A speech data interactive feedback control processing method based on large language model

Abstract

The present invention relates to the field of large language model technology and discloses a method for interactive feedback control and processing of voice data based on a large language model. The method comprises: collecting industrial voice commands through an AMR master controller, obtaining standardized vectors through industrial vocabulary matching and intent classification; inputting the industrial large language model for inference processing to generate an AMR execution plan; performing distributed coordination and task allocation on the AMR cluster to form a control command sequence; and the motion controller executing monitoring and handling exceptions, outputting a feedback strategy. This invention achieves accurate classification and semantic understanding of complex industrial commands, improves the ability to process industrial domain expertise, and simultaneously implements high-precision real-time status monitoring.

Description

Technical Field

[0001] The present invention relates to the technical field of large language models, and in particular to a speech data interactive feedback control processing method based on a large language model. Background Art

[0002] Current industrial voice interaction technologies generally suffer from poor domain adaptability. General-purpose speech recognition models cannot accurately understand industrial terminology and complex instructions, and lack a deep understanding of industrial production processes and equipment operating specifications. Traditional task allocation algorithms struggle to achieve global optimization in multi-machine coordination scenarios, prone to resource conflicts and low execution efficiency. Furthermore, existing AMR systems rely primarily on preset rules for exception handling, lacking intelligent analysis and adaptive adjustment capabilities. This makes it impossible to dynamically optimize control strategies based on actual operating conditions, making them difficult to cope with the complex and ever-changing industrial production environment. Summary of the Invention

[0003] The present invention provides a voice data interactive feedback control processing method based on a large language model. The present invention realizes the accurate classification and semantic understanding of complex industrial instructions, improves the processing ability of professional knowledge in the industrial field, and realizes high-precision real-time status monitoring.

[0004] In a first aspect, the present invention provides a method for interactive feedback control and processing of speech data based on a large language model, the method comprising:

[0005] Collecting industrial voice command data through the AMR master controller, and performing industrial vocabulary matching and voice intent classification on the industrial voice command data to obtain a standardized voice intent vector;

[0006] Inputting the standardized speech intention vector into the industrial Internet large language model for industrial domain expertise reasoning and context perception processing to obtain an AMR execution plan;

[0007] Performing distributed consistency coordination and task allocation on the AMR cluster network based on the AMR execution scheme to obtain a distributed control instruction sequence;

[0008] The distributed control instruction sequence is input into the motion controller of the AMR master controller for multi-threaded execution monitoring and exception processing to obtain an exception feedback processing strategy.

[0009] Optionally, in a first implementation of the first aspect of the present invention, collecting industrial voice command data through an AMR master controller, and performing industrial vocabulary matching and voice intent classification on the industrial voice command data to obtain a standardized voice intent vector includes:

[0010] Collecting industrial voice command data through a voice data acquisition module in the AMR main controller, and inputting the industrial voice command data into an audio digital signal processor for voice activity detection to obtain voice feature data;

[0011] Synchronously reading data and extracting parameters from the lidar sensor, inertial measurement unit sensor, and vision sensor of the AMR master controller to obtain a multi-source sensor data set;

[0012] Based on the reception time of the voice feature data, the multi-source sensor data set is timestamped and synchronized to obtain a spatiotemporal synchronized data group;

[0013] Associating, binding, and encoding the speech feature data with the spatiotemporal synchronization data group to obtain multi-dimensional fused speech data;

[0014] The multi-dimensional fused voice data is input into the edge reasoning module of the AMR main controller for industrial vocabulary matching and voice intent classification to obtain a standardized voice intent vector.

[0015] Optionally, in a second implementation of the first aspect of the present invention, the step of inputting the multi-dimensional fused voice data into an edge inference module of an AMR controller for industrial vocabulary matching and voice intent classification to obtain a standardized voice intent vector includes:

[0016] Inputting the multi-dimensional fusion voice data into the edge reasoning module in the AMR main controller to extract voice feature vectors and perform data format standardization processing to obtain voice feature vectors;

[0017] Inputting the speech feature vector into a hybrid neural network architecture for feature learning and speech pattern recognition to obtain an intermediate representation vector;

[0018] Performing vocabulary matching and semantic similarity calculation on a pre-built industrial voice command classification vocabulary based on the intermediate representation vector to obtain vocabulary matching data;

[0019] Speech intent classification calculation and confidence evaluation are performed based on the vocabulary matching data to obtain a standardized speech intent vector.

[0020] Optionally, in a third implementation of the first aspect of the present invention, inputting the speech feature vector into a hybrid neural network architecture for feature learning and speech pattern recognition to obtain an intermediate representation vector includes:

[0021] Inputting the speech feature vector into a convolutional neural network in a hybrid neural network architecture to perform multi-scale feature extraction and local speech pattern recognition to obtain convolution feature data;

[0022] Inputting the convolution feature data into the long short-term memory network in the hybrid neural network architecture to perform temporal dependency modeling and speech context semantic learning to obtain sequence representation data;

[0023] Based on the sequence representation data, linear transformation and nonlinear mapping are performed to obtain a high-level semantic feature vector, and attention weight analysis is performed on the high-level semantic feature vector to obtain an intermediate representation vector.

[0024] Optionally, in a fourth implementation of the first aspect of the present invention, inputting the standardized speech intention vector into the industrial Internet large language model for industrial domain expertise reasoning and context-aware processing to obtain an AMR execution solution includes:

[0025] Inputting the standardized speech intention vector into the Transformer encoder of the industrial Internet large language model for mapping processing to obtain industrial speech intention encoding data;

[0026] Based on the industrial speech intention coding data, professional terminology query and semantic relationship analysis are performed on the pre-built industrial knowledge graph to obtain industrial domain professional knowledge reasoning results;

[0027] The industrial domain expertise reasoning results are integrated with the current production line status data, AMR cluster location distribution information, and task queue status information through attention mechanism fusion and context perception to obtain comprehensive semantic reasoning data;

[0028] The comprehensive semantic reasoning data is input into the Transformer decoder of the industrial Internet large language model to generate task decomposition sequences, construct parameter configuration tables, and sort execution priorities to obtain the AMR execution plan.

[0029] Optionally, in a fifth implementation of the first aspect of the present invention, inputting the standardized speech intention vector into the Transformer encoder of the industrial Internet large language model for mapping processing to obtain industrial speech intention encoded data includes:

[0030] Performing dimension expansion and numerical normalization processing on the standardized speech intention vector to obtain vector dimension conversion data;

[0031] Inputting the vector dimension conversion data into the input embedding layer of the Transformer encoder for linear transformation and weight matrix mapping to obtain input embedding vector data;

[0032] Sine-cosine position coding calculation and vector fusion are performed based on the sequence position of the input embedded vector data to obtain a position coding fusion vector, and multi-head attention calculation is performed on the position coding fusion vector to obtain industrial speech intention coding data.

