to create a more suitable signal representation for agricultural scenarios. For example, when constructing a dedicated speech representation for agricultural scenarios, contextual information can be combined with signal features through feature integration to generate a more targeted representation.

[0180] In one embodiment, the frequency domain features of the speech signal can be combined with semantic tags to form a multi-dimensional feature vector, suitable for speech processing of different dialects in agricultural environments. For example, when organizing the correlation between time steps and vector sequences, the final set of agricultural scene speech vectors can be normalized through time alignment. It should be noted that this organization method helps with adaptability in subsequent processing, such as ensuring stable parsing of voice commands on different devices. Through the above multi-faceted design, the system can better adapt to complex environments in agricultural scenarios, improving the practicality of speech processing.

[0181] Step S103: Based on the generated agricultural scenario-specific speech vector sequence, combined with the pre-constructed vocabulary of water and fertilizer farmland terms and vector dimension compression principal component analysis and sparse coding dimensionality reduction strategy, word segmentation is performed. At the same time, the adaptability to low-quality speech is enhanced by adversarial generative network to obtain a preliminary dialect farmland operation instruction word unit sequence.

[0182] By using a preliminary sequence of dialect farmland operation instruction units, a hierarchical parsing method is employed to match the instruction units with water and fertilizer terms in agricultural scenarios, obtain a set of semantic units corresponding to the farmland vocabulary, and determine the preliminary semantic mapping results.

[0183] Based on the semantic mapping results, the dialect operation features in the speech vector are grouped according to the preset classification rules to obtain the classified instruction unit set, and their semantic consistency in the agricultural scenario is judged.

[0184] If there is a discrepancy between the classified instruction unit set and the semantic units in the farmland vocabulary, the semantic boundaries of the instruction units are adjusted by comparing the contextual information of water and fertilizer terms, and the adjusted instruction unit sequence is obtained to determine its semantic accuracy.

[0185] In combination with the dialect operation requirements in agricultural scenarios, a batch processing method is adopted to associate the adjusted instruction unit sequence with the adaptability of low-quality speech, resulting in an optimized instruction unit combination.

[0186] Based on the optimized instruction unit combination, information extraction tools are used to perform deep analysis on the potential semantic information in the speech vector to obtain a set of semantic instructions highly relevant to the agricultural scenario and determine their matching degree in actual operation.

[0187] If the set of semantic instructions is inconsistent with the actual needs of dialect operations, the instruction set is processed a second time through a pre-built semantic correction mechanism to obtain the final sequence of farmland operation instructions and determine its applicability in agricultural scenarios.

[0188] By combining the semantic information of water and fertilizer terminology, data storage tools are used to structurally save the final farmland operation instruction sequence, obtain an instruction database that can be called, and determine its availability in subsequent processing.

[0189] For example, when processing dialect farmland operation instruction word sequences, a hierarchical parsing approach can be used to match instruction units with water and fertilizer terms in agricultural scenarios. The core of hierarchical parsing lies in decomposing complex instructions into smaller semantic units and analyzing their relevance to the farmland vocabulary layer by layer. Suppose a dialect instruction is "hurry up and irrigate the land." The first layer of parsing breaks it down into two units: "hurry up" and "irrigate the land." The second layer maps "irrigate the land" to "irrigate" in the vocabulary, forming a preliminary semantic mapping. This method can progressively refine the semantic expression of the instruction, ensuring its fit with the agricultural scenario. For instance, when grouping dialect operation features in speech vectors, instruction units can be categorized into two main types based on operation type: "water management" and "fertilizer application," according to predefined classification rules.

[0190] Suppose a set of instructions contains two operations: "watering" and "fertilizing." Using classification rules, "watering" is categorized under "water management," and its pitch variation features are extracted from its speech vector. This grouping helps in subsequent semantic consistency assessments, providing a foundation for precise operations. For example, when adjusting the semantic boundaries of instruction units, if a discrepancy is found between the categorized instruction units and the farmland vocabulary, it can be corrected using contextual information.

[0191] Suppose a dialect instruction is "water more," but it's initially mapped as "irrigate a small amount." By analyzing the contextual information "the land is extremely dry," the semantic boundary can be adjusted to "irrigate heavily." This method improves the semantic accuracy of the instruction and adapts to the actual needs of farmland. For example, when considering the dialect operation requirements in agricultural scenarios and associating instruction units with the adaptability to low-quality speech, a batch processing approach can be used.

