Power grid dispatching voice recognition method and system based on multi-scale feature fusion

Abstract

The invention discloses a power grid dispatching voice recognition method and system based on multi-scale feature fusion, relates to the technical field of voice recognition, and is technically characterized in that the dispatching adaptability is guaranteed through a full-process technical scheme, power grid operation state parameters are synchronously acquired in a data acquisition link to provide scene context, and the power grid dispatching voice recognition efficiency is improved. In the feature fusion link, the weight is dynamically adjusted according to a scheduling scene, in the model identification link, illegal instructions such as non-electricity-testing, namely switching-on and the like are eliminated through a logic map, and in the result verification link, cross-dimension linkage verification of accuracy, logic compliance and scene consistency is implemented. Through quadruple logic verification and two-dimensional scene matching, it is ensured that an identification result is completely matched with a scheduling procedure and a power grid state; and meanwhile, adaptive adjustment of parameters is supported, switching of the operation state of the power grid can be quickly responded, complex and changeable scheduling scenes are perfectly adapted, and scheduling safety risks are avoided from the technical level.

Description

Technical Field

[0001] This invention relates to the field of speech recognition technology, and in particular to a speech recognition method and system for power grid dispatching based on multi-scale feature fusion. Background Technology

[0002] Power grid dispatching is a core component in ensuring the safe and stable operation of the power system. The accurate and efficient transmission of dispatching instructions directly determines the reliability of power grid operation and dispatching efficiency. As the power grid transforms towards intelligence and digitalization, speech recognition technology, with its advantages of freeing up hands and improving interaction efficiency, has been widely applied in power grid dispatching scenarios. The aim is to automate the input, parsing, and execution of dispatching instructions, reduce human error, and support the intelligent upgrade of dispatching operations. Currently, the mainstream speech recognition solutions used in power grid dispatching scenarios are mainly divided into two categories: one is a dedicated solution based on the combination of traditional speech features and hidden Markov models, and the other is a solution that directly transfers and adapts general speech recognition models. Both types of solutions have been piloted in simple dispatching scenarios.

[0003] However, power grid dispatching scenarios present significant unique challenges, including strong electromagnetic interference, noise coupling from multiple devices, multi-scale voice information, and stringent dispatching logic. In practical implementation, the aforementioned existing technical solutions have gradually revealed numerous intractable flaws, failing to meet industrial-grade dispatching requirements: First, existing solutions generally employ a separate processing strategy of noise reduction followed by feature extraction, such as traditional noise reduction methods based on spectral subtraction and Wiener filtering. During pilot implementations, these solutions struggle to accurately distinguish noise from voice signals in the face of strong electromagnetic interference and multi-device noise superposition in the power grid dispatching environment. This easily leads to attenuation of voice feature amplitude and frequency shift, resulting in a significant drop in recognition accuracy, especially in areas with strong interference such as substations. Second, existing solutions often focus on single-scale voice feature extraction. Traditional solutions only extract short-term voice features, while general transfer models, although capable of extracting long-term features, lack specificity. When processing dispatching voice, they either fail to accurately capture closing and opening signals or... The existing speech recognition model suffers from several problems. First, it lacks robustness. Second, it is difficult to fully parse long instructions containing operational conditions and objects, leading to frequent omissions and semantic misinterpretations during pilot implementations. Third, the existing speech recognition model does not integrate the professional logic of power grid dispatching and relies solely on the speech signal itself for recognition. In actual dispatching operations, it often produces recognition results that conflict with dispatching procedures, such as misidentifying non-compliant instructions like closing the circuit breaker without verification of power or performing maintenance before power outages, failing to meet the rigid safety requirements of power grid dispatching. Fourth, the existing solution's feature extraction, model recognition, and result verification are independent of each other and lack a linkage optimization mechanism. During pilot implementations, when the power grid operating state changes or the noise type changes, the parameters of each link cannot be adaptively adjusted, resulting in insufficient robustness of the recognition results and poor coordination between system modules, making it difficult to adapt to complex and ever-changing power grid dispatching scenarios. Therefore, we propose a power grid dispatching speech recognition method and system based on multi-scale feature fusion. Summary of the Invention

[0004] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a power grid dispatching speech recognition method and system based on multi-scale feature fusion, thus solving the aforementioned technical problems.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: A power grid dispatch speech recognition method based on multi-scale feature fusion is described below: The system acquires multi-source voice data and related parameters of power grid operation in power grid dispatching scenarios. The multi-source voice data includes voice information related to dispatching instructions, equipment prompts, environmental background, and personnel interaction. The related parameters of power grid operation include power grid load, bus voltage, equipment operating status, and fault alarm information. Voice data is collected synchronously through voice acquisition equipment adapted to the power grid scenario. During the acquisition process, the time synchronization accuracy of multi-source data and the acquisition resource allocation strategy are dynamically adjusted according to the power grid operating status. In fault scenarios, priority is given to ensuring the acquisition priority and data quality of dispatching instruction voice. To address the complex noise characteristics of power grid dispatching scenarios, a preprocessing scheme that uses noise and speech co-modeling is adopted to process multi-source speech data in an integrated manner, extracting multi-scale speech features adapted to the dispatching scenario. These multi-scale speech features include short-term features for capturing dispatching keywords, medium-term features for capturing sentence segments, and long-term features for capturing the semantic association of complete instructions. During the preprocessing process, a speech distortion compensation mechanism is constructed to reduce the interference of noise on speech features and ensure the effective preservation of key speech information. Combining the representational characteristics of multi-scale speech features and the speech association rules of power grid dispatching terms, a scheme for extracting directional features and term associations is adopted to extract feature vectors corresponding to features at each scale. After extraction, invalid features are removed and re-extracted by verifying feature validity and semantic relevance. A multi-dimensional fusion scheme combining scene-adaptive weight adjustment, effective feature focusing, and semantic redundancy filtering is adopted to dynamically fuse feature vectors at different scales, thereby achieving complementary enhancement and redundancy suppression of features at different scales and obtaining fused feature vectors. The fused feature vector is input into a preset speech recognition model. The model integrates a power grid dispatching logic reasoning mechanism and combines power grid operation-related parameters to achieve dual recognition of scenario adaptation and logic verification, and outputs speech recognition results. The speech recognition results are subjected to cross-dimensional linkage verification of accuracy, scheduling logic compliance and scenario consistency. The verification standards are dynamically adjusted according to the verification results to determine whether the recognition results meet the preset requirements. If the identification result meets the standard, the result is output and stored, and it is linked with the power grid dispatching system to support closed-loop management of dispatching operations. If the identification result does not meet the standard, the adaptive adjustment mechanism of feature processing, model parameters and preprocessing strategy is triggered. The relevant parameters are optimized according to the type of non-compliance and then returned for reprocessing. If the standard is still not met after repeating the process a preset number of times, manual intervention is triggered.

[0006] Preferably, the preprocessing scheme employing noise and speech co-modeling to perform integrated processing on multi-source speech data and extract multi-scale speech features adapted to the scheduling scenario follows these steps: An integrated processing scheme for multimodal noise identification, separation, and distortion compensation is adopted to identify the type and intensity of noise in multi-source speech data, distinguishing between electromagnetic interference noise from power grid equipment, environmental background noise, and redundant voice noise from personnel. An adaptive noise suppression mechanism is constructed by combining a power grid equipment noise feature library. The power grid equipment noise feature library contains the noise characteristics of various power grid equipment under different operating conditions and their correlation with speech distortion information, and supports dynamic updates. Based on the noise type and intensity level, a graded and adaptive noise reduction strategy is adopted to achieve accurate noise separation and suppression; at the same time, a speech distortion compensation mechanism is used to correct the speech feature distortion caused by noise, ensuring that the quality of the processed speech signal meets the requirements of subsequent recognition. Based on the priority of dispatch instructions and the power grid operation status, the noise-reduced voice data is subjected to adaptive frame segmentation processing, and the frame length and frame shift parameters of different duration frames are dynamically configured to adapt to the feature extraction requirements of different types of dispatch instructions. Emergency instructions are given priority to configure short duration frames to accurately capture keywords, regular instructions are configured with medium duration frames to balance keywords and sentence integrity, and maintenance instructions are configured with long duration frames to fully capture semantic information. Feature extraction was performed on speech data of different duration frames to obtain short-time, medium-time, and long-time speech features. The three types of features were normalized to eliminate differences in dimensions and amplitude deviations. After feature validity verification, standardized multi-scale speech features were obtained.

[0007] Preferably, the extraction scheme using directional features and term association, and the steps for extracting feature vectors corresponding to features at each scale are as follows: For short-term speech features, an extraction strategy combining temporal analysis, keyword template matching, and term pronunciation association is adopted to extract instantaneous speech features, which are then double-filtered to form a short-term feature vector. The keyword template contains standard speech feature information of core power grid dispatching operation terms, equipment names, and status description terms. For mid-time speech features, an extraction strategy combining frequency domain analysis, feature difference enhancement, and term and word combination association is adopted to extract speech spectrum features and construct mid-time feature vectors after enhancing the effective feature discrimination. For long-term speech features, an extraction strategy combining time-frequency joint analysis, semantic association filtering, and instruction structure matching is adopted to extract speech time-frequency features, which are then filtered by scheduling instruction semantic structure rules to form a long-term feature vector. The three types of feature vectors are subjected to dimension normalization and sparsification to remove invalid and redundant feature components, ensuring that the dimensions of each feature vector are uniform and meet the requirements of subsequent fusion. After feature discrimination and semantic relevance verification, the qualified feature vectors are obtained.

[0008] Preferably, a multi-dimensional fusion scheme combining scene-adaptive weight adjustment, effective feature focusing, and semantic redundancy filtering is adopted to dynamically fuse feature vectors at various scales. The steps to obtain the fused feature vector are as follows: Construct a feature weight evaluation model that combines scheduling scenario type, power grid operation status and command priority, and calculate the initial weight of feature vectors at each scale; configure an appropriate weight allocation strategy for different scheduling scenarios, give priority to ensuring the weight ratio of core features in the current scenario, and support manual fine-tuning and self-learning optimization of weight parameters; Based on the initial weights, an initial fused feature vector is obtained by combining weighted fusion and feature focusing mechanisms. By dynamically adjusting the focusing range, the feature regions with high correlation to scheduling instructions are strengthened, and residual noise feature interference is suppressed. Semantic constraint rules from the power grid dispatching domain are introduced. Based on semantic relevance, feature validity, and instruction structure integrity, the initial fused feature vector is screened to remove redundant, invalid, and noise-residual features, thereby reducing the computational complexity of the model. The filtered feature vectors are subjected to cross-scale feature association enhancement and semantic calibration to strengthen the correlation of effective features at different scales, ensure consistency between feature representation and scheduling logic, and obtain the final fused feature vector.