[0033] Optionally, in a sixth implementation of the first aspect of the present invention, performing distributed consistency coordination and task allocation on the AMR cluster network based on the AMR execution scheme to obtain a distributed control instruction sequence includes:

[0034] The task decomposition operation sequence in the AMR execution scheme is split into task units and the workload is evaluated. At the same time, the status data of each AMR master controller in the AMR cluster network is synchronized and information is updated through industrial Ethernet to obtain the AMR cluster global status table;

[0035] Performing load balancing processing on the task decomposition operation sequence based on the AMR cluster global state table to obtain a set of task allocation candidate solutions;

[0036] Input the set of task allocation candidate solutions into a distributed consensus algorithm for multi-objective optimization solution and feasibility verification to obtain the optimal task allocation solution;

[0037] A control instruction sequence is constructed and an urgent task priority arbitration process is performed according to the optimal task allocation scheme to obtain a distributed control instruction sequence.

[0038] Optionally, in a seventh implementation of the first aspect of the present invention, inputting the set of candidate task allocation solutions into a distributed consensus algorithm for multi-objective optimization solution and feasibility verification to obtain an optimal task allocation solution includes:

[0039] Performing total execution time calculation, energy consumption evaluation, and path conflict probability statistics on each candidate task allocation scheme in the set of task allocation candidate schemes to obtain a multi-objective function numerical matrix;

[0040] Based on the multi-objective function numerical matrix, the set of task allocation candidate solutions is subjected to load limit boundary check, power consumption threshold verification and safety distance interval verification processing to obtain a subset of feasible candidate solutions;

[0041] The subset of feasible candidate solutions is input into a distributed consensus algorithm for optimization and solution to obtain an optimal task allocation solution.

[0042] Optionally, in an eighth implementation of the first aspect of the present invention, inputting the distributed control instruction sequence into the motion controller of the AMR master controller for multi-threaded execution monitoring and exception handling to obtain an exception feedback handling strategy includes:

[0043] The distributed control instruction sequence is input into the main execution thread of the motion controller in the AMR main controller for instruction parsing and hardware drive control, and the monitoring thread is started to collect status data and manage cache to obtain hardware control instructions and real-time status monitoring data stream;

[0044] Based on the real-time status monitoring data stream, encoder position feedback, force sensor load feedback, and tilt sensor posture feedback are periodically sampled and time series data cached to obtain an AMR execution status data packet;

[0045] Performing position error threshold detection, speed deviation threshold detection, and load change anomaly detection based on the AMR execution status data packet to obtain an execution anomaly detection result;

[0046] According to the execution anomaly detection result, data is uploaded and analyzed by the industrial Internet large language model through industrial Ethernet to obtain an abnormal feedback processing strategy.

[0047] Optionally, in a ninth implementation of the first aspect of the present invention, the distributed control instruction sequence is input into the main execution thread of the motion controller in the AMR master controller for instruction parsing and hardware driver control, and a monitoring thread is started to perform status data collection and cache management to obtain hardware control instructions and real-time status monitoring data streams, including:

[0048] Inputting the distributed control instruction sequence into the motion controller of the AMR master controller for instruction format parsing and thread resource allocation, thereby obtaining a multi-thread execution scheduling scheme including a main execution thread task queue and a monitoring thread task queue;

[0049] Based on the multi-threaded execution scheduling scheme, the action type instruction and execution parameter values ​​in the main execution thread task queue are input into the servo motor driver, the steering mechanism controller, and the lifting mechanism operator to generate pulse width modulation signals and perform power amplification processing to obtain hardware control instructions;

[0050] According to the multi-threaded execution scheduling scheme, the monitoring thread is started to perform high-frequency data reading and parameter conversion on the encoder sensor, inertial measurement unit sensor, and tilt angle sensor of the AMR main controller to obtain a real-time status monitoring data stream.

[0051] The technical solution provided by this invention achieves precise association between voice commands and environmental states through the spatiotemporal synchronization of a digital microphone array, lidar, inertial measurement unit, and visual sensor, significantly improving the accuracy of command understanding and environmental adaptability. Synchronous alignment processing using a unified time base solves the problem of time inconsistency in multi-source sensor data, ensuring the real-time and reliability of control decisions. Based on a lightweight industrial speech recognition neural network and hybrid neural network architecture, high-precision recognition of industrial terminology is achieved, effectively solving the low accuracy problem of traditional speech recognition in industrial environments. Pre-built industrial speech command classification vocabulary matching enables accurate classification and semantic understanding of complex industrial commands, improving the system's ability to process industrial domain expertise. The Transformer architecture of encoders and decoders enables deep reasoning and context-aware processing of industrial domain expertise, significantly enhancing the intelligent level of task planning. Through terminology query and semantic relationship analysis within the industrial knowledge graph, the integrated application of multi-dimensional knowledge, such as manufacturing processes, equipment control, and quality management, is achieved, improving the professionalism and feasibility of the execution plan. By improving the distributed consensus algorithm to perform Pareto front search and non-dominated sorting, a multi-objective global optimization of total execution time, energy consumption, and conflict probability is achieved, effectively solving the local optimality problem of traditional task allocation algorithms. Feasibility verification based on load restrictions, power constraints, and safe distance constraints ensures the safety and executability of the task allocation scheme and improves the stability of AMR cluster coordination. The parallel processing architecture of the main execution thread and the monitoring thread enables simultaneous instruction execution and status monitoring, improving the system's real-time responsiveness and execution efficiency. High-frequency sampling and circular storage in a ring buffer enable high-precision real-time status monitoring. Intelligent anomaly analysis and processing based on the large language model of the industrial Internet enables the automatic generation of adaptive control strategies such as trajectory correction, speed adjustment, and path replanning, significantly improving the level of intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0053] Figure 1 The figure is a schematic diagram of an embodiment of a method for interactive feedback control of speech data based on a large language model in an embodiment of the present invention. DETAILED DESCRIPTION

[0054] An embodiment of the present invention provides a method for interactive feedback control and processing of speech data based on a large language model. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0055] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 In one embodiment of the present invention, a method for interactive feedback control of speech data based on a large language model includes:

[0056] Step S101: Collect industrial voice command data through the AMR master controller, perform industrial vocabulary matching and voice intent classification on the industrial voice command data, and obtain a standardized voice intent vector;

[0057] It is understandable that the execution subject of the present invention can be a voice data interactive feedback control processing system based on a large language model, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking the server as the execution subject as an example.

[0058] Specifically, the AMR controller integrates a voice data acquisition module, which includes a high-precision digital microphone array, audio signal acquisition circuitry, and a dedicated digital signal processor (DSP). This module captures and preprocesses voice commands issued by operators in industrial environments with high fidelity. The raw audio signal is then fed into the DSP in real time for voice activity detection. This process uses energy threshold variations and time-frequency feature boundary extraction to identify the start and end boundaries of the voice command. It then performs preliminary signal segmentation and voice feature extraction to obtain voice feature data representing the command. Simultaneously, the AMR controller activates its lidar sensor, inertial measurement unit, and visual perception module to synchronously sense and read data from the industrial site environment. This module obtains a multi-source sensor data set, including current position coordinates, heading angle, velocity and acceleration, obstacle distance, and image frame features. Using the moment the voice feature data is received as a unified time anchor, each item in the multi-source sensor data set is timestamped with high precision, and a unified synchronization alignment mechanism is implemented to ensure consistency between the moment of speech and the state of the environment. Through data structure mapping and channel encoding, voice feature data is associated and bound to the corresponding sensor dataset at that moment, forming a spatiotemporally synchronized data set. This data is then encoded into multidimensional fused voice data in a unified format. This multidimensional fused voice data is then fed into the edge inference module deployed within the AMR controller. This module consists of a lightweight neural network model consisting of a cascade of six convolutional units and two long short-term memory (LSTM) layers, offering low latency and local deployment. This module is preloaded with an industrial voice vocabulary containing 2048 industry-specific terms, 512 AMR control actions, and 256 cluster coordination commands. Using feature matching and semantic similarity calculations, it rapidly performs keyword recognition, semantic clustering, and intent classification on the input multidimensional voice data. It ultimately outputs a standardized voice intent vector consisting of four fields: command type, target object code, operation parameter value, and confidence score, expressed in 32-bit floating-point format.