[0192] Assuming the speech quality is poor in a noisy farmland environment, batch processing can be used to integrate instruction units from multiple time periods, extract common features, and form optimized instruction combinations. This approach enhances the robustness of instructions in complex environments. For example, when deeply analyzing the latent semantic information in speech vectors, information extraction tools can be used for customized analysis tailored to agricultural scenarios.

[0193] Suppose a set of instructions includes "water the fields tomorrow." Tools can be used to extract semantic information in two categories: "time" and "operation," forming a highly correlated set of semantic instructions. This deep analysis helps improve the matching degree between instructions and actual operations. For example, if the set of semantic instructions is inconsistent with dialectal operation requirements, secondary processing can be performed using a pre-built semantic correction mechanism.

[0194] If an instruction is misinterpreted as "fertilize" instead of "irrigate," a correction mechanism compares it with historical instruction data to readjust it to the correct semantics. This method ensures the accuracy of the final farmland operation instruction sequence. For example, when storing the final instruction sequence in a structured manner, data storage tools can be used to associate and store the instructions with the semantic information of water and fertilizer terms.

[0195] Suppose a database stores three-dimensional information on "irrigation-water volume-time" for easy retrieval later. This storage method improves the availability of the instruction database and facilitates subsequent processing.

[0196] Step S104 involves inputting the word unit sequence obtained from word segmentation into the dependency parsing model, combining it with the dependency parsing template library that stores the phrase structure rule set for the control context of the integrated water and fertilizer machine, and using the attention mechanism to calculate the dependency relationship of each word unit under the specific agricultural dialect syntax preference annotation data, to obtain a syntactic tree structure representation with dependency arcs.

[0197] By using word unit sequences as basic data and combining them with pre-constructed phrase structure rules in the control context of integrated water and fertilizer machines, a dependency syntax model is used for preliminary parsing to obtain preliminary dependency association information between word units.

[0198] Based on the preliminary dependency association information, an attention mechanism is applied to weight the syntactic annotation information in agricultural dialect data to obtain the dependency weight distribution of each word unit in a specific context and determine its hierarchical relationship in the syntactic tree structure.

[0199] By combining the phrase structure rules constructed from the template library, the hierarchical relationships and various dependency associations in the syntactic tree structure are refined to generate intermediate syntactic structures with dependency arcs, and their logical consistency under the control context is judged.

[0200] If there is a discrepancy between the dependency arc generation of the intermediate syntactic structure and the annotation information of the agricultural dialect data, the hierarchical relationship between word units is adjusted by preset rules to obtain the corrected syntactic tree structure.

[0201] Based on the revised syntax tree structure, and considering the results of dependency association analysis, data storage tools are used to structurally save the information of each node in the syntax tree structure, resulting in a syntax parsing database that can be used for subsequent calls.

[0202] Based on the specific instruction requirements of the integrated water and fertilizer machine control context, a batch processing method is applied to map the syntactic parsing database with the actual operation context, obtain the final control instruction sequence, and determine its matching degree in the agricultural scenario.

[0203] In one possible implementation, preliminary parsing of word unit sequences using a dependency syntax model can clearly reveal the dominance and subordination relationships among the various parts of agricultural dialect instructions.

[0204] For example, when a farmer says, "Spread the fertilizer evenly first, then boil water," the model first identifies "spread evenly" as the core action of "fertilizing," while "boiling water" depends on the preceding "then," thus forming a preliminary dependency arc of "spread evenly → fertilizer" and "boiling water → then." This analysis lays the structural foundation for subsequent processing.

[0205] Specifically, by applying an attention mechanism to weight the syntactic annotation information, the importance of each word unit within the control context of the integrated water and fertilizer machine is quantified. Assuming the instruction is "More fertilizer, less water, take it slow," the attention layer assigns higher weights to "more" and "less," as they directly determine the device's execution parameters, while "take it slow" has a relatively lower weight, serving only a modifying function. After weighting, the system obtains a dependency weight distribution: "more" weight 0.38, "less" weight 0.35, and "take it slow" weight 0.12, thus clearly indicating that the control parameter nodes in the syntactic tree are located at a higher level.

[0206] Preferably, refining dependency arcs by combining them with a pre-defined phrase structure template library can significantly improve the accuracy of intermediate syntactic structures.

[0207] For example, when processing "quickly rinse the cornfield", the initial dependency arc may incorrectly connect "rinse" to "cornfield". However, after the "rinse → increase water volume" rule in the template library is triggered, the system automatically adjusts to the correct structure of "rinse → increase water volume" as the dominant element and "cornfield" as the application object, ensuring that the logic conforms to the sprinkler irrigation logic of the water and fertilizer integrated machine.