[0009] Preferably, the steps for inputting the fused feature vector into the preset speech recognition model, combining it with power grid operation-related parameters to achieve dual recognition for scenario adaptation and logical verification, and outputting the speech recognition result are as follows: The preset speech recognition model is an adaptive recognition model driven by both scene and logic. It adopts a hybrid network structure that includes feature mapping, temporal enhancement and logical reasoning. It is trained and optimized with sufficient power grid scheduling speech samples, which cover different accents, noise scenes and fault scenes. The logical reasoning mechanism in the model integrates a power grid dispatching logical rule system. This system is built based on power grid dispatching-related procedures and historical operation cases, and includes dispatching instruction terminology specifications, operation logic rules, equipment associations, and emergency response rules, and supports dynamic updates. The fused feature vectors are input into the model, and the nonlinear representation capability of the features is improved through feature mapping. The temporal enhancement process strengthens the temporal correlation of the features. Then, the semantic calibration, logical reasoning and conflict elimination are performed through the logical reasoning mechanism to ensure that the recognition results conform to the scheduling logic. Through model classification and matching mechanisms, combined with historical scheduling instruction data for auxiliary matching, the results that meet the confidence and logical fit standards are selected as the final speech recognition results; the recognition results include key information such as scheduling instruction content, speech source, generation time, logical compliance identifier, scene matching degree and instruction type.

[0010] Preferably, the steps for cross-dimensional joint verification of speech recognition results in terms of accuracy, scheduling logic compliance, and scenario consistency are as follows: Extract key information from the speech recognition results, such as scheduling instruction keywords, semantic logic, priority identifiers, scene matching degree, and confidence level. Accuracy verification: Key information is compared with information in the power grid dispatch standard voice database in multiple dimensions, the information matching degree is calculated to determine the recognition accuracy, the accuracy threshold is preset, and results that do not meet the confidence or logical fit are judged to need to be reviewed. Dispatch logic compliance verification: Based on the power grid dispatch logic rule system, a multi-verification mechanism is adopted, including keyword combination, operation sequence, equipment association and logic conflict, to verify whether the identification results comply with the power grid dispatch safety specifications and operation logic, and to identify non-compliant scenarios and corresponding handling solutions. Scenario consistency verification: Construct a matching model between operating status and instruction type, and perform multi-dimensional matching between the instruction information in the recognition results and the power grid operation-related parameters to determine whether the instruction content, type and priority are compatible with the current power grid operating status; Cross-dimensional linkage verification: Associate the results of three verification items, dynamically adjust the verification threshold and review strategy, and comprehensively determine whether the recognition result is qualified; record relevant information and generate an optimization report for unqualified results, and trigger the manual verification process for results that need to be reviewed.

[0011] A power grid dispatching speech recognition system based on multi-scale feature fusion, comprising: The multi-source data acquisition module is used to synchronously acquire multi-source voice data and power grid operation-related parameters in power grid dispatching scenarios. It integrates voice acquisition equipment and data interfaces adapted to power grid scenarios, has dynamic priority allocation and time synchronization functions for multi-source data, prioritizes the quality of voice acquisition of dispatching instructions in fault scenarios, and has built-in equipment status monitoring and fault self-diagnosis mechanisms. The scene-adaptive preprocessing module is used to perform integrated processing of multi-source speech data based on the noise characteristics of power grid dispatching scenarios. It adopts a preprocessing scheme of noise and speech co-modeling and a frame segmentation strategy of instruction priority adaptation to extract standardized multi-scale speech features. It has a built-in power grid equipment noise feature library, distortion compensation model and preprocessing parameter optimization mechanism, and has processing quality assessment and self-optimization functions. The directional feature extraction module is used to extract feature vectors corresponding to multi-scale speech features by adopting an extraction scheme that associates directional features with terms, combined with scheduling term-related rules and templates; it has a built-in feature selection rule system and has the functions of feature validity and semantic relevance verification and re-extraction. The multi-dimensional feature fusion module is used to dynamically fuse feature vectors at various scales using a multi-dimensional fusion scheme that combines scene-adaptive weight adjustment, effective feature focusing, and semantic redundancy filtering, and outputs a fused feature vector. It has a built-in feature weight evaluation model, semantic constraint rules, and fusion quality evaluation mechanism, and supports weight adjustment and self-learning optimization. The dual-drive recognition module is used to input the fused feature vector into the preset speech recognition model, and combine it with power grid operation-related parameters to achieve dual recognition of scenario adaptation and logic verification, and output the speech recognition result; it has a built-in power grid dispatching logic rule system to support model status monitoring, online optimization and parameter updates. The cross-dimensional verification module is used to perform cross-dimensional linkage verification of speech recognition results in terms of accuracy, scheduling logic compliance, and scenario consistency. It dynamically adjusts the verification standards and review strategies to determine whether the recognition results are qualified. It has a built-in standard speech database, verification rule system, and manual verification interface, and has the functions of generating verification reports and recording non-compliance items. The result processing and linkage module is used to output and store qualified identification results and link them with the power grid dispatch system; for unqualified results, it triggers an adaptive parameter adjustment mechanism, instructs relevant modules to reprocess, records processing logs and optimization reports, and supports model iteration and manual intervention processes.

[0012] Preferably, the scene adaptive preprocessing module includes: The multimodal noise processing unit is used to identify noise type and intensity level, construct an adaptive noise suppression model, and achieve accurate noise processing and speech distortion correction by adopting an integrated processing scheme of noise identification, separation and distortion compensation; it has the functions of model self-learning update, processing effect evaluation and parameter adaptive adjustment. The priority framing unit is used to perform adaptive framing processing on the noise-reduced voice data based on the priority of scheduling instructions and the power grid operation status. It dynamically configures framing parameters, supports manual configuration and automatic adaptation of parameters, and has the functions of framing parameter verification and scene adaptability detection. The multi-scale feature extraction unit includes short-time, medium-time and long-time feature extraction sub-units, which extract features from speech data of different duration frames to obtain corresponding multi-scale speech features. The feature normalization unit is used to normalize multi-scale speech features, eliminate dimensional differences and amplitude deviations, and output standardized multi-scale speech features after feature validity and semantic relevance verification; it also triggers a re-extraction process for abnormal features.

[0013] Preferably, the directional feature extraction module is as follows: The short-time vector extraction unit is used to extract feature vectors corresponding to short-time speech features by combining time-domain analysis, keyword template matching and term pronunciation association. It has the functions of feature matching degree and pronunciation association degree evaluation and screening threshold adjustment. The mid-time vector extraction unit is used to extract feature vectors corresponding to mid-time speech features by employing a strategy that combines frequency domain analysis, feature difference enhancement, and term-lexical combination association. The long-term vector extraction unit is used to extract feature vectors corresponding to long-term speech features by employing a strategy that combines time-frequency joint analysis, semantic association filtering, and instruction structure matching. The feature normalization unit is used to normalize and sparsify the three types of feature vectors, remove invalid and redundant features, and ensure that the vector dimensions are consistent. It has the functions of dimension verification, feature discrimination and semantic relevance detection, and triggers the re-extraction process for features that do not meet the standards.

[0014] Preferably, the multi-dimensional feature fusion module includes: The weight calculation unit is used to construct a feature weight evaluation model. It calculates the initial weights of feature vectors at each scale by combining scheduling scenarios, power grid operation status, and command priorities. It supports dynamic weight updates, manual fine-tuning, and self-learning optimization functions. The feature focusing fusion unit is used to obtain an initial fused feature vector based on the initial weights by combining weighted fusion and feature focusing mechanisms; it dynamically adjusts the focusing range to enhance the effective feature region and suppress noise interference. The semantic filtering unit is used to filter the initial fused feature vector based on the semantic constraint rules of power grid dispatch, remove redundant, invalid and noise residual features, and reduce computational complexity. The feature enhancement unit is used to perform cross-scale association enhancement and semantic calibration on the selected feature vectors to improve the stability of feature representation and scene adaptability; it has a fusion feature quality assessment function, and features that do not meet the standards trigger a weight readjustment process.

[0015] (III) Beneficial Effects 1. By constructing a noise feature library for power grid equipment, accurately identifying and classifying three types of noise (including electromagnetic interference), and combining differentiated noise reduction strategies with a speech distortion compensation mechanism, a combined notch filtering and spectrum restoration process is used for high-level electromagnetic interference during implementation. Inverse corrections are applied to amplitude attenuation and frequency shift features, effectively solving the problems of inaccurate noise-speech differentiation and feature distortion in existing solutions. This ensures high retention rates of key speech features and a good signal-to-noise ratio. Even in areas with strong interference in substations, the recognition error rate is well controlled, significantly improving recognition stability in noisy environments. Secondly, short-term, medium-term, and long-term feature vectors are extracted specifically, and invalid redundant features are eliminated through L1-L2 hybrid regularization. In the fusion stage, weights are dynamically allocated based on the scenario, and attention mechanisms are used to focus on key features, while semantic rules are used to filter effective information, achieving multi-scale feature complementarity. Simultaneously, a scheduling logic inference layer is embedded in the model recognition stage, and training and optimization are performed using a dedicated dataset. This accurately captures keywords such as "closing" and fully parses long-sentence commands, improving recognition accuracy.

[0016] 2. The system ensures dispatch adaptability through a comprehensive technical solution. In the data acquisition stage, it synchronously acquires power grid operation status parameters to provide scenario context. In the feature fusion stage, it dynamically adjusts weights according to the dispatch scenario. In the model recognition stage, it uses logic graphs to eliminate illegal commands such as closing the circuit without verification of power. In the result verification stage, it implements cross-dimensional linkage verification of accuracy, logical compliance, and scenario consistency. Through four-fold logical verification and two-dimensional scenario matching, it ensures that the recognition results are fully compatible with the dispatch procedures and power grid status. At the same time, it supports adaptive parameter adjustment, which can quickly respond to the switching of power grid operation status and perfectly adapt to complex and ever-changing dispatch scenarios, thus avoiding dispatch security risks from a technical perspective.

[0017] 3. The multi-source data acquisition module has built-in equipment fault self-diagnosis and backup switching functions to ensure uninterrupted data acquisition; the preprocessing and feature extraction stages have abnormal feature feedback and reprocessing mechanisms to improve data quality; the result processing stage triggers precise parameter adjustments for unqualified results, and triggers manual intervention after a maximum of 3 retries to avoid misjudgments affecting dispatching; at the same time, it seamlessly links with the power grid dispatching system through the IEC61850 standardized interface, reducing linkage response delay and significantly improving dispatching efficiency. Attached Figure Description

[0018] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0019] Figure 1 This is an overall flowchart of an embodiment of the present invention; Figure 2 This is a flowchart of the speech preprocessing process in an embodiment of the present invention; Figure 3 This is a flowchart of multi-scale feature extraction and filtering in an embodiment of the present invention. Detailed Implementation

[0020] This application provides a power grid dispatching speech recognition method and system based on multi-scale feature fusion, which effectively solves the technical problems existing in the current power grid dispatching speech recognition scheme, realizes accurate recognition of power grid dispatching speech, improves the noise robustness, dispatching logic compliance and scenario adaptability of the recognition results, ensures the accuracy and efficiency of dispatching command transmission, and supports the automated operation of power grid dispatching.