[0059] Step S102: Input the standardized speech intent vector into the Industrial Internet Big Language Model to perform industrial domain expertise reasoning and context perception processing to obtain an AMR execution plan;

[0060] Specifically, a standardized speech intent vector is input into the large language model of the industrial internet platform, and semantic mapping and embedding of the vector are performed through the encoder portion of its Transformer architecture. The standardized speech intent vector, which includes fields such as command type, target object, operation parameters, and confidence, is mapped into a high-dimensional semantic representation after input into the encoder. Its internal structure is reconstructed through a multi-head attention mechanism and a feedforward network, generating industrial speech intent encoding data that captures industrial semantic associations and control intent. After obtaining the encoding representation, the large language model then invokes its pre-trained industrial knowledge graph module. This graph encompasses over 30,000 pieces of industrial process data, 50,000 pieces of equipment operating manuals, and 80,000 fault diagnosis cases, forming a complex knowledge network composed of semantic nodes and relationship edges. Leveraging the keywords and semantic patterns contained in the speech intent encoding data, the graph is subjected to specialized terminology query and semantic path expansion analysis, exploring the device control logic, operation sequence, safety specifications, and fault tolerance strategies associated with the input command. Logical connections are established between terms based on semantic nesting and hierarchical relationships, resulting in targeted and specialized inference results based on industrial domain expertise. The above inference results are fused with real-time industrial field operation data. This includes current production line status information such as workstation occupancy, task completion rate, and work-in-process progress; spatial distribution data of the AMR cluster, such as each robot's location coordinates, power level, and availability; and queue status data from the task scheduling system, such as the priority of waiting tasks, expected execution time, and resource conflicts. By inputting dynamic context information and knowledge reasoning results into the attention mechanism unit of the Transformer architecture, the large language model can establish a cross-domain context-aware model, automatically learning which production conditions, resource status, and knowledge relationships are most critical for current task execution. Through weighted attention fusion, it generates comprehensive semantic reasoning data that is task-adaptive, environmentally relevant, and semantically consistent, ultimately forming a deep control semantic representation that fully understands the context and is execution-oriented. The comprehensive semantic reasoning data is then fed into the decoder module of the Transformer architecture, where a step-by-step decoding strategy is used to generate a task decomposition sequence, construct a parameter configuration table, and prioritize execution. The task decomposition sequence explicitly converts complex voice intent into a set of executable basic action units. The parameter configuration table assigns specific execution parameters to each action, such as target coordinates, movement speed, steering angle, and safety margin. Prioritization ensures that tasks are executed in an orderly manner within the multi-task queue based on urgency and resource scheduling. The AMR execution plan is output in a structured JSON format.

[0061] Step S103: performing distributed consistency coordination and task allocation on the AMR cluster network based on the AMR execution scheme to obtain a distributed control instruction sequence;

[0062] Specifically, a task-decomposition operation sequence is extracted from the AMR execution plan generated by the large language model. This sequence, encoded as a logical action chain or structured JSON format, indicates the underlying physical action unit corresponding to each voice command. This operation sequence is then refined, and a task unit splitting strategy is used to map high-level semantic instructions into fine-grained control task units. A workload assessment mechanism is also introduced to quantitatively evaluate each task unit based on factors such as the execution time required for the action, path length, energy budget, and operation complexity, resulting in a clearly structured set of task subunits. Furthermore, through a high-speed communication interface deployed on Industrial Ethernet, status data of all master controllers in the AMR cluster network is synchronized and updated in real time. Multi-dimensional parameters such as the current position, power level, load status, execution progress, and fault flags of each AMR are collected and structured, integrating these state parameters to dynamically construct a global state table for the AMR cluster. Based on the AMR cluster global state table, load balancing is performed on the task-decomposition operation sequence. During the processing process, the scheduling module compares the resource requirements of each task unit with the available capacity of the AMR node, taking into account factors such as path conflict probability, remaining battery power, and the current task queue length. This module then makes a preliminary task allocation, forming a set of candidate solutions, each representing a different set of AMR task assignment paths. To improve allocation quality and system stability, the candidate task allocation solutions are input into the distributed consensus algorithm module, which utilizes a modified Paxos or Raft consensus framework and integrates a multi-objective optimization solver. This module uses a cost function minimization mechanism to perform a weighted balance between execution time, energy consumption, path overlap, and task dependency constraints. It then performs feasibility verification to eliminate conflicting or overloaded configurations and select the optimal task allocation solution that balances performance, security, and execution efficiency. Based on this optimal task allocation plan, the scheduling system initiates a control instruction sequence construction program, mapping the tasks of each AMR node in the plan into a standardized control instruction sequence. Each instruction includes the target AMR number, action type, target location, execution speed, task dependencies, and a scheduling timestamp. Simultaneously, based on the system's emergency task response strategy, priority arbitration is performed on the task sequence, allowing high-priority tasks with emergency markings to interrupt or reschedule currently executing non-critical tasks, thereby ensuring real-time response and system flexibility. The resulting distributed control instruction sequence is structurally encapsulated and distributed to each AMR master controller via Industrial Ethernet, achieving synchronous instruction deployment and efficient closed-loop cluster scheduling.

[0063] Step S104: input the distributed control instruction sequence into the motion controller of the AMR master controller for multi-threaded execution monitoring and exception handling to obtain an exception feedback handling strategy.

[0064] Specifically, a distributed control instruction sequence is sent to each AMR master controller and connected to the master controller's internal motion controller. The controller's main execution thread performs real-time parsing and parameter decoding of the received control instructions, identifying key information such as the target action type, execution path, speed parameters, and target coordinates required by each instruction. These decoded results are converted into drive signals and sent via an interface protocol to the AMR's internal servo motor, steering mechanism, and lift actuator, completing the issuance of the underlying hardware control commands. Simultaneously, the motion controller launches a monitoring thread in parallel. This thread, as the data processing submodule in the master controller responsible for operational status monitoring, samples the AMR's various sensor status channels in real time, including encoder feedback channels, force sensor feedback channels, and attitude and tilt sensor channels. The sampling results are packaged into a structured data cache stream at a certain frequency (e.g., every 10 milliseconds), enabling continuous collection and dynamic recording of instruction execution status. This forms a closed control execution loop consisting of a parallel hardware control instruction stream and a real-time status monitoring data stream. During the acquisition phase, the monitoring thread periodically samples wheel position feedback data from the encoders, load response values ​​from the force sensors, and pitch and roll angle information from the tilt sensors. The sampling points are arranged by timestamp, forming a highly continuous time series data set suitable for anomaly modeling. This data is then packaged into an AMR execution status data packet. This packet includes key indicators such as the current coordinates, velocity vector, motor torque, tilt angle data, payload mass, and power status. It is uploaded in a fixed format to the anomaly detection module on the main controller. This module internally maintains a set of error judgment rules and threshold parameters, such as a position error upper limit of ±5cm, a velocity deviation upper limit of ±10%, and an anomaly if the load sudden change exceeds 20%. After parsing the execution status data packet, the system performs error detection operations one by one. This involves comparing the actual encoder position with the target command position to calculate the position error, calculating the velocity deviation by the difference between the current velocity and the command velocity, and then determining the load sudden change based on the rate of change of the load value between the previous cycle and the current cycle. This generates an execution anomaly detection result, which includes structured output information such as the anomaly type, occurrence time, deviation magnitude, and impact level. Anomaly detection results are reported to the Industrial Internet platform in real time via Industrial Ethernet. A large language model deployed in the cloud performs in-depth semantic analysis and policy generation. Based on a large number of fault cases and device behavior patterns acquired through previous training, the large language model semantically categorizes anomaly types and their causes and conducts causal chain reasoning. It determines whether the current anomaly stems from environmental disturbances, hardware fatigue, path conflicts, task parameter mismatches, or system scheduling imbalances. It then incorporates contextual information (such as upstream and downstream task dependencies, the AMR's physical location, and the execution time period) to formulate anomaly feedback and handling strategies.