[0208] It should be noted that when there is a discrepancy between the intermediate syntactic structure and the dialect annotation, the hierarchical relationship can be effectively corrected by adjusting the rules.

[0209] For example, the phrase "the land is too dry, water it thoroughly first" was initially interpreted as "water thoroughly" governing "first". However, according to the context rules, "water thoroughly first" is the complete intention. After adjustment, the dependency arc becomes "first → water thoroughly → the land is too dry", and the hierarchical relationship is more in line with the actual emergency irrigation needs of farmland.

[0210] In one embodiment, the modified syntactic tree structure is structurally saved using a data storage tool, forming a syntactic parsing database containing node, arc relationships, and weight information. The saved content may include "Node: Thoroughly watered, Arc: First → Thoroughly watered, Weight: 0.41, Context: Drought-stricken water replenishment," facilitating rapid subsequent retrieval. It is understood that batch mapping of this syntactic parsing database with actual operational needs can generate a final control command sequence. For example, for three consecutive commands "Fertilizer applied, Water not too much, Apply evenly," the system batch-matches and outputs a structured sequence: Fertilizer application rate 120%, Water flow limit 80%, Even mode enabled. This mapping method significantly improves the execution matching accuracy of commands on the integrated water and fertilizer machine, ensuring the complete transmission of operational intentions.

[0211] Step S105: Based on the root node and dependency path information in the syntactic tree structure representation, and combined with the dependency path constraint rules of the water and fertilizer integrated machine operation parameter slots built into the template library, extract the core control predicate and related parameter slot filling content. At the same time, through the context-based predicate completion prediction algorithm and the multi-turn dialogue context dependency relationship completion strategy, determine the set of structured operation instruction triples.

[0212] By using the root node information in the syntax tree structure and combining it with dependency path rules, we can obtain the preliminary basis for dividing the operation parameter slots and determine the priority order of controlling predicate extraction.

[0213] Based on the priority order of the control predicate extraction, the preset rules in the template library constraints are used to fill the operation parameter slots with content, thus obtaining the preliminary parameter slot content.

[0214] Based on the initial parameter slot content and relevant contextual completion information, a predicate completion strategy is used to supplement the missing control predicates. It is then determined whether the supplemented parameter slot content conforms to the structural requirements of the operation instruction set. If the supplemented control predicates are inconsistent with the dependency relationship completion information in the multi-turn dialogue, correction is performed using a preset template library constraint to obtain the adjusted parameter slot content.

[0215] Based on the adjusted or supplemented parameter slot content, combined with dependency path rules and root node information, a complete set of operation instructions is constructed to determine its logical integrity in the water and fertilizer integrated machine control scenario.

[0216] The syntax tree root node information, dependency path information, and weight information in the operation instruction set are structured and associated with the corresponding parameter slot content to obtain instruction mapping data.

[0217] By combining contextual completion and multi-turn dialogue information, semantic integration and optimization of instruction mapping data are performed to generate the final set of operation instruction triples, and their applicability in actual agricultural scenarios is determined.

[0218] For example, in agricultural scenarios controlled by integrated water and fertilizer machines, dividing the operation parameter slots using root node information and dependency path rules in a syntactic tree structure provides a clear logical framework for subsequent instruction generation. The root node typically represents the core intent of the instruction, such as "watering" or "fertilizing," while dependency path rules help identify the hierarchical relationships between parameters. Suppose an instruction is "water more in the western field," the root node is identified as "water," the dependency path shows "more" modifying "water," and "western field" as the target location, thus initially dividing the parameter slots: action "water," degree "more," and target "western field." This division lays the foundation for subsequent parameter extraction.

[0219] For example, when filling slots based on the priority order of control predicate extraction and predefined rules in the template library, core action words can be extracted first. Assuming the instruction is "Let go of the fat, hurry up and do it," the system first extracts "let go" as the control predicate using the priority rules. Combining this with the "let go-quantity-degree" structure in the template library, the slot is filled as "Action: Let go, Object: Fat, Degree: Less," thus initially forming the parameter content. This method ensures that the core intent is not overlooked.

[0220] For example, in context completion and predicate completion strategies, if the instruction is "Don't forget the water," the initial slot may lack a specific action. The system analyzes the context of "irrigation" mentioned in the multi-turn dialogue and completes the predicate to "irrigate," forming the complete slot "Action: Irrigate, Object: Water." If the completed slot does not match the dependencies in the multi-turn dialogue, such as if the user later mentions "just check the water," it is corrected using a template library to "Action: Check, Object: Water," ensuring logical consistency.