[0021] Example: Reference Figures 1 to 3 As shown, the technical solution in this application embodiment effectively solves the technical problems existing in the current power grid dispatch voice recognition solution. The overall idea is as follows: To address the problems existing in the prior art, this invention provides a power grid dispatch speech recognition method based on multi-scale feature fusion. This recognition method comprises seven steps: multi-source data acquisition, scene-adaptive preprocessing, directional feature extraction, triple-linkage feature fusion, scene and logic dual-driven recognition, cross-dimensional linkage verification, result processing, and linkage. The details are as follows: The pre-construction preparation phase requires the deployment of a dispatch center microphone array and a substation industrial-grade voice acquisition terminal, which are connected to the SCADA system via an IEC 61850 standard interface. The microphone array must be electromagnetically interference resistant and have a pickup range of 0.5 to 5 meters. It employs MEMS microphone units with spatial beamforming technology to focus on the target sound source and suppress sidelobe noise, capturing dispatch voice information in the 100Hz to 20kHz core frequency band at a high sampling rate, suitable for dynamic pickup scenarios of dispatchers from 0.5 to 5 meters. Electromagnetic interference is mitigated through a metal shielding layer and grounding to avoid interference from the strong electromagnetic environment of the power grid, and it is used to collect dispatch command voices. The substation industrial-grade voice acquisition terminal has an IP65 protection rating, adaptable to operating conditions from -40℃ to 70℃ to suit the harsh environment of outdoor substations, and is used to collect on-site voice and equipment prompts. Wide-temperature processing is achieved through industrial-grade component selection and optimized heat dissipation structure, covering both the frigid conditions of the north and the high temperatures of the south. The terminal has a built-in anti-interference chip to filter out near-field electromagnetic interference from substation equipment, ensuring the quality of on-site voice acquisition. IEC As an international standard in the field of power grid automation, the 61850 standard interface adopts object-oriented modeling and abstract communication service interface, which can realize interoperability between equipment from different manufacturers. When connecting to SCADA system, it transmits power grid operation parameters through MMS protocol, ensuring the real-time and reliable transmission of data and avoiding parameter loss caused by protocol incompatibility.

[0022] During the preparatory phase, four major databases and feature libraries were constructed simultaneously: a dedicated dataset for power grid dispatching voice, a feature library for electromagnetic noise of power grid equipment, a standard voice database for power grid dispatching, and a causal logic graph for power grid dispatching operations, as detailed below: The dedicated dataset for power grid dispatch voice includes 200,000 real dispatch voice recordings, 100,000 synthesized scene voice recordings, and 50,000 fault scenario-specific voice recordings. Each data point is annotated with instruction content, priority, semantic logic, and power grid status for model training and matching. The electromagnetic noise feature library for power grid equipment covers noise data for 12 types of equipment, 3 operating conditions, and 5 intensity levels, including feature parameters such as spectral peak values ​​and supporting online updates for noise identification and suppression. The standard voice database for power grid dispatch contains 200,000 annotated standard voice recordings and feature parameters, updated quarterly for accuracy verification. The causal logic graph for power grid dispatch operations includes over 500 terms, over 300 rules, and over 200 sets of equipment relationships for logical reasoning and conflict resolution.

[0023] A CNN-LSTM hybrid model is employed to input speech time-domain and frequency-domain features, outputting noise type, intensity level, and distortion trend for accurate noise identification. The CNN extracts spatial features of the speech, including time-domain amplitude and frequency-domain spectral density, using three convolutional layers and two pooling layers to achieve feature dimensionality reduction and noise feature capture. The LSTM captures the temporal features of the speech, utilizing a gating mechanism to memorize the temporal correlations of continuous noise, avoiding misjudgments due to transient interference. The output provides a precise basis for subsequent noise reduction. A blind source separation algorithm (FastICA) is used to separate the target speech from noise, based on the principle of independent component analysis, assuming that the target speech and noise are statistically independent. The signal separation mechanism maximizes the non-Gaussianity of the signal to achieve source signal separation, effectively separating mixed noise superimposed on speech, especially suitable for substations with multiple noise superposition scenarios; the adaptive notch filter has a center frequency of 50Hz and 150Hz, with a notch depth of not less than 40dB, for targeted filtering of power grid interference; the core frequency of power grid interference is 50Hz, and 150Hz is the third harmonic. The notch filter adaptively adjusts the filtering parameters to deeply attenuate the 50Hz and 150Hz frequency components while preserving the speech signal. The notch depth of not less than 40dB ensures that the amplitude of the interference signal is reduced without attenuating the core frequency band of the speech.

[0024] The multi-source data acquisition process employs a dynamic timestamp alignment algorithm, based on the NTP protocol, to calibrate the time of each acquisition device and ensure data synchronization. The NTP protocol achieves time synchronization of each acquisition device with millisecond-level timestamp accuracy. The dynamic alignment algorithm adjusts the timestamp according to the data acquisition delay to ensure time consistency between voice data and power grid parameters. The synchronization accuracy in fault conditions does not exceed 20ms (e.g., 18ms) and in normal conditions does not exceed 50ms (e.g., 45ms). In fault conditions, a dynamic bandwidth allocation algorithm increases the core data transmission bandwidth by 50%, prioritizing the transmission of fault command voice and power grid fault parameters to avoid data congestion. The total acquisition delay does not exceed 80ms. Data transmission uses the TCP / IP protocol and sets a CRC-32 data checksum to ensure data transmission without loss or error.

[0025] The scene-adaptive preprocessing stage employs an integrated approach of joint noise and speech modeling, using a CNN-LSTM model to identify noise type and intensity. Differentiated noise reduction strategies are employed for different levels of interference. High-level interference (noise intensity not less than 60dB) severely impairs speech, requiring notch filtering with spectrum restoration. First, notch filtering removes power frequency interference, then spectrum restoration replenishes the missing spectrum. Medium-level interference (noise intensity between 40 and 60dB) is primarily power frequency interference, and a single notch filter suffices for noise reduction. Environmental noise is handled using adaptive threshold filtering with spectral subtraction. Adaptive threshold filtering dynamically adjusts the filtering threshold based on noise intensity. Spectral subtraction reduces noise by subtracting the noise spectrum, adapting to dynamic changes in environmental noise. Redundant speech employs a speech activity detection plus semantic filtering scheme. First, speech activity detection removes non-speech segments, and then semantic filtering removes speech unrelated to scheduling, such as idle chatter. Subsequently, a distortion compensation model corrects amplitude attenuation and frequency shift to ensure that speech distortion is below 5%. This distortion compensation is based on a speech distortion model, correcting amplitude attenuation through gain adjustment and frequency shift through phase calibration, ensuring that speech distortion is controlled within 5% and does not affect command recognition.

[0026] Targeted feature extraction involves extracting 256-dimensional short / medium / long-term feature vectors based on four template rule bases. First, multi-scale features are extracted through time domain, frequency domain, and time-frequency analysis. Then, regularization and sparsification are performed to ensure feature discriminativeness and semantic relevance, providing high-quality feature input for subsequent fusion stages.

[0027] The preprocessing stage dynamically frames instructions based on priority. Level 1 instructions (fault handling instructions) use a 12ms frame length and a 6ms frame shift to accurately capture core keywords such as fault and isolation. Fault instructions require rapid response; the short frame length ensures accurate keyword capture, and the 6ms frame shift ensures no keyword loss. Level 2 instructions (routine adjustment instructions) use a 60ms frame length and a 30ms frame shift to balance keyword recognition and sentence completeness. Routine instructions need to consider both keyword and sentence semantics, and this frame length and frame shift parameter adaptation is required. Level 3 instructions (maintenance instructions) use a 300ms frame length and a 150ms frame shift, focusing on complete semantic information. Maintenance instructions are longer, and the longer frame length allows for comprehensive semantic content capture. Frame parameter adjustments are adjusted within 30ms. Multi-scale features are extracted, including short-time, medium-time, and long-time features. Short-time features include instantaneous amplitude, instantaneous frequency, and instantaneous phase, with 16 feature values ​​per frame, focusing on key information. The system extracts detailed audio information, reflecting instantaneous changes in speech, to suit keyword recognition. Mid-time features include four categories: spectral peak, spectral bandwidth, spectral centroid, and spectral entropy. Extracted based on a 20ms Fourier transform, these features focus on speech segment information, reflecting the frequency distribution of speech segments to suit sentence segment recognition. Long-time features include three categories: time-frequency entropy, energy, and correlation. Extracted based on a 4-scale wavelet transform, these features focus on speech semantic information, reflecting the overall time-frequency characteristics of speech to suit complete semantic recognition. Finally, Min-Max adaptive normalization maps the features to the [0.05, 0.95] interval, eliminating abnormal features with a correlation coefficient below 0.7, outputting standardized parameters, with a preprocessing delay not exceeding 100ms. After preprocessing, a feature quality assessment step is added, evaluating the feature signal-to-noise ratio (SNR) and feature completeness. Features with an SNR of at least 25dB and completeness of at least 95% can proceed to the next stage; otherwise, they are returned for reprocessing.

[0028] The targeted feature extraction stage is based on four template and rule bases: a core keyword speech template base, a pronunciation association rule base, a terminology combination rule base, and a scheduling instruction structure template base. It extracts 256-dimensional short-term, medium-term, and long-term feature vectors respectively. Short-term vectors are constructed through temporal analysis, template matching, and pronunciation rules; medium-term vectors are constructed through frequency domain analysis, feature enhancement, and vocabulary rules; and long-term vectors are constructed through time-frequency analysis, semantic filtering, and structural matching. L1-L2 hybrid regularization is used to remove invalid and redundant features. Features with a variance below 0.005 that show little change and carry little information are considered invalid. Features with a cross-correlation coefficient above 0.9 are highly similar and contain information redundancy. Regularization ensures that the feature discrimination is not less than 0.6, outputting a 256-dimensional regularized vector. The extraction latency is no more than 70ms. A feature validity feedback mechanism is used: if the valid feature retention rate is less than 90%, the regularization parameters (L1 weight ±0.05, L2 weight ±0.05) are adjusted, and the feature is extracted again.