[0065] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0066] The industrial voice command data is collected through the voice data acquisition module in the AMR main controller, and the industrial voice command data is input into the audio digital signal processor for voice activity detection to obtain voice feature data;

[0067] Synchronously read data and extract parameters from the AMR master controller's lidar sensor, inertial measurement unit sensor, and vision sensor to obtain a multi-source sensor data set;

[0068] Based on the reception time of the voice feature data, the multi-source sensor data set is timestamped and synchronized to obtain a spatiotemporal synchronized data set;

[0069] Associating, binding, and encoding the speech feature data with the spatiotemporal synchronization data group to obtain multi-dimensional fused speech data;

[0070] The multi-dimensional fused voice data is input into the edge inference module of the AMR main controller for industrial vocabulary matching and voice intent classification to obtain a standardized voice intent vector.

[0071] Specifically, a high-precision voice data acquisition module is integrated into the AMR main controller. This module's internal structure consists of an array digital microphone, an audio front-end signal amplifier, a high-speed sampling ADC component, and a low-power audio DSP chip. This module continuously monitors the voice signal channel in the industrial field and enhances the spatial directionality of external sound sources through a multi-channel microphone array to improve voice capture capabilities in noisy production environments. When a voice signal is detected, the module initiates the audio signal preprocessing process. The front-end circuit performs high-fidelity amplification and filtering on the raw analog voice, then converts it into a digital audio stream through an analog-to-digital converter. This audio stream is then transmitted to the DSP core for real-time processing. After receiving the audio stream, the audio DSP chip uses an adaptive filtering algorithm to suppress background noise and cancel echoes. It also uses a VAD (Voice Activity Detection) algorithm to identify the start and end boundaries of speech segments, eliminating non-speech segments and redundant ambient noise, thereby extracting continuous and pure speech command intervals. The DSP calculates feature parameters for each speech interval, extracting audio descriptors in multiple dimensions, such as MFCC (Mel-Frequency Cepstral Coefficients), pitch contour, energy envelope, and zero-crossing rate. These features are then packaged according to a unified coding standard to form a speech feature data vector, which represents the basic semantic structure of the speech segment in the time and frequency domains. Simultaneously, the system simultaneously activates the various external perception sensors configured in the AMR master controller to synchronously collect the robot's physical environment and motion posture parameters. The LiDAR module collects 2D or 3D environmental point cloud data, providing distance distribution information and obstacle outline recognition. The inertial measurement unit (IMU) uses a three-axis accelerometer and gyroscope to provide the AMR's linear acceleration, angular velocity, and spatial attitude change data. The visual sensor, based on a CMOS image acquisition array, captures image frames of the scene in front of the AMR, extracting visual features such as environmental texture, color distribution, object boundaries, and landmark patterns. After collection, data from all these sensing channels is transmitted to the AMR controller's multi-source perception fusion interface. After format standardization, numerical normalization, and frame rate unification, a consistent multi-source sensor data set is formed, forming a comprehensive physical information system describing the AMR's local environment and motion state. To ensure temporal consistency between the voice signal and external sensor data, the multi-source sensor data set is timestamped, using the reception time of the voice feature data as a synchronization anchor. A unified system clock is established in the main controller. Through a high-precision clock synchronization mechanism, the corresponding reception time of each data from the radar, IMU or visual channel is recorded, and this time is aligned and compared with the start time of the voice feature data. A sliding window matching algorithm is used to search for sensor data that is simultaneous with the voice command in multiple time segments, forming a set of sensor states that are highly consistent with the time of the voice segment.On this basis, the synchronized sensor data from various types is merged to generate a unified spatiotemporal synchronized data set. This data set contains information such as the precise spatial position, velocity vector, attitude angle, environmental point cloud, and image frame of the AMR at the moment of speech. This speech feature data is deeply fused with the spatiotemporal synchronized data set. Through an association binding mechanism, the speech vector is used as the primary key to match the multi-source sensor data entry at the time of speech occurrence. These data are then jointly encoded in a fixed format. Following a unified data structure design specification, the system concatenates the 128-byte speech feature vector, the 64-byte position state vector, and the 32-byte environmental perception vector into a 224-byte multidimensional fused speech data packet. Compressed sensing algorithms and quantization coding techniques are used to reduce redundant dimensions and optimize the dimensionality, forming a compact fused data structure required for input into the edge inference module. This multidimensional fused speech data is then fed into the edge inference module deployed in the AMR controller for semantic analysis. This module is a low-power, highly responsive, and lightweight deep neural network system that utilizes a hybrid neural network architecture consisting of six convolutional layers and two LSTM layers and embeds an industry-specific speech vocabulary optimized with domain knowledge. The system extracts representative local features from speech fusion data through convolutional layers, and then uses an LSTM structure to capture temporal dependencies within speech segments, completing the feature learning process. The resulting intermediate feature representation is matched against a local semantic lexicon consisting of 2,048 industrial terms, 512 motion control keywords, and 256 cluster scheduling commands. The system performs keyword recognition and semantic clustering by calculating cosine similarity and semantic distance metrics between word vectors. The parsed results are ultimately output in a structured form as a standardized speech intent vector. This vector contains four fields: instruction type identifier, target object code, operation parameter value, and confidence score, and is uniformly encoded in 32-bit floating-point format.

[0072] In a specific embodiment, the step of inputting the multi-dimensional fused speech data into the edge inference module of the AMR master controller for industrial vocabulary matching and speech intent classification to obtain a standardized speech intent vector may specifically include the following steps:

[0073] The multi-dimensional fused voice data is input into the edge inference module of the AMR main controller to extract the voice feature vector and standardize the data format to obtain the voice feature vector;

[0074] The speech feature vector is input into the hybrid neural network architecture for feature learning and speech pattern recognition to obtain an intermediate representation vector;

[0075] Based on the intermediate representation vector, the pre-built industrial voice command classification vocabulary is matched and semantic similarity is calculated to obtain vocabulary matching data;

[0076] Speech intent classification calculation and confidence evaluation are performed based on the vocabulary matching data to obtain a standardized speech intent vector.