[0221] For example, when constructing a complete set of operation instructions, the system integrates the adjusted slot content into structured instructions by combining the root node and dependency path rules. For the instruction "More fertilizer, less water," the final set becomes "Fertilizer application: More, Water application: Less," which logically meets the equipment control requirements. The structured storage tool then associates and saves key information from the syntax tree with the slot, such as "Node: Fertilizer, Path: More, Weight: 0.6," facilitating subsequent retrieval.

[0222] For example, when generating the final set of operation instruction triplets, the system combines context and multi-turn dialogue information to generate triplets like "fertilize first, then water" for the "fertilize first, then water" instruction, ensuring a reasonable execution order in actual agricultural scenarios. This approach improves the applicability of the instructions.

[0223] Step S106: A structured set of operation instruction triples is adopted, combined with a fast intent completion module based on historical instruction sequence pattern matching priority sorting and user voice tone emotion tendency judgment logic, and mapped to the predefined control protocol format of the water and fertilizer integrated machine. It is then determined whether it meets the semantic consistency verification requirements. If the verification passes, the final executable device instruction sequence is generated.

[0224] Obtain the core operational intent of the current round of dialogue by analyzing the historical instruction sequence.

[0225] The historical instruction sequences are sorted according to pattern matching priority to determine the three most matching templates.

[0226] The pronunciation characteristics of the current user (the user corresponding to this control voice) are analyzed using voice intonation parameters to obtain the emotional tendency judgment result.

[0227] Based on the sentiment judgment results, the intent is quickly completed for the best matching template, generating a set of candidate operation instructions.

[0228] The candidate operation instruction set is mapped one by one to the predefined control protocol format of the water and fertilizer integrated machine to obtain the protocol instruction sequence.

[0229] Perform semantic consistency checks on the protocol instruction sequence. If the check passes, output the final device instruction sequence. If it fails, discard the instruction and return to the previous step to complete it again.

[0230] An instruction execution record is established based on the final device instruction sequence, and stored by associating it with the current round dialogue identifier.

[0231] Specifically, in agricultural scenarios controlled by integrated water and fertilizer machines, obtaining the core operational intent of the current round of dialogue through historical command sequences is a crucial step. For example, if a user mentions "more fertilizer," "less water," and "fertilize first this time" three times consecutively, the system extracts the historical core as an intent chain of "fertilization priority with a higher fertilizer amount and lower irrigation amount." Specifically, the system sorts historical command sequences by pattern matching priority, prioritizing templates from the most recent rounds, with the highest frequency and semantic similarity. Assuming three templates are matched: the first, "fertilize first, then irrigate," has a match score of 0.92; the second, "increase fertilizer amount, decrease irrigation" has a match score of 0.85; and the third, "adjust fertilizer amount only" has a match score of 0.67, then the first two are selected as the primary references.

[0232] Preferably, the user's current pronunciation characteristics are analyzed in conjunction with speech intonation parameters. If the user speaks quickly, with an upward pitch and emphasis on the word "hurry up," the emotional tendency is judged as "urgent and slightly dissatisfied." In this case, the matching template is quickly completed to prioritize generating a set of candidate instructions that favor faster execution and increased adjustment. For example, candidates include "immediately increase fertilization by 20%, decrease irrigation by 15%", "quickly fertilize first, then water, increase fertilizer and reduce irrigation", and "only increase fertilizer amount this time, keep everything else unchanged." In one possible implementation, the above candidate operation instructions are mapped one by one to the predefined control protocol format of the water and fertilizer integrated machine to obtain a protocol instruction sequence. For example, the first instruction is mapped as "valve 1 opening +20%, valve 2 opening -15%, execution order 1 priority", and the second instruction is "sequence control: fertilization → irrigation, increase fertilizer concentration, decrease irrigation flow". It should be noted that semantic consistency verification needs to be performed on the protocol instruction sequence. The verification mainly checks whether the execution order is consistent with the historical intent. Figure 1 The system checks whether the adjustment range exceeds the equipment's safety threshold and whether there are conflicting actions. For example, if the historical requirement is "water cannot be completely shut off," but a candidate shows "irrigation volume reduced to zero," then that check fails and is discarded. The system returns to the previous step to complete other candidates. For example, after successful verification, the final equipment instruction sequence might be "fertilizer valve opening increased to 125% of the original level, irrigation valve opening decreased to 80% of the original level, execution order: fertilize first, then irrigate." An instruction execution record is created based on this sequence and stored in conjunction with the current round's dialogue identifier for easy subsequent traceability and conflict detection. Understandably, this processing method, based on historical sequences, sentiment indicators, and rigorous verification, significantly improves instruction continuity, user satisfaction, and equipment safety, avoiding operational deviations caused by single-round misunderstandings, making it particularly practical in actual greenhouse precision management.