[0029] The triple-linkage feature fusion process constructs a feature importance assessment model driven by scenario, state, and priority. Subjective weights are determined using the analytic hierarchy process (AHP), and objective weights are determined using the entropy weight method, with a subjective weight ratio of 0.5:0.5. Subjective weights adapt to subjective needs such as scenario and priority, while objective weights adapt to the feature's inherent discriminative ability. This ratio balances subjective scenario requirements with objective feature attributes. Different scenarios have different weight allocations: in emergency scenarios, short-term features have a weight of 0.7, medium-term features 0.2, and long-term features 0.1, because emergency scenarios require rapid keyword capture. Short-term features have the highest weight. In typical scenarios, short-term features have a weight of 0.2, medium-term features 0.6, and long-term features 0.2. Medium-term features have the highest weight because typical scenarios need to consider both keywords and sentence fragments. In maintenance scenarios, short-term features have a weight of 0.1, medium-term features 0.2, and long-term features 0.7. Long-term features have the highest weight because maintenance scenarios require complete semantics. Weights can be manually adjusted by ±0.1 to adapt to individual on-site needs. Weights are automatically optimized monthly, and the consistency check error does not exceed 0.05. Consistency checks ensure that there are no logical conflicts in weight allocation, and the error is controlled within 0. The weights are kept within 0.05 to ensure reasonable weighting. An initial vector is obtained by fusion using weighted summation and a dynamic attention mechanism. The dynamic attention mechanism has a window range of 32 to 64 dimensions. Features with a relevance of at least 0.85 are given a weight of 1.2 times to strengthen effective features, while noisy features with a relevance below 0.7 have their weights reduced to weaken ineffective features. Redundant features are removed according to semantic constraints (relevance at least 0.85, discriminability at least 0.65, and matching at least 0.7), resulting in a filtered vector with a dimension not exceeding 256. The Pearson correlation coefficient is used to... Cross-scale feature association enhancement is performed with a correlation coefficient of at least 0.75. Cross-scale features with a correlation coefficient of at least 0.75 have strong correlations. Feature representation is enhanced through weighted fusion. After semantic calibration, a 256-dimensional fusion vector is output. The fusion vector is required to have a discriminability of at least 90% and a noise suppression ratio of at least 25dB. If the standard is not met, the weights are readjusted. The criteria for judging the failure of fusion quality are clearly defined. If the discriminability is less than 90% or the noise suppression ratio is less than 25dB, the attention window size (±5 dimensions) is adjusted first, and then the weight parameters are adjusted. The fusion delay is no more than 60ms.

[0030] The scene- and logic-driven recognition stage employs a hybrid network model. This model's hierarchical structure is: Input Layer → Feature Mapping Layer → Fusion Enhancement Layer → Logical Inference Layer → Classification Layer → Output Layer. The 256-dimensional fusion vector is processed by four convolutional layers, three pooling layers, and a 512-node BiLSTM layer. The convolutional layers extract local features, the pooling layers retain key features, and the BiLSTM layer enhances temporal correlations with a dropout setting of 0.2. The inference layer loads a causal logic graph and historical rules to achieve semantic consistency verification, instruction type prediction, and logical conflict elimination. A fully connected layer with a Softmax function matches 80 types of scheduling instructions. The fully connected layer maps features to the dimensions of the 80 instruction types. tmax calculates the probability of each category to classify instructions, and simultaneously associates the top 3 results with 6 months of historical data to improve recognition accuracy and avoid single-shot recognition errors. It prioritizes outputting results with a confidence score of at least 0.95 and a logical fit of at least 0.9. The output includes complete information items: instruction content, voice source, generation time (milliseconds), compliance identifier, power grid status matching degree, priority, confidence score, logical fit, and instruction type label. The recognition latency is no more than 90ms. An online model update mechanism is adopted, fine-tuning the model parameters monthly based on 5000 new voice data entries: learning rate ±0.0001 and dropout ratio ±0.05, ensuring the model adapts to new scenarios.

[0031] The cross-dimensional linkage verification process extracts key information with an accuracy rate of no less than 99.5%. The overall accuracy is calculated using a weighted average of 0.3 for edit distance, 0.4 for cosine similarity, and 0.3 for semantic similarity, with a threshold of 98%. Cosine similarity has the highest weight because feature similarity has the greatest impact on recognition accuracy; a review is triggered when the confidence level is below 0.95 or the logical fit is below 0.85, with a review threshold of 95%. Four-fold logical compliance verification and four-dimensional scenario consistency verification are performed. The four-fold logical compliance verification includes keyword combination verification, operation sequence verification, device association verification, and logical conflict verification. The four-dimensional scenario consistency verification includes fault and handling verification, load and regulation verification, equipment status and operation verification, and urgency and priority verification. If all three verifications pass, the system is deemed qualified. If one verification fails, semantic calibration is triggered. If two or more verifications fail, the system is deemed unqualified. The reasons for unqualified results are recorded, such as semantic deviation, logical conflict, feature mismatch, etc. An optimization report is generated, which includes parameter adjustment suggestions and triggers parameter adjustment or manual intervention. The verification delay does not exceed 50ms. The conditions for triggering manual intervention are clearly defined. In addition to failing to meet the standard after three reprocessing attempts, failure to identify core fault commands such as line faults and equipment tripping directly triggers manual intervention.

[0032] The results processing and linkage stages store qualified results and support 10-year backtracking. A MySQL + Redis storage architecture is used, with MySQL storing historical data and Redis caching recent data. Data backtracking queries for 10 years are supported, and data encryption uses the AES-256 algorithm, meeting power grid information security requirements. Qualified results are synchronized to the dispatch system via the IEC 61850 interface, with a linkage delay not exceeding 150ms. Synchronized data is transmitted using encryption to ensure data security. Unqualified results are returned for reprocessing up to three times; failure to meet the standard after three reprocessing attempts triggers manual intervention. A results statistical analysis function is employed to statistically analyze indicators such as identification accuracy, verification pass rate, and linkage response time daily, generating a daily report that is pushed to the dispatch management platform.

[0033] In the scene adaptive preprocessing process, five new technical features have been added: integrated processing of multimodal noise identification, separation and distortion compensation; online updating of the noise adaptive suppression model; dynamic frame division based on instruction priority; multi-scale feature extraction; and adaptive normalization double verification. The details are as follows: The pre-processing includes a multimodal noise processing model and a frame parameter configuration library. The multimodal noise processing model is a CNN-LSTM hybrid model for precise noise processing. Multimodal refers to the simultaneous input of speech time-domain and frequency-domain features. The model extracts multimodal features through CNN and captures temporal correlations through LSTM, achieving accurate identification of noise type and intensity, adapting to mixed noise scenarios in substations. The frame parameter configuration library pre-sets three types of instruction frame lengths and frame shift ranges, associating them with the power grid status for rapid parameter matching. The library stores the frame parameter ranges corresponding to different power grid states (normal, fault, and maintenance), allowing for rapid matching of optimal frame parameters without real-time calculation. The core algorithm used is consistent with the preprocessing stage. A new XGBoost algorithm is added to build an adaptive noise suppression model. This algorithm has strong classification and regression capabilities, adapts to the mapping between noise features and denoising parameters, and uses Gradient Boosting Tree (GBDT) to suppress overfitting through regularization. It supports parallel computation and can construct a mapping relationship between noise features (type, intensity) and denoising parameters (notch frequency, filter threshold), achieving adaptive adjustment of denoising parameters. The supplementary software dependency is XGBoost. Library 1.7.5 requires a CPU that supports the AVX2 instruction set to improve model training efficiency.

[0034] The multimodal noise integrated processing stage takes raw speech data as input and extracts time-domain features, including amplitude and zero-crossing rate, and frequency-domain features, including spectral density and spectral entropy, using a CNN-LSTM model to accurately identify noise type and intensity. Noise types are covered, including transformer noise, circuit breaker operation noise, and environmental wind noise, with intensity ranging from 40-80 dB and divided into 5 levels. The FastICA algorithm is used to separate the target speech from noise, improving the signal-to-noise ratio by at least 20 dB. FastICA effectively separates the target speech from mixed noise by maximizing the non-Gaussianity of the signal, resulting in a high signal-to-noise ratio. A gain of 20dB or more ensures speech clarity; a gain adjustment scheme is used to address amplitude attenuation, with a gain coefficient ranging from 1.1 to 1.3, dynamically adjusted according to the degree of amplitude attenuation. This range avoids speech distortion caused by excessive gain; a phase calibration scheme is used to address frequency shift, adjusting the speech signal phase by calculating the frequency shift, correcting the frequency shift to ensure the accuracy of speech frequency characteristics, and ensuring overall speech distortion is below 5%; a noise type recognition feedback mechanism is used, if the recognition accuracy is below 99%, supplementing the model with at least 500 training data points for each noise type and retraining the model.

[0035] The noise adaptive suppression model construction and update process uses a power grid equipment noise feature library as training data and employs the XGBoost algorithm to build the model. Input noise features include type, intensity, and spectral parameters, while output noise reduction parameters include notch frequency and filtering threshold. During model training, noise features are used as input and noise reduction parameters as output. Gradient boosting iterative optimization achieves accurate mapping of the noise reduction parameters. Model training uses 5-fold cross-validation, and deployment is only permitted after the validation set accuracy is no less than 98.5%. At least 1000 new data points are automatically collected monthly, including noise from new equipment and noise under extreme operating conditions. The feature library is updated and the model is retrained. New data containing noise features from new equipment and operating conditions improves the model's adaptability to new scenarios and ensures continuous and stable noise reduction performance. A model version management mechanism is used, retaining the last three model versions. If a new model is ineffective, it can be rolled back to the historical best version.

[0036] The adaptive framing stage for command priority combines voice keywords with grid status to determine priority. Voice keywords such as fault, isolation, and trip correspond to level 1 commands; adjustment and regulation correspond to level 2 commands; and maintenance and repair correspond to level 3 commands. Command priority is quickly determined through keyword matching, improving framing efficiency by eliminating the need for complete voice content parsing. Grid statuses correspond to level 1 commands for fault states, level 2 commands for normal states, and level 3 commands for maintenance states. Level 1 command frame lengths range from 10 to 15 ms (default 12 ms), with frame shifts ranging from 5 to 8 ms (default 6 ms). Short frame lengths allow for rapid capture of urgent keywords, ensuring real-time fault handling. Level 2 command frame lengths range from 50 to 80 ms. The default frame length is 60ms, with a frame shift range of 25 to 40ms and a default of 30ms. This moderate frame length balances keyword recognition and sentence integrity, adapting to common adjustment commands. The frame length range for level 3 commands is 200 to 400ms, with a default of 300ms. The frame shift range is 100 to 200ms, with a default of 150ms. Longer frame lengths can focus on maintenance semantics, adapting to the long sentence requirements of maintenance scenarios. Manual configuration of ±5ms is supported to adapt to personalized on-site needs, and the parameter adjustment response time does not exceed 30ms. A frame segmentation effect evaluation mechanism is adopted, which evaluates the performance based on two indicators: intra-frame keyword integrity of no less than 95% and inter-frame information redundancy of no more than 5%. If the standards are not met, the frame length and frame shift parameters are dynamically adjusted.