[0077] Specifically, the multi-dimensional fused speech data is input into the edge inference module of the AMR main controller. This module has data preprocessing capabilities for speech perception and an embedded efficient semantic understanding structure. Its processing chain has a complete logical flow, from raw data formatting and feature compression to deep semantic modeling. In this process flow, the preprocessing unit receives the fused speech data packet, which contains speech feature content, AMR spatial state information, and external perception environment parameters. To ensure that the model can focus on the semantic expression, the speech data is decoded and extracted from the structured fusion packet through a semantic separation mechanism. The speech feature vector is then extracted. Specifically, typical acoustic features such as MFCC parameters, zero-crossing rate, short-time energy, and spectral centroid are extracted. These features are obtained through frequency domain analysis and window function processing, then uniformly normalized to a fixed dimension and normalized to eliminate feature scale inconsistencies caused by different sound sources or channels. This step results in a speech feature vector. The speech feature vector is fed into a hybrid neural network architecture deployed in the edge inference module for deep feature learning and speech pattern recognition. This neural network architecture utilizes a composite structure consisting of a six-layer convolutional neural network (CNN) and a two-layer long short-term memory (LSTM) neural network in series. The first six convolutional layers extract local pattern features from the speech signal at multiple scales, such as boundary changes, plosives, high-frequency resonances, and weak pitch changes. Convolutional kernels with varying receptive field sizes identify the combination of short-term and long-term variations, while simultaneously compressing the spatial hierarchy and enhancing generalization capabilities. The second two LSTM layers perform temporal modeling on the high-dimensional feature tensor extracted by the CNN. Leveraging its hidden state memory mechanism, they extract contextual dependencies and semantic continuity from the speech signal. This is particularly important for recognizing semantic continuity despite factors such as speech rate variations, pause insertions, and intonation fluctuations. This multi-layer network outputs an intermediate representation vector that combines the time series evolution of speech with semantic cues. The intermediate representation vector is input into the industrial voice command classification vocabulary matching subsystem deployed in conjunction with the edge inference module. The vocabulary system consists of three types of special instructions, including 2048 industrial control terms, 512 AMR motion command keywords, and 256 AMR cluster coordination and scheduling commands. Each entry has a corresponding word vector space pre-built through a semantic vectorization model (such as Word2Vec or BERT embedding).The intermediate representation vector is matched against each command type in the aforementioned word vector space. The matching process is based on a semantic similarity metric, such as cosine similarity, which measures similarity by calculating the cosine of the angle between the intermediate vector and each term vector. A value closer to 1 indicates greater semantic consistency. The system also introduces a distance constraint threshold and a semantic distribution mapping function to prevent mismatches and semantic drift while maintaining recognition sensitivity. The system then outputs vocabulary matching data containing the highest matching value, candidate terms, command categories, and confidence indicators. Voice intent classification and confidence assessment are performed based on the vocabulary matching data. The system utilizes the candidate command types and matching confidence information provided in the vocabulary matching data, combined with contextual features of the task environment, to make classification voting decisions. A set of soft classifiers calculates the probability distribution of each candidate intent and determines the optimal intent type label. Building on this foundation, the system introduces a confidence assessment mechanism that comprehensively assesses the confidence level of the current intent classification result by combining voice signal strength stability, matching result aggregation trends, and historical command distribution probabilities. This generates a numerical confidence score, which serves as the weighting for the system's response to the voice command. The edge inference module outputs the results of this round of speech analysis as a standardized speech intent vector. The vector structure consists of four core fields: instruction type identifier, target object encoding, operation parameter value and confidence score. Each field is encoded in 32-bit floating point format.

[0078] In a specific embodiment, the step of inputting the speech feature vector into the hybrid neural network architecture for feature learning and speech pattern recognition to obtain the intermediate representation vector may specifically include the following steps:

[0079] The speech feature vector is input into the convolutional neural network in the hybrid neural network architecture to perform multi-scale feature extraction and local speech pattern recognition to obtain convolution feature data;

[0080] The convolution feature data is input into the long short-term memory network in the hybrid neural network architecture to perform temporal dependency modeling and speech context semantic learning to obtain sequence representation data;

[0081] Based on the sequence representation data, linear transformation and nonlinear mapping are performed to obtain a high-level semantic feature vector, and attention weight analysis is performed on the high-level semantic feature vector to obtain an intermediate representation vector.

[0082] Specifically, the speech feature vector is input into a hybrid neural network structure deployed in the AMR main controller. This structure consists of a front-end convolutional neural network (CNN) module and a back-end long short-term memory (LSTM) module. Its overall architecture is designed to address the dual needs of modeling the local aggregation of spatial patterns and temporal evolutionary structures in speech signals. During this process, the CNN receives the speech feature vector output by the edge speech processing module. This feature vector includes physically interpretable acoustic features such as multidimensional Mel-Frequency Cepstral Coefficients (MFCCs), speech energy envelope, zero-crossing rate, and spectral centroid. It is organized into a two-dimensional time-frequency matrix to match the CNN input requirements. The system designs a multi-scale convolution kernel structure for this convolution module, using receptive fields of varying sizes to simultaneously capture local feature variations at different time scales in the speech signal, such as rapidly changing plosives in the high-frequency band, resonant structures in the low-frequency region, and structural transitions between pauses and speech flow. By stacking multiple convolutional layers, applying nonlinear transformations to activation functions, and performing dimensionality reduction compression using the max pooling operation, convolutional neural networks gradually extract highly recognizable convolutional feature data from input features. This convolutional feature data is then fed into the long short-term memory (LSTM) network within the hybrid neural network architecture to model the temporal contextual dependencies of speech signals. As a special type of recurrent neural network, the LSTM network's internal structure includes a forget gate, an input gate, and an output gate. These gating mechanisms dynamically adjust the updating and retention of the unit state vector, effectively capturing long-term dependencies and avoiding the long-term information loss problem caused by vanishing gradients in standard RNNs. In speech semantic modeling tasks, LSTMs can effectively identify temporal logical relationships between speech segments, such as the spacing between verbs and objects, the continuation of intonation in control sentences, and the time-domain propagation of intent semantics within the context. In its implementation, convolutional feature data is fed into each time step of the LSTM in a time-slice sequence. The forward propagation mechanism continuously updates the hidden and memory states, while outputting a temporary semantic state encoding vector at each step. After the entire speech segment undergoes temporal modeling processing within the LSTM structure, its output is serialized into a set of time-stamped semantic representation data. These data, with time as the primary axis, record the semantic evolution path from the start to the end of the speech. High-level semantic features are extracted from this sequence representation data. A linear transformation structure is introduced, mapping the high-dimensional time-step feature vectors to a unified high-order semantic space by setting a weight matrix, thereby achieving the dual goals of dimensionality increase and feature compression. A nonlinear mapping mechanism is then introduced, using nonlinear activation functions such as ReLU, Tanh, or Swish to enhance the separability of semantic boundaries and the expressive power of the model. This results in a high-level semantic feature vector that preserves the temporal information and spatial patterns in the speech signal while forming a more discriminative deep representation through interleaving linear and nonlinear processing.This high-level semantic feature vector is input into the attention weight analysis module for attention modeling. Based on the self-attention mechanism, this module calculates the weight distribution for each dimension or feature component of the semantic feature vector at each moment. By calculating the correlation matrix between features, it determines which features are most critical for speech intent recognition and which components can be ignored or weakened. The attention mechanism allows the system to dynamically select the most discriminative information paths from the overall representation of the speech signal, thereby forming a non-uniform information distribution focusing structure in the feature space, enabling the model to focus on key morphemes of instructions, central sentence structures, or inflection points of intonation. The attention-fused semantic vector, i.e., the intermediate representation vector, is calculated based on the weighted attention distribution.