[0233] Step S107 involves sending the final executable device instruction sequence to the agricultural irrigation execution module, while simultaneously using the dynamic update mechanism of the dependency syntax template library to record the new dialect syntax pattern and the feedback data from the agricultural scene background noise filtering module, thereby determining the real-time adaptive adjustment requirements during instruction execution.

[0234] The device transmits instructions to the agricultural irrigation control unit to obtain the initial status information of the issued instructions and determine whether the instructions have been received correctly.

[0235] Based on the initial state information, the response data of the agricultural irrigation control unit is compared using pre-established command parsing rules to determine the initial completeness of command execution.

[0236] If the initial integrity of instruction execution does not reach the preset threshold, environmental interference data is obtained by combining feedback information from the noise filtering module to determine the scope of interference's impact on instruction execution.

[0237] By leveraging the real-time update function of the syntactic template, real-time environmental interference data is obtained from the background noise filtering module in the agricultural scenario, and newly entered dialect expression data is acquired. It is then determined whether the instruction expression method needs to be adjusted to adapt to the scenario requirements. If the instruction expression method needs to be adjusted, a new instruction format suitable for dialect expression is generated through syntactic structure analysis, resulting in optimized instruction content.

[0238] Based on the optimized instructions, the data is transmitted to the agricultural irrigation control unit for re-execution, and updated feedback information is obtained to determine whether the dynamic adjustment meets the requirements of the scenario.

[0239] The updated feedback information records the final state of the command execution and stores it in the system log database, completing the closed-loop processing of the agricultural irrigation task.

[0240] For example, in the command transmission and execution process of an agricultural irrigation control unit, obtaining initial state information is a crucial step to ensure correct command reception. Suppose that in a greenhouse irrigation task, the system issues a command to "open the main water valve to 50% flow." The device returns initial state information indicating successful reception, but the flow rate is only 30%. Using pre-established command parsing rules, the system compares the data and finds that the flow rate deviation exceeds a preset threshold of 10%, indicating that the initial integrity is not met. At this point, combined with environmental interference data from the noise filtering module, strong electromagnetic interference is detected, potentially causing unstable signal transmission and affecting the main water valve control area. Specifically, to address the interference's impact, the system uses the real-time update function of the syntactic template to obtain newly entered dialect expression data. Suppose the user is accustomed to using the local expression "water halfway open," while the original command is the standard phrase "open the main water valve to 50% flow." The system determines that the command expression needs to be adjusted to adapt to the local scenario. Through syntactic structure analysis, a new command format adapted to the dialect is generated, such as "water halfway open, execute immediately," and transmitted to the control unit for re-execution. The updated feedback shows that the traffic has been adjusted to 48%, close to the expected value, and the dynamic adjustment meets the needs of the scenario. For example, in the final status recording stage of command execution, the system stores information such as the traffic adjustment to 48% and the execution time in the system log database, forming a closed-loop process. Such recording facilitates subsequent traceability, and if similar interference issues are found, historical data can be quickly retrieved to optimize the command format. This approach effectively improves the stability of command execution, especially in complex environments, reducing misoperations caused by inappropriate expression. Specifically, for the implementation method of each technical topic, the noise filtering module can automatically trigger the command retransmission mechanism by collecting environmental signal strength, such as when the electromagnetic interference value reaches a certain critical level. The syntax template update relies on the accumulation of users' daily conversations. For example, the system recognizes that "water less" often means "reduce traffic by 20%" locally, and then includes it in the template library. This accumulation and adaptation makes the commands closer to user habits and reduces communication errors. The final status recording uses timestamps and task identifiers to ensure that each command is traceable, providing data support for subsequent optimization. Such a closed-loop process can significantly improve the accuracy and reliability of task execution in agricultural irrigation.