[0037] The multi-scale speech feature extraction stage employs three algorithms: time-domain analysis, Fourier transform, and wavelet transform, to extract short-term, mid-term, and long-term features, respectively. Short-term features include instantaneous amplitude, instantaneous frequency, and instantaneous phase, with 16 feature values ​​per frame focusing on detailed information. Time-domain analysis extracts features such as instantaneous amplitude and zero-crossing rate, reflecting instantaneous changes in speech to suit keyword recognition. Mid-term features include spectral peak value, spectral bandwidth, spectral centroid, and spectral entropy, with 16 feature values ​​per frame focusing on segment information. Fourier transform converts the speech to the frequency domain, extracting features such as spectral peak value and bandwidth, reflecting the frequency distribution of speech segments. Long-term features include time-frequency entropy, energy, and correlation, with 16 feature values ​​per frame focusing on semantic information. Wavelet transform enables joint time-frequency analysis, extracting features such as time-frequency entropy and energy, reflecting the overall semantics of the speech, with an extraction delay of no more than 25ms. Detailed feature extraction parameters are specified: the Fourier transform window is 20ms, the wavelet transform scale is 4 levels, and the wavelet basis function is db4, ensuring the reproducibility of the extraction process.

[0038] The adaptive normalization and dual verification process employs Min-Max adaptive normalization, dynamically adjusting the mapping interval based on the feature distribution type. Features deviating from the normal distribution are mapped to the [0.05, 0.95] interval, while normally distributed features are mapped to the [0.1, 0.9] interval. This setting avoids excessive compression of extreme values ​​of normally distributed features. Feature validity verification uses the 3σ criterion and analysis of variance. The 3σ criterion considers values ​​exceeding the mean ± 3σ as outliers, effectively eliminating abnormal features caused by acquisition errors. Analysis of variance removes features with a variance lower than 0.005, as these features carry little information and are deemed invalid. Semantic relevance verification uses cosine similarity calculation, removing features with a relevance lower than 0.7, as these features have weak relevance to scheduling services, reducing subsequent processing pressure. Overall, the effective feature retention rate is required to be no less than 90%, and the processing delay should not exceed 10ms. A failure handling mechanism is adopted: if the effective feature retention rate is lower than 90%, normalization is re-executed to adjust the mapping interval to ± 0.05; if it still fails to meet the standard, the process returns to the noise processing stage for reprocessing.

[0039] In the targeted feature extraction process, the following new rules were added for constructing short / medium / long-term feature vectors, feature regularization and L1-L2 hybrid regularization sparsity processing, and dual detection technology features, as detailed below: The pre-processing includes four main template and rule libraries: a core keyword speech template library, a pronunciation association rule library, a terminology and vocabulary combination rule library, and a scheduling instruction structure template library. The core keyword speech template library contains over 80 scheduling terms and pronunciation variations for keyword matching. The library stores standard spectral templates for each term and variant templates for different accents, using cosine similarity matching for rapid keyword localization, adapting to different accent scenarios. The pronunciation association rule library defines phoneme transition and stress features to improve accent adaptability. It includes rules for phoneme connection in Mandarin Chinese and stress shift rules for different accents, correcting pronunciation deviations caused by accents and ensuring accurate feature extraction. The terminology and vocabulary combination rule library includes rules such as "device + operation" to ensure feature relevance. These rules define valid terminology combinations and exclude invalid combinations, filtering feature components relevant to scheduling operations. The scheduling instruction structure template library contains three types of structure templates for semantic filtering. These templates contain grammatical constraints on scheduling instructions, filtering feature vectors that conform to the instruction structure and eliminating invalid features.

[0040] A time-domain waveform analysis algorithm is used to extract short-term features, extracting instantaneous features such as peak value and zero-crossing rate by analyzing the speech time-domain waveform to reflect the instantaneous changes in speech. A Fourier transform algorithm is used to extract mid-time features, converting the speech signal from the time domain to the frequency domain and extracting frequency features such as spectral peak value and bandwidth to reflect the frequency distribution of speech segments. A wavelet transform algorithm is used to extract long-term features, achieving joint time-frequency analysis and simultaneously capturing the time and frequency features of speech to reflect the overall semantics of the speech. A cosine similarity algorithm is used for keyword matching and semantic association detection, calculating the cosine similarity between the feature vector and the template vector to measure their similarity. A variance enhancement algorithm with an enhancement coefficient of 1.2 is used to strengthen effective features, amplifying the effective features by multiplying the variance by 1.2. The differences in effectiveness of features enhance feature discriminative power; the Word2Vec model is used for semantic similarity calculation, mapping words to low-dimensional vectors, which are closer to semantically similar word vectors, and is used to detect the semantic correlation between features and scheduling terms; the L1-L2 hybrid regularization algorithm is used for feature sparsity, with L1 regularization achieving feature selection and L2 regularization preventing overfitting, and the combination of the two balancing sparsity and model generalization ability; the variance ratio calculation algorithm is used for discriminative power detection, calculating the variance ratio of different categories of features, with a larger variance ratio indicating stronger feature discriminative power; a supplementary template library and rule library maintenance mechanism is implemented, adding no less than 5 new scheduling terms, updating no less than 10 pronunciation variant templates, and optimizing no less than 5 word combination rules every quarter to ensure the timeliness of the library content.

[0041] The short-time feature vector extraction stage employs a time-domain waveform analysis algorithm to extract peak value, zero-crossing rate, instantaneous amplitude, instantaneous frequency, and instantaneous phase features. The extracted time-domain features are then normalized and mapped to the [0.05, 0.95] interval. A keyword template library is loaded to calculate cosine similarity, requiring a matching degree of at least 85%. This matching degree ensures accurate keyword recognition and avoids mismatches. Feature components with a correlation higher than 0.8 are retained based on the pronunciation rule library. These feature components match pronunciation rules and can effectively adapt to accent scenarios. A feature splicing method is used... The formula integrates and constructs a 256-dimensional vector, with each feature accounting for 30% of instantaneous amplitude, 30% of instantaneous frequency, and 40% of instantaneous phase. This proportion can highlight the instantaneous phase feature, as the phase feature contributes more to keyword recognition, and the extraction delay does not exceed 20ms. The feature concatenation order is clearly defined as instantaneous amplitude → instantaneous frequency → instantaneous phase to ensure the consistency of vector construction. A keyword matching failure handling mechanism is added. If the matching degree is less than 85%, the matching range of the template library is expanded to include similar term templates. If it still does not meet the standard, it returns to the preprocessing stage to re-extract features.

[0042] The mid-time feature vector extraction stage employs a 20ms window Fourier transform with a Hanning window function to analyze frequency features and extract spectral peaks, bandwidth, centroid, and entropy features. Spectral peaks reflect the peak position of the frequency distribution, bandwidth reflects the frequency range, centroid reflects the center of the frequency distribution, and entropy reflects the uniformity of the frequency distribution. A variance enhancement algorithm with an enhancement coefficient of 1.2 is used to strengthen feature differences, amplify the variance differences of effective features, and improve feature discriminability. The enhanced features are then reduced to 64 dimensions using PCA. Feature components matching "device + operation" are selected based on a vocabulary combination rule base, and feature components related to scheduling services are selected to eliminate invalid features. A 256-dimensional vector is then constructed, with each of the four types of features accounting for 25%, and the even distribution ratio can comprehensively reflect the mid-time frequency features with an extraction delay of no more than 20ms. Details of the Fourier transform parameters are supplemented, with 1024 sampling points and a frequency resolution of 100Hz to ensure the reproducibility of the extraction process. A feature enhancement effect evaluation mechanism is added, with evaluation based on a feature variance ratio of no less than 1.5. If the standard is not met, the enhancement coefficient is adjusted to the range of 1.1 to 1.3 for re-enhancement.

[0043] The long-term feature vector extraction stage employs a 4-scale wavelet transform with db4 as the wavelet basis function. Time-frequency joint analysis is performed to extract time-frequency entropy, energy, and correlation features. Time-frequency entropy reflects the complexity of the time-frequency distribution, energy reflects the intensity of the speech signal, and correlation reflects the degree of association between features at different scales. The extracted long-term features are smoothed with a sliding window of 5ms. Semantic similarity is calculated using a Word2Vec model (vector dimension 128, window size 10 words, distance metric using Euclidean distance). Features with a semantic similarity higher than 0.8 are retained, as these features are strongly associated with scheduling terms. To preserve the semantic validity of features, three types of templates are matched based on the instruction structure template library. Feature components that conform to the structure are retained, and feature vectors that conform to the scheduling instruction structure are selected to improve the adaptability of features to business. A 256-dimensional vector is constructed by integrating these vectors, with each feature accounting for 30% of time-frequency entropy, 30% of energy, and 40% of relevance. Relevant features are highlighted because they contribute more to the complete semantic recognition, and the extraction latency does not exceed 20ms. If the proportion of features with semantic similarity below 0.8 exceeds 20%, Word2Vec model training data is supplemented (no less than 1,000 new scheduling term corpora are added), and the model is retrained.

[0044] In the feature normalization and sparsity processing stage, a dimension normalization algorithm is used to unify the three types of vectors to 256 dimensions. When the dimension is insufficient, zero vectors are used for padding; when the dimension exceeds the limit, core features are pruned. Pruning follows the maximum variance priority principle, retaining the 256 feature components with the largest variance. This processing ensures that the vector dimensions are uniformly adapted to the subsequent fusion process. Sparsity processing is performed through L1-L2 hybrid regularization, with an L1 weight of 0.4 and an L2 weight of 0.6. The regularization loss function is calculated as follows: Regularization loss = Original loss function + λ1×||w||1 + λ2×||w||2 2 Where λ1 and λ2 are the L1 and L2 regularization coefficients, respectively, λ1=0.01 and λ2=0.001, w is the feature weight vector, ||w||1 is the L1 norm (the sum of the absolute values ​​of all elements in vector w), and ||w||2 is the L2 norm. 2 The L2 norm is the square of the vector w (the sum of squares of all elements in the vector w). The sparsity threshold is set to 0.005, meaning that feature components with an absolute weight value lower than 0.005 are directly removed, achieving precise filtering of invalid and redundant features. An adaptive adjustment mechanism for regularization parameters is added. If the proportion of effective features is lower than 70% after sparsification, the L1 weights are adjusted to the range of 0.3 to 0.5 and the L2 weights to the range of 0.5 to 0.7 for re-sparserization.

[0045] The dual detection process calculates the feature discrimination index using variance ratio, which is calculated as between-class variance / within-class variance. The formula is: Discrimination Index = Sum of Between-Class Variances / Sum of Within-Class Variances. A discrimination index of at least 0.6 is required to ensure that features can effectively distinguish different scheduling instructions. Semantic relevance is detected using cosine similarity, requiring a semantic relevance of at least 0.7. This relevance ensures that features are consistent with the scheduling business semantics. If the criteria are not met, features are re-extracted, with a processing delay of no more than 10ms. The dual detection judgment logic is clearly defined: both indicators must meet the criteria for a successful detection; if one indicator fails, it is considered a suspected failure and requires re-extraction once; if both indicators fail, it is considered a failure and requires reprocessing in the preprocessing stage. A detection result recording mechanism is added to record in detail the type, number, and reason for features that fail to meet the criteria, providing a basis for subsequent parameter optimization.