[0083] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0084] The standardized speech intent vector is input into the Transformer encoder of the Industrial Internet large language model for mapping processing to obtain industrial speech intent encoding data;

[0085] Based on industrial speech intent coding data, professional terminology query and semantic relationship analysis are performed on the pre-built industrial knowledge graph to obtain industrial domain professional knowledge reasoning results;

[0086] The results of industrial domain professional knowledge reasoning are integrated with the current production line status data, AMR cluster location distribution information, and task queue status information through attention mechanism and context perception to obtain comprehensive semantic reasoning data.

[0087] The comprehensive semantic reasoning data is input into the Transformer decoder of the Industrial Internet large language model to generate task decomposition sequences, construct parameter configuration tables, and sort execution priorities to obtain the AMR execution plan.

[0088] Specifically, a standardized speech intent vector is input into the Transformer encoder structure of a large language model deployed on the industrial internet platform. This standardized speech intent vector consists of four fields: a command type identifier, a target object code, an operation parameter value, and a confidence score. Each field is uniformly encoded in 32-bit floating-point format and undergoes dimensionality expansion and vector normalization to ensure effective model reception and analysis. The Transformer encoder performs a linear transformation on the vector through an embedding layer, projecting the values ​​of each dimension into a high-dimensional semantic space. A learnable embedding weight matrix is ​​introduced to construct a semantic representation. A positional encoding vector generated by sine and cosine functions is then fused with the original embedding vector based on its sequential position in the control context to preserve the temporal and structural dependencies of the semantic units. The fused vector is then fed into a multi-layer stack of self-attention modules. At each layer, the correlation weights between the semantic components are calculated by transforming the query, key, and value vectors. This allows for deep modeling of the grammatical relationships, logical coupling, and control direction within the input semantic vector. The representation is then updated layer by layer, outputting a set of industrial speech intent encoded data with contextual semantics and structural abstraction capabilities. Based on this encoded data, the system invokes a pre-built industrial knowledge graph module. This graph stores 30,000 industrial production processes, 50,000 equipment operating manuals, and 80,000 common fault diagnosis records in a graph-like structure. Nodes contain entity information such as industrial terminology, equipment components, control parameters, and workflows, while edges represent semantic, operational, dependency, and causal relationships between them. Based on the keywords, control logic, and object references extracted from the encoded data, the system performs terminology queries on the knowledge graph. The entity recognition subsystem locates the control semantics and target objects involved in the voice commands. Using a semantic path search mechanism, the system conducts adjacency queries on relevant nodes within the graph, extracting information about hyponyms, parallel actions, and causal chains. Furthermore, the system uses a semantic relationship analysis module to reason about the graph structure between the extracted terms, identifying industrial rules such as "a certain operation must be performed first under certain conditions" or "a certain parameter adjustment should be triggered when a certain state occurs." This generates structured inference results based on industrial domain expertise. To enhance the adaptability of task generation, this industrial knowledge inference result is integrated with real-time production line status data, AMR cluster location distribution information, and task queue status information. Production line status data includes the current production line rhythm, workstation occupancy, work-in-progress processing progress and bottleneck node status; AMR cluster distribution information includes the current coordinates, power level, task execution status and communication connectivity of each mobile robot; task queue status information covers task items queued for execution, priority sorting, dependencies and task completion.The system integrates this heterogeneous information using the attention mechanism within the Transformer architecture. It uses industrial knowledge reasoning results as the main sequence input and various contextual information as external context input. A cross-attention structure is used to model these associations. Supported by a multi-layered attention weighting mechanism, the system automatically focuses on the core production nodes, resource states, or constraint variables upon which the current voice command depends. This generates comprehensive semantic reasoning data encompassing control semantics, resource constraints, and contextual dependencies. This data structure, expressed as a tensor, preserves the logical order of operations, parameter dependency paths, and execution risk assessment information. This comprehensive semantic reasoning data is input into the Transformer decoder of the Industrial Internet Big Language Model to execute the task generation logic. Within the decoder, the model uses a position-aware mechanism to structure the input data and, at each step, predicts the next most appropriate operation step based on the generated partial task sequence. This process utilizes the decoder's built-in multi-head attention mechanism to achieve continuous focus and content selection on the input data. The output of task generation is represented as a logical sequence of operation instructions, known as a task decomposition sequence, where each element corresponds to a basic operation unit recognizable by the AMR controller, such as "move to coordinate (x, y)," "turn θ degrees," and "complete loading." In parallel with task decomposition and generation, the model constructs a parameter configuration table that lists the execution speed, safety margin, resource requirements, power estimate, and action time window required for each operation. These parameters are dynamically derived and bound to the corresponding operation steps through the conditional generation module in the model. Furthermore, the system introduces an execution priority ranking mechanism that evaluates each task based on its urgency, dependency on upstream processes, and AMR resource availability, and assigns ranking tags to form a prioritized task execution chain. All task data is encapsulated as an AMR execution plan, and the output format is structured JSON. Fields include task ID, action type, target coordinates, parameter table, execution time window, priority identifier, and security policy.

[0089] In a specific embodiment, the step of inputting the standardized speech intent vector into the Transformer encoder of the industrial Internet large language model for mapping processing to obtain industrial speech intent encoded data may specifically include the following steps:

[0090] Perform dimension expansion and numerical normalization on the standardized speech intention vector to obtain vector dimension conversion data;

[0091] Input the vector dimension conversion data into the input embedding layer of the Transformer encoder for linear transformation and weight matrix mapping to obtain the input embedding vector data;

[0092] Sine and cosine position encoding calculation and vector fusion are performed based on the serial position of the input embedded vector data to obtain a position encoding fusion vector. Multi-head attention calculation is performed on the position encoding fusion vector to obtain industrial speech intent encoding data.