[0241] This invention presents a robust voice control method for integrated water and fertilizer machines designed for agricultural dialects. It addresses the unique challenges of traditional speech recognition systems, such as low accuracy, difficulty in capturing dialect rhythm features, and command parsing failures due to non-standard syntactic structures, when farmers issue water and fertilizer operation commands in local dialects under complex noise conditions in farmland. This invention extracts a preliminary general speech representation using a hybrid feature model combining Mel-frequency cepstral coefficients and log-energy spectral density. Then, it utilizes a multi-layer transform residual network to fuse semantic weights related to farmland operation contexts to generate a speech vector sequence specific to agricultural scenarios. Furthermore, it employs a generative adversarial network to enhance the adaptability of low-quality speech and combines a dedicated vocabulary for water and fertilizer operations in farmland with dimensionality reduction strategies to segment dialect commands. Subsequently, it constructs a syntax tree based on a dependency syntax template library and an attention mechanism, extracts core control predicates and parameter slots, and forms structured operation command triples through multi-round contextual completion. Finally, it maps these triples to the device protocol and dynamically updates the template library for continuous adaptation. This method effectively solves the problem of accurate recognition and control of agricultural dialect speech under strong noise and non-standard syntax, significantly improving the robustness, accuracy, and practicality of voice interaction in integrated water and fertilizer machines.

[0242] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A robust voice control method for integrated water and fertilizer machines based on agricultural dialects, characterized in that, include: Step 1: Collect the voice signal controlling the integrated water and fertilizer machine, filter the voice signal using the agricultural scene background noise filtering module, perform hybrid feature calculation on the filtered voice signal to obtain the initial acoustic feature vector, and input the initial acoustic feature vector into the pre-established embedding model for feature space mapping to obtain preliminary general voice features. Step 2: The preliminary general speech features are enhanced by using multi-layer transformation mapping combined with residual connection network. At the same time, the farmland operation context is embedded through semantic weight adjustment mechanism to generate a speech vector sequence for agricultural scenarios containing time step markers. Step 3: Based on the agricultural scenario-specific speech vector sequence, combine the pre-constructed vocabulary of water, fertilizer, and farmland terms with the vector dimension compression strategy to perform word segmentation processing, and enhance the adaptability to low-quality speech through adversarial generative networks; obtain the preliminary word unit sequence, perform hierarchical parsing, dialect operation feature classification and batch optimization on the preliminary word unit sequence, and then perform deep semantic parsing to obtain the dialect farmland operation instruction word segmentation unit sequence; Step 4: The sequence of word segments for the dialect farmland operation instructions is parsed using a dependency parsing model. Combined with a dependency parsing template library that stores the rule set of phrase structures for the control context of the integrated water and fertilizer machine, the dependency relationship of the word segments in the agricultural dialect context is calculated using an attention mechanism to obtain a syntactic tree structure with dependency arcs. Step 5: Based on the root node and dependency path information in the syntax tree structure, and combined with the dependency path constraint rules in the dependency syntax template library, extract the predicates of the control speech and the related water and fertilizer integrated machine slot filling data. Then, through the predicate completion prediction algorithm, combined with the context of the control speech and the preset rules of the dependency syntax template library, obtain the set of structured operation instruction triples. Step Six: Use the fast intent completion module to map the structured set of operation instruction triples to the predefined control protocol format of the water and fertilizer integrated machine to obtain the protocol instruction sequence; perform semantic consistency verification on the protocol instruction sequence. If the verification passes, obtain the final executable device instruction sequence and execute Step Seven; if it fails, discard the protocol instructions that fail the verification and return to Step Five. Step 7: Send the final executable device instruction sequence to the agricultural irrigation execution module of the water and fertilizer integrated machine, use the agricultural scene background noise filtering processing module to obtain real-time environmental interference data, obtain real-time adaptive adjustment data during instruction execution, and realize voice control of the water and fertilizer integrated machine.

2. The robust voice control method for integrated water and fertilizer machines based on agricultural dialects according to claim 1, characterized in that, In step one, the specific method for obtaining preliminary general speech features is as follows: Step 11: The collected control voice signal is filtered using an agricultural scene background noise filtering module to remove non-voice segments and background interference, resulting in a clean voice data stream. Steps 1 and 2: Extract features from the clean speech data stream and calculate the Mel frequency cepstral coefficients and log energy spectrum of the extracted features to obtain a mixed feature set; Step 13: Using the aforementioned hybrid feature set, construct an initial acoustic feature vector to obtain structured features; Step 14: Input the structured features into a pre-established embedding model, which is used to map the structured features into intermediate feature vectors adapted to agricultural noise environments; Step 15: Using the intermediate feature vector, a feature space mapping is used to obtain general speech features. When the discriminative power of the general speech features in different agricultural noise environments is lower than a preset threshold, the dimensions of the general speech features are adjusted until the discriminative power of the general speech features in different agricultural noise environments reaches the preset threshold, thus obtaining optimized features. The optimized features are then used as the final general speech features.