[0046] In the process of triple-linkage feature fusion, the following features have been added: a triple-driven feature importance evaluation model, a dynamic attention mechanism, semantic constraint selection rules, and cross-scale feature enhancement techniques. The pre-processing includes a feature importance assessment index system, a set of semantic constraint screening rules for power grid dispatching, and a cross-scale feature association enhancement algorithm. The feature importance assessment index system comprises four indicators: scenario adaptability, semantic relevance, recognition sensitivity, and noise resistance, with weights of 0.3, 0.3, 0.2, and 0.2 respectively. This index system covers four dimensions: scenario, semantics, recognition, and anti-interference. The weight allocation is adapted to the needs of power grid dispatching scenarios, with scenario and semantics having the highest weights because they have the greatest impact on the fusion effect. The set of semantic constraint screening rules for power grid dispatching sets screening thresholds of no less than 0.85 for relevance, no less than 0.65 for discrimination, and no less than 0.7 for template matching, ensuring that the screened features possess high relevance, high discrimination, and high adaptability. The cross-scale feature association enhancement algorithm is based on the Pearson correlation coefficient, identifying strongly correlated features by calculating the Pearson correlation coefficient of cross-scale features and performing weighted fusion enhancement.

[0047] The Analytic Hierarchy Process (AHP) is used for subjective weight calculation. By constructing a hierarchical model, it invites scheduling experts to score and determine the subjective weight of features, adapting to subjective needs such as scenarios and priorities. The entropy weight method is used for objective weight calculation. Based on feature information entropy, the smaller the entropy value, the stronger the feature's discriminative ability and the greater the weight, reflecting the objective attributes of the features. The weighted fusion algorithm is used for weighted summation of features. It integrates multi-scale features by weighting the elements of the feature vector according to the weight coefficient. The dynamic attention mechanism algorithm is used to focus on effective features. It calculates the correlation between feature vectors based on the self-attention model and assigns high weights to effective features. The semantic constraint rule matching algorithm is used for feature screening. It performs rule matching on feature vectors one by one to eliminate invalid features that do not meet the rules. The Pearson correlation coefficient algorithm is used for cross-scale correlation detection. It measures the linear correlation of cross-scale features to provide a basis for correlation enhancement. The feature correlation enhancement algorithm is used for cross-scale feature enhancement. It performs weighted fusion on strongly correlated cross-scale features to improve feature expression capabilities. A supplementary indicator system maintenance mechanism evaluates the rationality of indicator weights every six months and adjusts the weights by ±0.05 according to the fusion effect to ensure that the indicator system adapts to business needs.

[0048] The feature importance weight calculation stage employs a triple-driven evaluation model. Input scenario types include emergency, routine, and maintenance; power grid status includes fault, normal, and maintenance; and command priorities include first-level, second-level, and third-level features and three types of feature vectors. Subjective weights are determined using the analytic hierarchy process (AHP), with 5 to 8 dispatching experts scoring the results. The consistency test CR < 0.1 is used to determine objective weights. Based on feature information entropy, a smaller entropy value indicates stronger feature discrimination ability and a larger weight, reflecting the objective attributes of the features. The weighting ratio of subjective to objective weights is 0.5:0.5, balancing subjective scenario requirements with objective feature attributes to ensure reasonable weight allocation. The final weight allocation is as follows: emergency scenario: short-term 0.7, medium-term 0.2, long-term 0.1; routine scenario: short-term 0.2, medium-term 0.6, long-term 0.2; maintenance scenario: short-term 0.1, medium-term 0.2, long-term 0. 7. This weight allocation adapts to the core needs of different scenarios: emergency scenarios focus on keywords, routine scenarios focus on sentence fragments, and maintenance scenarios focus on complete semantics. The weights support manual fine-tuning of ±0.1 with a response time of no more than 5ms, adapting to personalized on-site needs. The weight calculation cycle is 25ms, and it is automatically optimized monthly based on historical fusion results, including fusion feature recognition and noise suppression ratio. The consistency verification error does not exceed 0.05, ensuring that the weight allocation has no logical conflicts and that the error is controlled within 0.05 to ensure the rationality of the weights. An abnormal weight calculation handling mechanism is added: if the consistency verification error exceeds 0.05, experts are invited to re-score or the entropy weight method calculation window is adjusted by ±50 feature samples for recalculation. A clear weight storage mechanism is defined: the current weights are cached in Redis, and historical weights are stored in MySQL and retained for 12 months, supporting weight backtracking.

[0049] The dynamic attention fusion process involves weighted summation by multiplying the weight coefficients by the feature vector elements to obtain a 256-dimensional initial fusion vector. The attention window is adaptively adjusted based on the scene and feature effectiveness, ranging from 32 to 64 dimensions: 40 dimensions for emergency scenarios, 50 dimensions for routine scenarios, and 60 dimensions for maintenance scenarios. The window size adapts to the feature distribution of different scenarios; a smaller window in emergency scenarios focuses on keywords, while a larger window in maintenance scenarios focuses on complete semantics. The attention weight calculation uses a softmax function to ensure the sum of weights is 1. Features with a correlation of at least 0.85 are given a 1.2x enhancement weight to strengthen effective feature expression, while noisy features with a correlation of less than 0.7 are given a 0.5x suppression weight to weaken invalid feature interference. The fusion delay is no more than 20ms. A new real-time evaluation mechanism for the fusion effect is added, evaluating the initial fusion vector based on a signal-to-noise ratio of at least 22dB. If this standard is not met, the attention window size is adjusted by ±5 dimensions for re-fusion. The weighted summation calculation order is explicitly defined as short-term features → medium-term features → long-term features to ensure computational consistency.

[0050] The semantic constraint filtering process loads a set of filtering rules and performs rule matching on each feature component of the initial vector one by one, eliminating feature components that do not meet any of the rules. If a feature does not meet even one rule, it is determined to be invalid, ensuring the high quality of the filtered features. The output is a filtered vector with a dimension not exceeding 256 (the default dimension is 200 to 220), and the filtering latency is not more than 10ms. The removal rate of redundant, invalid, and residual noise features is not less than 85%, which can effectively reduce the pressure on subsequent processing and improve recognition efficiency. A filtering dimension control mechanism is added. If the dimension of the filtered vector is less than 180, the filtering rule thresholds are lowered to a correlation of not less than 0.8, a discriminant of not less than 0.6, and a template matching of not less than 0.65 for re-filtering to avoid excessive feature loss. The feature component removal rules are clearly defined, prioritizing the removal of noisy features, then redundant features, and finally invalid features, ensuring that core features are not lost.

[0051] The cross-scale feature enhancement and semantic calibration process calculates the Pearson correlation coefficients of features at different scales, including short-time-medium-time, medium-time-long-time, and short-time-long-time. Feature pairs with correlation coefficients no lower than 0.75 are enhanced. The enhancement target is clearly defined as each pair of cross-scale features (short-time-medium-time, medium-time-long-time, and short-time-long-time are treated as independent feature pairs). The weighted summation formula is: Enhanced Feature = 0.5 × Feature A + 0.5 × Feature B (Feature A and Feature B are the two types of features in each pair of cross-scale features), with equal weights to ensure that the feature information of both sides is preserved. After enhancement, the features are normalized and mapped to the [0.05, 0.95] interval. Features that conflict with scheduling logic are corrected in conjunction with semantic constraint rules, and features that do not conform to the scheduling logic are corrected. To ensure semantic coherence of features, the following steps are taken: The fusion feature discriminability and noise suppression ratio are calculated. The fusion feature discriminability is calculated using the feature-standard template matching rate, with the formula: Discriminability = Number of successfully matched feature components / Total number of feature components × 100%. A discriminability of at least 90% and a noise suppression ratio of at least 25dB are required. A discriminability of at least 90% ensures accurate feature recognition by subsequent models, while a noise suppression ratio of at least 25dB ensures effective suppression of noise interference in the features. If these standards are not met, the process returns to the weight calculation step for readjustment, with an enhancement delay not exceeding 15ms. The final output is a 256-dimensional fusion vector. A mechanism for handling association enhancement failures is added. If the feature discriminability is still below 90% after association enhancement, the process returns to the semantic constraint filtering stage to adjust the filtering rule threshold and re-filter.

[0052] In the dual-driven recognition process of scene and logic, a new hybrid network model structure, scheduling logic reasoning mechanism, model training optimization strategy, and technical features of recognition result information items are added, as follows: The preliminary preparation includes a hybrid network model, a set of historical instruction reasoning rules, and model training parameters. The hybrid network model has a hierarchical structure of input layer → feature mapping layer → fusion enhancement layer → scheduling logic reasoning layer → classification layer → output layer. This hierarchical structure adapts to the entire process of feature processing, logical reasoning, and classification output, ensuring recognition accuracy and logical consistency. The set of historical instruction reasoning rules is built based on 6 months of historical data and includes rules for associating fault and adjustment instructions. The rule set contains the relationships between historical scheduling instructions, such as the correspondence between fault instructions and handling instructions, to assist in logical reasoning. The model training parameters are set with an initial learning rate of 0.001, decaying to 0.0001 after 100 rounds, a gradient clipping threshold of 0.5, and an early stopping patience of 20. The initial learning rate of 0.001 ensures fast convergence, and decaying to 0.0001 after 100 rounds avoids overfitting. The gradient clipping threshold of 0.5 prevents gradient explosion. The early stopping patience of 20 means that training stops if there is no improvement in accuracy on the validation set after 20 rounds to avoid overfitting.

[0053] Convolutional algorithms are used to extract local features, capturing spatial correlations by sliding convolution kernels to extract local features from the feature map. Pooling algorithms are used for dimensionality reduction and noise reduction, using MaxPooling / AvgPooling to reduce the dimensionality of the feature map, retaining key features and removing redundant information. BiLSTM algorithms are used to enhance temporal correlations, with bidirectional LSTM simultaneously capturing the positive and negative temporal correlations of features, adapting to the temporal characteristics of speech signals. Logical reasoning algorithms are used for semantic verification and conflict resolution, based on causal logic graphs and historical rules, eliminating logical conflicts by using the semantic consistency of reasoning features. Softmax activation functions are used for class probability calculation, mapping the model output to class probabilities to achieve instruction classification. The Adam optimizer is used for model training, combining momentum and adaptive learning rate for fast convergence and avoiding local optima. Cross-entropy loss functions address sample imbalance by balancing weights, improving the model's accuracy in recognizing niche instructions. Dropout strategies are used to prevent overfitting, randomly discarding some neurons during training to reduce over-dependence and improve generalization ability.