[0093] Specifically, the standardized speech intent vector is preprocessed. The vector's original structure consists of four fields: a command type identifier, a target object code, an operation parameter value, and a confidence score. This vector undergoes dimensionality expansion and numerical normalization. The dimensionality expansion process involves defining a set of learnable upscaling matrices or employing an interpolation-based expansion mechanism to elevate the original vector from a low-dimensional space to the high-dimensional feature space required by the model, for example, from 4 to 128 dimensions. Each dimension not only retains its original meaning after expansion but also incorporates a projected space of contextual features learned through training to represent the potential relevance of that field in the industrial semantic space. The numerical normalization process uses maximum-minimum scaling or Z-score normalization techniques to normalize the values ​​of each dimension so that they are evenly distributed within the range of [-1, 1] or 0 with a mean and unit variance. This prevents bias in the subsequent weight matrix projection caused by dimensional differences between fields. This step results in a set of dimensionally expanded and numerically smoothed vectors, known as the vector dimension-transformed data. The dimension-transformed vectors are then fed into the input embedding layer of the Transformer encoder architecture, which converts the numerical vectors into embedding representations that can be processed within the model. The embedding layer uses a linear transformation mechanism, multiplying the input vector by a set of learnable weight matrices and adding a bias vector to project it into the model's embedding space. This operation achieves dimensionality mapping and, through training, automatically captures implicit correlations between different dimensions in the input data. This linear transformation helps reconstruct this information, particularly when weak structural dependencies or semantically implied relationships exist between different fields. After completing the linear projection, the embedding layer semantically embeds certain discrete fields (such as instruction type numbers and object codes) using an internally constructed word embedding lookup table or continuous variable embedding function. This creates a representation structure in the vector space that is closer to that understood by the language model, thereby generating a set of input embedding vectors. The Transformer encoder introduces a positional encoding mechanism to enhance the model's ability to perceive the position of each element in the input sequence. Because the Transformer architecture, unlike traditional recurrent neural networks, lacks the natural ability to model sequential sequences, the system explicitly incorporates positional information into the embedding representation to maintain the order of the semantic structure. Each dimension of the input embedding vector is assigned an encoding value related to its sequence position. These position encoding values ​​are calculated by combining sine and cosine functions, defined as the position encoding using the sine function for the even-numbered dimension and the cosine function for the odd-numbered dimension, with the frequency varying exponentially with the dimension. The calculation method based on trigonometric functions ensures a predictable offset relationship between any two position encodings, making it easier for the model to learn the relative relationship between positions during training. The system performs element-by-element summing and fusion of this position encoding vector with the original input embedding vector to obtain a position encoding fusion vector. This position encoding fusion vector is input into the multi-head attention mechanism of the Transformer encoder for semantic modeling.Within this module, the fused vector is subjected to three weight mappings: query, key, and value. These are then fed into multiple attention heads. Each attention head calculates the dot-product similarity between the query and key, scales it, and performs softmax normalization to generate an attention weight matrix. This weighted matrix is ​​then used to linearly combine the values ​​to generate the corresponding attention output representation for each head. These heads focus on different semantic dimensions and feature subspaces, enabling the parallel modeling of diverse relationships, such as the preferential dependencies between action targets, the control coupling between parameters and objects, and the quantitative connection between confidence and execution risk. Finally, the outputs of each attention head are concatenated and redimensionalized through a set of linear transformation layers, ultimately generating a unified semantic vector, the industrial speech intent encoding data. This encoding data integrates the deep semantic structure extracted from the input intent, the logical path between execution instructions, the interaction between control semantics and context, and the nonlinear relationship between confidence level and intent strength.

[0094] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0095] The task decomposition operation sequence in the AMR execution plan is split into task units and the workload is evaluated. At the same time, the status data of each AMR master controller in the AMR cluster network is synchronized and information is updated through industrial Ethernet to obtain the AMR cluster global status table;

[0096] Based on the global state table of the AMR cluster, the task decomposition operation sequence is load balanced to obtain a set of candidate task allocation solutions;

[0097] The set of task allocation candidate solutions is input into the distributed consensus algorithm for multi-objective optimization solution and feasibility verification to obtain the optimal task allocation solution;

[0098] According to the optimal task allocation scheme, the control instruction sequence is constructed and the priority arbitration of urgent tasks is performed to obtain a distributed control instruction sequence.

[0099] Specifically, the task decomposition operation sequence in the AMR execution plan is parsed and restructured. This task decomposition operation sequence consists of multiple logical operation units, each corresponding to an action instruction that can be materialized by a specific AMR actuator, such as moving to a specified location, loading cargo, avoiding obstacles, steering, or path inspection. The system first performs a task unit splitting operation on this sequence. Using a semantic parsing engine, the system extracts the action type, spatial parameters, required resources, and expected execution time of each operation instruction, and analyzes its task dependencies and execution preconditions. During the task splitting process, a granularity minimization strategy is adopted to split macro instructions into independent control units. This ensures that each split task has the smallest indivisible execution logical boundary, ensuring maximum flexibility in task distribution and scheduling. Furthermore, the system introduces a workload assessment module to quantitatively calculate each task unit. This assessment comprehensively considers the required driving distance, control complexity, power load, time consumption, and equipment wear parameters. A normalized scoring mechanism is used to generate a workload indicator vector for each unit, which is used to guide the subsequent task allocation weight control. After completing task splitting and workload assessment, the system synchronizes the status data of each master controller in the AMR cluster network currently within the task scheduling domain via Industrial Ethernet. This process is initiated by the scheduling management server, which requests and receives the online operating status of each AMR in real time via Ethernet. This includes data indicators in multiple dimensions, such as current location coordinates, battery charge percentage, task execution progress, current load, sensor health status, and communication link quality. A unified timestamp marking mechanism ensures the temporal consistency of status data uploaded by different master controllers. The collected data is mapped and organized into a multi-field state structure table, the AMR cluster global state table, which uses the AMR device number as the index field. The real-time data of all master controllers is integrated and cached in a structured matrix format into the distributed state synchronization module, forming the real-time environmental information foundation for task scheduling. Based on this AMR cluster global state table, load balancing is performed on the task decomposition operation sequence completed in the previous stage. The scheduling engine evaluates the adaptability and resource matching of each task unit by matching the workload vector with the current status indicator of each AMR, and constructs a task-resource matching scoring matrix based on the current task priority, spatial location information, job density, and available path conflict prediction results. It then constructs multiple task assignment schemes through a combination algorithm. These schemes represent candidate task allocation paths under different scheduling strategies, such as the shortest workload priority strategy, the shortest path strategy, the optimal power strategy, or the comprehensive weighted scoring strategy. The system ultimately forms a set of task assignment candidate schemes. Each item in the set contains the assignment target AMR number, estimated start time, execution duration, task dependencies, and post-completion status prediction of the specific task unit.To select the optimal solution from a set of candidate solutions, the set is fed into a distributed consensus algorithm module for multi-objective optimization and feasibility verification. Based on a modified Raft or Paxos consensus mechanism, this module establishes a consensus voting and negotiation process among multiple AMR master controllers to ensure system-wide state acceptability and resource non-conflict. During the multi-objective optimization process, a cost function set is introduced to calculate the cost of each candidate solution. This cost function simultaneously considers multiple objectives, including total execution time, total power consumption, path overlap conflict risk, resource usage balance, and task overdue risk. The optimal solution is searched for using a weighted linear combination or Pareto frontier approximation algorithm. During the solution process, the system performs constraint verification on each candidate solution, including verifying hard constraints such as load limit, power threshold, safety distance lower limit, and task timing dependency consistency. Infeasible or conflicting solutions are eliminated, and the allocation strategies that meet all constraints and achieve the best global performance are retained. Ultimately, a single optimal task allocation solution is output. Based on this optimal task allocation scheme, the control instruction sequence construction module is activated, converting the execution information of each task unit into a standard control instruction structure. Each instruction contains the target AMR number, action type, control parameters, execution start and end times, path tags, environmental warning parameters, and task dependency identifiers. These instructions are then sorted on a timeline to form a task execution sequence. At the same time, the system introduces an emergency task arbitration mechanism. If a higher-priority emergency task request (such as fault handling or urgent shipment transfer) occurs during execution, the system will suspend the current non-critical task via a soft interrupt, placing it in a waiting state queue. The system then dynamically rearranges the current instruction sequence and reallocates resources, generating a distributed control instruction sequence with priority tags and pre-scheduling information.