3. The robust voice control method for integrated water and fertilizer machines based on agricultural dialects according to claim 1 or 2, characterized in that, In step two, the method for generating a speech vector sequence specifically for agricultural scenarios, including time step markers, is as follows: Step 2: First, a network-based processing framework is constructed using a combination of multi-layer transformation and mapping. This framework is then used to enhance the detail information of the initial general speech features to obtain the initial speech signal. Step 22: Optimize the initial speech signal using a residual connection mechanism, retaining the detailed information in the original signal to obtain the optimized speech signal; Steps 2 and 3: Combine the farmland operation context to calculate the semantic attention weight of the optimized speech signal, and use the semantic attention weight to perform weighted adjustment on the optimized speech signal to obtain a set of speech signals with contextual information. Step 24: Perform time-step segmentation and labeling on the speech signal set with contextual information to generate a structured intermediate vector sequence; Step 25: Perform detail feature enhancement processing on the structured intermediate vector sequence to highlight the speech parts related to farmland operations, and obtain the enhanced speech signal sequence; Step 26: Perform temporal correlation coding and feature aggregation on the enhanced speech signal sequence to generate the final speech vector sequence for agricultural scenarios, which includes time step markers.

4. The robust voice control method for integrated water and fertilizer machines based on agricultural dialects according to claim 3, characterized in that, In step three, the method for obtaining the sequence of segmented units for dialect farmland operation instructions is as follows: Step 3:

1. Use a vocabulary list of terms for water and fertilizer farmland, and combine principal component analysis or sparse coding to perform dimensionality compression and initial word segmentation on the speech vector sequence for the agricultural scenario.

2. Use an adversarial generative network to adaptively enhance low-quality speech and obtain a preliminary sequence of dialect farmland operation instruction word units. Step 3.2: Using a hierarchical parsing method, match the instruction units in the preliminary dialect farmland operation instruction unit sequence with the water and fertilizer farmland terminology vocabulary to obtain the semantic unit set corresponding to the water and fertilizer farmland terminology vocabulary. Step 3: Based on the dialect operation features in the speech vector sequence, classify the semantic unit set using a preset classification rule to obtain the classified instruction unit set; Steps 3 and 4: Combining the dialect operation requirements in agricultural scenarios, a batch processing method is used to associate each unit in the instruction unit set with the low-quality speech adaptation capability enhanced by adversarial generative network, so as to obtain an optimized instruction unit combination. Step 35: Perform deep analysis on the optimized instruction unit combination, extract potential semantic information, and obtain the dialect farmland operation instruction word segmentation unit sequence.

5. The robust voice control method for integrated water and fertilizer machines based on agricultural dialects according to claim 1, characterized in that, In step four, the method for obtaining the syntax tree structure with dependency arcs is as follows: Step 41: Combining the phrase structure rules under the pre-constructed control context of the integrated water and fertilizer machine, the dependency parsing model is used to perform preliminary analysis on the word segmentation unit sequence of dialect farmland operation instructions to obtain preliminary dependency association information between each word unit; Step 4.2: Based on the preliminary dependency association information, apply an attention mechanism to weight the importance of the syntactic role of each word unit in the word unit sequence in the agricultural dialect context, obtain the dependency weight distribution of all word units, and determine the hierarchical relationship of all word units in the syntactic tree structure based on the dependency weight distribution. Step 4.3: Use the dependency syntax template library of the control context phrase structure rule set of the integrated water and fertilizer machine to refine the hierarchical relationship in the syntax tree structure and generate a syntax tree structure with dependency arcs.

6. The robust voice control method for integrated water and fertilizer machines based on agricultural dialects according to claim 1, characterized in that, Step five involves obtaining the structured set of operation instruction triples: Step 51: Based on the root node information in the syntax tree structure and the parameter slot dependency path constraint rules in the dependency syntax template library, obtain the preliminary division basis of the operation parameter slots and determine the priority order of control predicate extraction. Step 52: According to the priority order of the control predicate extraction, the operation parameter slots are filled with content using the preset rules in the dependency syntax template library to obtain the preliminary parameter slot content. Step 53: Based on the statistical correlation between the control predicates and the context in the instruction sequence during multi-round historical voice control, a predicate completion prediction algorithm is used to fill in the missing control predicates in the preliminary parameter slot content to obtain the completed parameter slot content. Step 54: Based on the completed parameter slot content, combined with the dependency path rules and root node information in the syntax tree structure, construct a complete set of operation instructions; Step 55: Structure and store the root node information, dependency path information, and weight information of the syntax tree in the set of operation instructions with the corresponding parameter slot content to obtain instruction mapping data; Steps 5 and 6: Use the preceding instruction information from the multi-turn historical dialogue of voice control to perform semantic integration and optimization on the instruction mapping data, and generate the final set of operation instruction triples.