[0054] In the model training and optimization phase, a dedicated dataset for power grid dispatching voice is loaded and divided into training, validation, and test sets in an 8:1:1 ratio. This ratio balances the training sample size with the validation and testing performance, ensuring the model is adequately trained and effectively evaluated. The Adam optimizer and cross-entropy loss function are used to train the model, combined with gradient pruning and early stopping strategies to prevent overfitting. The model is only deployed after achieving a recognition accuracy of no less than 98% and a logical fit of no less than 95% on the test set. Strict deployment thresholds ensure the model's recognition quality. During the online iteration phase, parameters are fine-tuned monthly based on no less than 5,000 new voice data entries, including new scenarios and techniques. For speech recognition, fine-tuning parameters can improve the model's adaptability to new scenarios and optimize inference rules. The activation functions for each layer of the hybrid network model are as follows: all four convolutional layers use the ReLU activation function (calculated as f(x)=max(0,x)) to enhance the model's non-linear expressive ability; the BiLSTM layer has no additional activation function, only performing linear transformation through the output layer; the fully connected layer uses the LeakyReLU activation function (calculated as f(x)=max(0.01x,x)) to avoid the gradient vanishing problem; the classification layer uses the Softmax activation function for class probability calculation.

[0055] The feature mapping process inputs a 256-dimensional fused vector into the model, which is then processed through four convolutional layers to extract local features. The parameters of the convolutional layers are Conv1: 3×3×256, Conv2: 5×5×128, Conv3: 3×3×64, and Conv4: 5×5×32, with a stride of 1 and Same padding. Alternating between 3×3 and 5×5 convolutional kernels captures local features at different scales, while Same padding ensures that the feature map size is not reduced. Gradually decreasing the number of channels achieves feature dimensionality reduction. Three pooling layers retain key features and smooth features. The pooling layer parameters are MaxPool1 / 2 and AvgPool3, with a window size of 2×2 and a stride of 2. MaxPool1 / 2 retains key features, while AvgPool3 smooths features and avoids interference from local extrema. The final output is a 32×32 feature map with a processing latency of no more than 20ms.

[0056] The bidirectional temporal correlation enhancement step flattens the feature map into a 1024-dimensional vector, inputs it into a 512-node BiLSTM layer, sets dropout to 0.2 and recurrent_dropout to 0.1; the 512 nodes ensure the ability to capture temporal features, dropout=0.2 and recurrent_dropout=0.1 are applied to the input layer and the recurrent layer respectively to prevent overfitting, enhance the temporal correlation of multi-scale features and suppress residual noise, output a 512-dimensional temporal feature vector, and the processing delay does not exceed 25ms.

[0057] The scheduling logic reasoning and conflict resolution process loads causal logic graphs and historical rules to verify the consistency between temporal features and semantic terminology, and to verify the semantic matching degree between features and scheduling terms to avoid semantic deviations. Semantic deviations are corrected by logical reasoning to ensure the accuracy of feature semantics. Reasonable instruction types are predicted in conjunction with power grid status, and instruction types that conform to the current state are associated to improve recognition accuracy. Feature mapping results that do not conform to logical rules are eliminated, as are results that conflict with scheduling logic, such as closing instructions under fault conditions. 256-dimensional logically consistent feature data is output with a processing delay of no more than 30ms.

[0058] The category matching and result output stage maps features to 80 categories of scheduling instructions through a fully connected layer, which in turn maps 256-dimensional features to 80 instruction dimensions to achieve instruction classification. Softmax is used to calculate the probability of each category of instruction, and the result with the highest probability is selected as the candidate result. The top three results are then compared with six months of historical data to improve recognition accuracy and avoid single-shot errors. Results with a confidence level of at least 0.95 and a logical fit of at least 0.9 are prioritized for output. High confidence and logical fit ensure reliable results. The output includes complete information items: instruction content, voice source, generation time (milliseconds), compliance identifier, grid status matching degree, priority, confidence level, logical fit, and instruction type label. This information covers all business elements and meets scheduling business requirements. If no results meet the criteria, the top three are output for review, with a processing latency of no more than 15ms.

[0059] The cross-dimensional linkage verification process incorporates a new triple verification mechanism, accuracy calculation weights, cross-dimensional linkage judgment, and technical features for handling unqualified results, as detailed below: The system includes pre-processing for accuracy calculation weight configuration, a four-fold verification mechanism for scheduling logic compliance, a four-dimensional matching model for scenario consistency, and a verification report template. The accuracy calculation weight configuration is 0.3 for edit distance, 0.4 for cosine similarity, and 0.3 for semantic similarity, with a comprehensive accuracy threshold of 98%. Cosine similarity has the highest weight because feature similarity has the greatest impact on recognition accuracy. Edit distance reflects text differences, and semantic similarity reflects semantic consistency; the combination of these three comprehensively evaluates accuracy. The 98% threshold ensures strict verification standards. The four-fold verification mechanism for scheduling logic compliance includes keyword combination verification, operation sequence verification, equipment association verification, and logical conflict verification. These four verifications cover the core logical dimensions of scheduling instructions, ensuring logical compliance. The four-dimensional matching model for scenario consistency includes fault-handling matching, load-regulation matching, equipment status-operation matching, and urgency-priority matching. This four-dimensional matching correlates the power grid status with the instructions, ensuring that the instructions are adapted to the scenario. The verification report template includes verification results, non-compliance types, reasons, and optimization suggestions. The standardized template outputs verification reports, providing a basis for parameter adjustment and manual intervention.

[0060] Regular expression parsing algorithms are used for key information extraction, quickly extracting core content by matching scheduling terms, operation instructions, and other key information using regular expressions; semantic parsing algorithms are used for semantic logic analysis, analyzing the semantic logic relationships of instructions to verify semantic consistency; edit distance algorithms are used for text difference assessment, calculating the number of character operations between the recognized text and the standard text to measure text differences; cosine similarity algorithms are used for feature similarity assessment, calculating the similarity between recognized features and standard features to measure the degree of feature matching; the Word2Vec model is used for semantic similarity calculation, mapping words to vectors to measure the semantic consistency between the recognized text and the standard text; and rule matching algorithms are used for logical compliance verification, matching the recognition results with logical rules to verify logical compliance.

[0061] The key information extraction process is based on regular expressions and semantic parsing algorithms to extract scheduling instruction keywords, semantic logic, contextual information, compliance identifiers, grid status matching degree, confidence level, logical adaptability, priority, and instruction type tags. Regular expressions quickly extract keywords, and semantic parsing algorithms analyze semantic logic to ensure comprehensive and accurate key information extraction with an accuracy rate of no less than 99.5%. This high accuracy rate ensures the reliability of subsequent verification, with a processing delay of no more than 10ms.

[0062] The accuracy verification process loads a standard speech database, compares the extracted information with the standard information, and calculates edit distance similarity, cosine similarity, and semantic similarity. Edit distance similarity reflects text differences; the smaller the edit distance, the smaller the text differences and the higher the similarity. Cosine similarity reflects feature similarity; the higher the cosine similarity, the higher the feature matching degree. Semantic similarity is calculated using the Word2Vec model; the higher the semantic similarity, the stronger the semantic consistency. The comprehensive accuracy calculation formula is: edit distance similarity × 0.3 + cosine similarity × 0.4 + semantic similarity × 0.3. A comprehensive accuracy of not less than 98% is considered satisfactory. A review is triggered when the confidence score is lower than 0.95 or the logical fit is lower than 0.85. The review threshold is 95%. Results with low confidence and low logical fit need to be reviewed to reduce the false positive rate. The 95% review threshold balances accuracy and efficiency, with a processing delay of no more than 15ms.

[0063] The dispatch logic compliance verification process performs four checks: keyword combination verification to check the matching of the operation with the preconditions (e.g., closing the circuit requires passing the voltage test to avoid violations without preconditions); operation sequence verification to check compliance with dispatch procedures (e.g., the operation sequence of power outage → voltage test → grounding → maintenance to avoid safety risks caused by incorrect sequence); equipment association verification to check the matching of the operation command with the current status of the equipment (e.g., shut-down equipment cannot be closed); and logic conflict verification to eliminate contradictory commands (e.g., simultaneously closing and opening the same equipment). If any check fails, it is judged as a logical non-compliance, the non-compliance type is recorded, and the processing delay does not exceed 15ms.

[0064] The scenario consistency verification process performs four-dimensional matching, and the fault-handling matching associates fault types with corresponding handling instructions to ensure compatibility. The mapping between 12 types of faults and handling instructions is clearly defined as follows: 1. Line short circuit fault → Isolate the faulty line + arrange emergency repair; 2. Transformer overload fault → Reduce transformer load + monitor temperature; 3. Circuit breaker failure to operate fault → Manually operate the circuit breaker + maintenance mechanism; 4. Disconnect switch abnormal fault → Stop operation + check insulation; 5. Busbar grounding fault → Isolate the faulty busbar + test for voltage grounding; 6. Voltage transformer fault → Remove voltage transformer + replace equipment; 7. Current transformer fault → Remove current transformer + inspection and verification; 8. Capacitor bank fault → Disconnect capacitor bank + check for short circuits; 9. Reactor fault → Remove reactor + test windings; 10. Cable fault → Locate the fault point + excavate and repair; 11. Surge arrester fault → Replace surge arrester + test insulation; 12. Fuse blown fault → Replace fuse + investigate overload cause. To prevent mismatches between fault types and handling instructions; load-regulation matching verifies the rationality of regulation instructions by combining real-time grid load data. For example, when the load is overloaded, a load reduction regulation instruction must be matched to avoid issuing a load reduction instruction when the load is insufficient; equipment status-operation matching verifies the compatibility between operation instructions and the current equipment status. For example, operating equipment can perform a tripping operation, while shut-down equipment cannot perform a closing operation to prevent conflicts between operation and equipment status; urgency-priority matching calibrates the instruction priority according to the grid urgency level. For example, an equipment explosion fault corresponds to a level 1 urgency level and requires a level 1 instruction, while routine inspections correspond to a level 3 urgency level and require a level 3 instruction. If any matching fails, it is determined to be a scenario inconsistency, and the dimension and reason for the mismatch are recorded, with a processing delay of no more than 15ms.

[0065] The verification result processing stage requires simultaneous data archiving and process linkage. Qualified results, linked to the verification log (including verification time, verification personnel, and values ​​of various verification indicators), are pushed to the result processing and linkage stage. Unqualified results, in addition to generating an optimization report, require the error samples to be stored in the abnormal sample library of the power grid dispatch voice dataset for subsequent model optimization training. Verification reports are summarized daily, including the daily pass rate, distribution of unqualified types, parameter adjustment records, and number of manual interventions, and are pushed to the dispatch management platform for operation and maintenance personnel to view. A verification log traceability mechanism is added, supporting 10-year backtracking of all verification process data. Log data is stored using AES-256 encryption, complying with power grid information security requirements.