7. The robust voice control method for integrated water and fertilizer machines based on agricultural dialects according to claim 1, characterized in that, In step six, the method for obtaining the final executable device instruction sequence is as follows: Step 61: Extract the instruction triplet set from the historical rounds of dialogue, and combine it with the operation instruction triplet set from the current round to extract the main operation intent of the current round of dialogue; Step 62: Match the current round's set of operation instruction triples with the instruction patterns in the historical instruction triples set, and correct the matching results according to the main body of the current round's dialogue operation intent to determine candidate intent templates that meet the preset matching requirements; Step 63: Analyze the pronunciation features of the current controlled speech using speech intonation parameters to obtain the emotional tendency judgment result in the current speech information; Step 64: Based on the sentiment tendency judgment result, perform intent completion on the candidate intent template to generate a set of candidate operation instructions; Step 65: Map each candidate instruction in the candidate operation instruction set to the predefined control protocol format of the water and fertilizer integrated machine to obtain the corresponding protocol instruction sequence; Step 66: Perform semantic consistency verification on the protocol instruction sequence to determine whether it meets the semantic constraints of the integrated water and fertilizer machine control scenario. If the verification passes, output the current protocol instruction sequence as the final executable device instruction sequence. If the verification fails, discard the current candidate instruction and return to Step 64 to perform intent completion again until an instruction sequence that meets the verification requirements is generated or all candidate instructions are traversed.

8. The robust voice control method for integrated water and fertilizer machines based on agricultural dialects according to claim 1, characterized in that, Step seven, the process of implementing voice control of the integrated water and fertilizer machine, is as follows: Step 71: Send the final executable device instruction sequence to the agricultural irrigation control unit, and obtain the initial status information of the instruction issuance from the agricultural irrigation control unit; Step 72: Based on the initial state information, analyze the response data after the agricultural irrigation control unit sends the command using the pre-established command parsing rules, and determine whether the initial integrity of the command execution has reached a preset threshold. If so, proceed to step seven or six; Otherwise, proceed to step seven-three; The instruction parsing rules include: comparing the status code returned by the device with the preset successful execution status code, and calculating the deviation between the actual execution parameters and the instruction issuance parameters; Step 73: Obtain real-time environmental interference data from the agricultural scene background noise filtering module, analyze the impact of the current noise environment on command execution, and determine the scope of interference. Step 74: Based on the noise type and intensity corresponding to the affected area, retrieve historical dialect expression data associated with the noise environment from the dependency syntax template library, and obtain the dialect instruction patterns successfully executed in the noise scenario; Step 75: Calculate the semantic similarity between the initial state information of the issued command and the dialect command pattern successfully executed in the noise scenario. If the semantic similarity is lower than a preset threshold, it is determined that the command expression method needs to be adjusted; otherwise, it is determined that no adjustment is needed. If the command expression method needs to be adjusted, the dependency parsing model is used to perform syntactic structure analysis on the dialect command pattern successfully executed in the similar noise scenario to generate a new command format adapted to the dialect expression habits, and obtain the optimized command content. The optimized command content is retransmitted to the agricultural irrigation control unit for execution, and the updated command state information is obtained. The updated command state information replaces the initial state information of the issued command, and step 76 is executed. Step 76: Determine whether the initial state information of the issued instruction meets the expected requirements of the agricultural scenario. If it does, generate instruction execution success status information. If it does not, determine whether the current number of adjustments has reached the preset maximum adjustment threshold. If it has not reached the threshold, return to step 73 and readjust the instruction expression method based on the current noise environment. If it has reached the threshold, generate instruction execution failure status information.

9. The robust voice control method for integrated water and fertilizer machines based on agricultural dialects according to claim 1, characterized in that, Step seven also includes step seven seven, which records the final status information of the instruction execution to the system log database, and simultaneously updates the dialect expression data involved in this execution process and its corresponding noise environment features and syntactic patterns to the dependency syntax template library.

Images