[0066] Finally, it should be noted that the above embodiments are merely examples for clearly illustrating the present invention and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A power grid dispatch speech recognition method based on multi-scale feature fusion, characterized in that, The evaluation method involves the following steps: Acquire multi-source voice data and power grid operation-related parameters in power grid dispatch scenarios, and dynamically adjust data time synchronization accuracy and acquisition resource allocation strategies according to power grid operation status; To address the noise characteristics of power grid dispatching scenarios, a preprocessing mechanism that uses noise and speech collaborative modeling is adopted to perform integrated processing on multi-source speech data and extract multi-scale speech features adapted to the dispatching scenario. Combining the representational characteristics of multi-scale speech features and the speech association rules of power grid dispatching terms, an extraction scheme of directional features and term association is adopted to extract the feature vectors corresponding to each scale feature, and then invalid features are removed after verification. A multi-dimensional fusion scheme is used to dynamically fuse feature vectors at various scales to obtain a fused feature vector. The fused feature vector is input into a preset speech recognition model, which integrates the power grid dispatching logic reasoning mechanism and combines power grid operation-related parameters to perform dual recognition of scenario adaptation and logic verification, and outputs speech recognition results. Perform cross-dimensional linkage verification on speech recognition results, dynamically adjust the verification standards based on the verification results, and determine whether the recognition results meet the preset requirements. If the recognition result meets the standard, the result is output and stored; otherwise, an adaptive adjustment mechanism is triggered to reprocess until the result is qualified or manual intervention is triggered.

2. The method according to claim 1, characterized in that, The following steps are taken to extract multi-scale speech features suitable for scheduling scenarios by using a preprocessing mechanism that combines noise and speech modeling to perform integrated processing on multi-source speech data: Noise in multi-source speech data is identified by type and intensity, distinguishing between electromagnetic interference noise from power grid equipment, environmental background noise, and redundant speech noise from personnel. An adaptive noise suppression mechanism is constructed by combining the noise feature library of power grid equipment. The noise feature library of power grid equipment contains the noise characteristics of various types of power grid equipment under different operating conditions and their correlation with speech distortion. Based on the noise type and intensity level, a graded and adaptive noise reduction process is used for noise separation and suppression; at the same time, a speech distortion compensation mechanism is used to correct the speech feature distortion caused by noise. Based on the priority of dispatch instructions and the power grid operation status, the noise-reduced voice data is subjected to adaptive frame segmentation processing, and the frame length and frame shift parameters of different duration frames are dynamically configured; among them, emergency instructions are configured with short duration frames, regular instructions are configured with medium duration frames, and maintenance instructions are configured with long duration frames. Feature extraction was performed on speech data of different duration frames to obtain short-time, medium-time, and long-time speech features. The three types of features were normalized to eliminate differences in dimensions and amplitude deviations. After feature validity verification, standardized multi-scale speech features were obtained.

3. The method according to claim 1, characterized in that, Combining the representational characteristics of multi-scale speech features and the speech association patterns of power grid dispatching terms, a scheme for extracting directional features and term associations is adopted. The steps for extracting the feature vectors corresponding to each scale feature are as follows: For short-term speech features, an extraction strategy combining temporal analysis, keyword template matching, and term pronunciation association is adopted to extract instantaneous speech features, which are then double-filtered to form a short-term feature vector. Among them, the keyword template contains standard speech feature information of core power grid dispatching operation terms, equipment names, and status description terms. For mid-time speech features, an extraction strategy combining frequency domain analysis, feature difference enhancement, and term and word combination association is adopted to extract speech spectrum features and construct mid-time feature vectors after enhancing the effective feature discrimination. For long-term speech features, an extraction strategy combining time-frequency joint analysis, semantic association filtering, and instruction structure matching is adopted to extract speech time-frequency features, which are then filtered by scheduling instruction semantic structure rules to form a long-term feature vector. The three types of feature vectors are subjected to dimension normalization and sparsification to remove invalid and redundant feature components; after feature discrimination and semantic relevance verification, the qualified feature vectors are obtained.

4. The method according to claim 1, characterized in that, The steps for dynamically fusing feature vectors at various scales using a multi-dimensional fusion scheme to obtain the fused feature vector are as follows: Construct a feature weight evaluation model that combines scheduling scenario type, power grid operation status and command priority, and calculate the initial weights of feature vectors at each scale; Based on the initial weights, an initial fused feature vector is obtained by combining weighted fusion and feature focusing mechanisms; and by dynamically adjusting the focusing range, the feature regions with high correlation to scheduling instructions are strengthened, and residual noise feature interference is suppressed. Semantic constraint rules from the power grid dispatching domain are introduced. Based on semantic relevance, feature validity, and instruction structure integrity, the initial fused feature vector is screened to remove redundant, invalid, and noise-residual features. The filtered feature vectors are subjected to cross-scale feature association enhancement and semantic calibration to strengthen the correlation of effective features at different scales, resulting in the final fused feature vector.

5. The method according to claim 1, characterized in that, The steps for inputting the fused feature vector into the preset speech recognition model, combining power grid operation-related parameters for scenario adaptation and logical verification, and outputting the speech recognition result are as follows: The preset speech recognition model is an adaptive recognition model driven by both scene and logic. It adopts a hybrid network structure that includes feature mapping, temporal enhancement and logical reasoning. It is trained and optimized by power grid dispatch speech samples, which cover different accents, noise scenes and fault scenes. The logical reasoning mechanism in the model integrates the power grid dispatching logical rule system, which is built based on power grid dispatching-related procedures and historical operations. It includes dispatching instruction terminology specifications, operational logic rules, equipment associations, and emergency response rules. The fused feature vectors are input into the model, and after feature mapping, temporal enhancement, and logical reasoning, feature category matching is performed. Historical dispatch instruction data is used for auxiliary matching, and the results that meet the confidence and logical fit standards are selected as the final speech recognition results. The recognition results include information such as dispatch instruction content, speech source, generation time, logical compliance identifier, scene matching degree and instruction type.

6. The method according to claim 1, characterized in that, The steps for cross-dimensional linkage verification of speech recognition results are as follows: Extract scheduling instruction keywords, semantic logic, priority identifiers, scene matching degree, and confidence level from the speech recognition results; Key information is compared with information in the power grid dispatch standard voice database in multiple dimensions to calculate the information matching degree, preset the accuracy threshold, and determine that results with insufficient confidence or logical fit need to be reviewed. Based on the power grid dispatching logic rule system, a multi-verification mechanism is adopted, which includes keyword combination, operation sequence, equipment association and logical conflict, to verify whether the identification results comply with the power grid dispatching safety specifications and operation logic, and to determine unqualified scenarios and corresponding handling solutions. Construct a matching model between operating status and instruction type, and perform multi-dimensional matching between the instruction information in the identification results and the power grid operation-related parameters to determine whether the instruction content, type and priority are suitable for the current power grid operating status. The system correlates the three verification results, dynamically adjusts the verification threshold and review strategy, and comprehensively determines whether the identification result is qualified. For unqualified results, relevant information is recorded and an optimization report is generated. For results that need to be reviewed, a manual verification process is triggered.

7. A power grid dispatching voice recognition system based on multi-scale feature fusion, characterized in that, The identification system includes: The multi-source data acquisition module is used to synchronously acquire multi-source voice data and power grid operation-related parameters in the power grid dispatching scenario, and to dynamically prioritize and synchronize the multi-source data. The scene-adaptive preprocessing module is used to perform integrated processing on multi-source speech data based on the noise characteristics of power grid dispatching scenarios, using a preprocessing mechanism that combines noise and speech modeling and a frame-segmentation strategy that adapts to instruction priority, and extracting standardized multi-scale speech features. The directional feature extraction module is used to extract feature vectors corresponding to multi-scale speech features by adopting an extraction scheme that associates directional features with terms, combined with scheduling term-related rules and templates. The multi-dimensional feature fusion module is used to dynamically fuse feature vectors at various scales using a multi-dimensional fusion scheme and output a fused feature vector. The dual-drive recognition module is used to input the fused feature vector into the preset speech recognition model, combine it with the power grid operation-related parameters to perform dual recognition for scenario adaptation and logic verification, and output the speech recognition result. The cross-dimensional verification module is used to perform cross-dimensional linkage verification of speech recognition results, dynamically adjust verification standards and review strategies, determine whether the recognition results are qualified, and output and store qualified recognition results when qualified; otherwise, it triggers the parameter adaptive adjustment mechanism, instructs relevant modules to reprocess, and records processing logs and optimization reports.

8. The system according to claim 7, characterized in that, The scene adaptive preprocessing module includes: The multimodal noise processing unit is used to identify noise type and intensity level, construct an adaptive noise suppression model, and perform precise noise processing and speech distortion correction by adopting an integrated processing scheme of noise identification, separation and distortion compensation. The priority framing unit is used to perform adaptive framing processing on the noise-reduced voice data based on the priority of scheduling instructions and the power grid operating status, and dynamically configure the framing parameters. The multi-scale feature extraction unit includes short-time, medium-time and long-time feature extraction sub-units, which extract features from speech data of different duration frames to obtain corresponding multi-scale speech features. The feature normalization unit is used to normalize multi-scale speech features, eliminate dimensional differences and amplitude deviations, and output standardized multi-scale speech features after feature validity and semantic relevance verification; it also triggers a re-extraction process for abnormal features.

9. The system according to claim 7, characterized in that, The targeted feature extraction module includes: The short-time vector extraction unit is used to extract feature vectors corresponding to short-time speech features by employing a strategy that combines temporal analysis, keyword template matching, and term pronunciation association. The mid-time vector extraction unit is used to extract feature vectors corresponding to mid-time speech features by employing a strategy that combines frequency domain analysis, feature difference enhancement, and term-lexical combination association. The long-term vector extraction unit is used to extract feature vectors corresponding to long-term speech features by employing a strategy that combines time-frequency joint analysis, semantic association filtering, and instruction structure matching. The feature normalization unit is used to normalize and sparsify the three types of feature vectors, remove invalid and redundant features, and perform dimension verification, feature discrimination and semantic relevance detection. Features that do not meet the criteria are triggered to re-extract.

10. The system according to claim 7, characterized in that, The multi-dimensional feature fusion module includes: The weight calculation unit is used to construct a feature weight evaluation model, and calculates the initial weights of feature vectors at each scale by combining scheduling scenarios, power grid operation status and command priority. The feature-focusing fusion unit is used to obtain an initial fused feature vector based on the initial weights by combining weighted fusion and feature-focusing mechanisms. The semantic filtering unit is used to filter the initial fused feature vector based on the semantic constraint rules of power grid dispatch, and remove redundant, invalid and noise residual features. The feature enhancement unit is used to perform cross-scale association enhancement and semantic calibration on the selected feature vectors, adjust the stability of feature representation and scene adaptability, and perform quality evaluation of fused features. If a feature fails to meet the standard, the weight readjustment process is triggered.

Images