[0100] In a specific embodiment, the execution step inputs the set of candidate task allocation solutions into a distributed consensus algorithm for multi-objective optimization and feasibility verification. The process of obtaining the optimal task allocation solution may specifically include the following steps:

[0101] For each candidate task allocation scheme in the set of task allocation candidate schemes, the total execution time calculation, energy consumption evaluation and path conflict probability statistics are performed to obtain a multi-objective function numerical matrix;

[0102] Based on the multi-objective function numerical matrix, the set of task allocation candidate solutions is subjected to load limit boundary check, power consumption threshold verification and safety distance interval verification, and a subset of feasible candidate solutions is obtained;

[0103] The subset of feasible candidate solutions is input into the distributed consensus algorithm for optimization and solution to obtain the optimal task allocation solution.

[0104] Specifically, performance metrics are calculated for a set of candidate task allocation solutions to form the basis for multi-objective optimization analysis. This set of candidate solutions consists of several task assignment combinations generated for different AMR resource states and task matching logic. Each candidate solution represents a feasible task allocation path with different strategies, including information such as the assigned AMR number, execution time period, path trajectory, load value, and execution order for each task unit. The system uses the scheduling engine to read each candidate task allocation plan in the set in turn and calculates the total execution time for each of them. That is, based on the execution timestamp data in the task operation sequence, combined with the AMR execution speed model, the task switching delay estimation function and the inter-task dependency constraint graph, the critical path length of the entire plan is calculated to obtain the upper bound of the time required to complete the batch of tasks as a whole, that is, the total execution time indicator of the candidate plan; the system evaluates the energy consumption of the same plan, and integrates and accumulates the energy consumption of each operation unit based on the energy consumption model of the AMR drive motor, the power-load curve and the task path distance to obtain the total energy consumption indicator of the plan; to evaluate the potential path interference risk during multi-machine parallel operation, the system introduces a path conflict probability calculation mechanism. By performing time synchronization analysis on the operation trajectory of each AMR during the task, the spatial and temporal overlap between trajectory segments is detected, and a conflict probability value is given in combination with the historical path interaction model to form a conflict indicator representing the spatial collaboration risk of the candidate plan. The three metrics of total execution time, energy consumption, and path conflict probability are assembled into a multi-objective function matrix. Each row of this matrix corresponds to a candidate solution, and each column represents an optimization objective. This matrix forms the core data structure for multi-objective scheduling performance evaluation. Based on this multi-objective function matrix, the system enters the task feasibility constraint verification phase, verifying the resource constraints and safety margins of each candidate task allocation solution.During this process, the load configuration of each candidate solution is checked for load limits. The total load corresponding to each assigned AMR in the task set is retrieved and compared with its designed maximum load threshold. If the cumulative load of an AMR within a task interval exceeds its rated load, the system will determine that the solution has a physical overload risk and immediately remove it. The total power consumed by each AMR to execute the task path required for the solution is compared with the current battery power. If the total energy consumption of the tasks required to be executed by an AMR in the solution exceeds its remaining available power, it indicates that there is a risk of failure during execution and is therefore eliminated. To ensure the safety of multiple AMRs moving in the same area, the path data is checked for safe distance intervals. Specifically, the Euclidean distance between the spatial positions of any two AMRs in the same time window is calculated and compared with the safety threshold. If the distance between the two AMRs at any time during operation is less than the set safety lower limit (e.g., 50 cm), it indicates a collision risk. The system marks the solution as unsafe and removes it from the set. After checking resource boundaries and safety constraints along the three dimensions, a subset of feasible candidate solutions is selected that simultaneously meet task feasibility, hardware resource loadability, and operational safety requirements. This subset of feasible candidate solutions is then fed into a multi-objective scheduling optimization solution framework based on a distributed consensus algorithm to conduct the final search for the optimal task allocation solution. Within this module, global state negotiation and task consensus voting are carried out across multiple AMR master nodes, using Paxos or Raft-like consensus protocols as the underlying communication and state synchronization mechanism. The system simultaneously runs a local task cost function calculator on each node. By evaluating the resource utilization, execution efficiency, and path risk of the current candidate solution for the corresponding task on that node, it outputs a local scoring result and synchronizes it to the global coordination node. The coordination node then performs a weighted summation, negotiation ranking, and consensus proposal process on the overall cost vector of the candidate solutions based on the scoring feedback from all master nodes. The core goal of the entire process is to find a scheduling strategy from a subset of candidate solutions that achieves the best overall performance under the three objectives of "shortest total execution time," "lowest system energy consumption," and "lowest path conflict probability," while ensuring that the strategy is free of resource conflicts, communication blockages, or security risks among all participating nodes, and that it meets the system's global consistency and stability requirements. During the algorithm's execution, the system introduces weight adjustment factors and penalty function strategies to dynamically guide the solution to converge toward the Pareto frontier, thereby balancing the trade-offs between multiple objectives. Ultimately, a consensus algorithm is used to achieve collaborative recognition and output a unique optimal task allocation solution. Structurally, the solution consists of multiple task-AMR mapping units, each of which records the task number, assignment object, expected start time, task duration, path trajectory, load status, and predictable risk identifier.

[0105] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0106] The distributed control instruction sequence is input into the main execution thread of the motion controller in the AMR main controller for instruction parsing and hardware drive control. At the same time, the monitoring thread is started to collect status data and manage cache to obtain hardware control instructions and real-time status monitoring data streams.

[0107] Based on the real-time status monitoring data stream, encoder position feedback, force sensor load feedback, and tilt sensor posture feedback are periodically sampled and time series data cached to obtain the AMR execution status data packet;

[0108] Based on the AMR execution status data packet, position error threshold detection, speed deviation threshold detection and load change anomaly detection are performed to obtain the execution anomaly detection result;

[0109] Based on the results of anomaly detection, data is uploaded through industrial Ethernet and analyzed by the industrial Internet large language model to obtain the anomaly feedback processing strategy.

[0110] Specifically, a distributed control command sequence is input into the main controller of each target AMR, where its integrated motion control module manages multi-threaded execution. Within the motion control module, this control command sequence is parsed into multiple task execution units and, based on a thread scheduling mechanism, fed into two task distribution channels: the main execution thread and the monitoring thread. As the core execution path of the control system, the main execution thread primarily parses the format, decodes parameters, and performs hardware translation for each control command. By analyzing elements such as the action type field, path coordinate parameters, execution speed value, and control time window, the system logically maps the command into a low-level, drivable form. Based on the action type, the system invokes hardware control interfaces such as the servo motor driver, steering mechanism actuator, and lift platform controller to generate the corresponding pulse control signals. During execution, the system encodes motor torque, servo angle, and lift level signals through PWM (pulse width modulation) control channels, outputting these signals to the actual mechanical actuators via the driver module. This completes the command delivery path from upper-level semantic control to physical motion control, forming a clear, closed-loop "voice → planning → scheduling → execution" link. At the same time, the system starts a monitoring thread and configures an independent memory cache and high-frequency sampling priority for it. This thread is mainly responsible for collecting the status data stream generated by the AMR during execution from various sensor channels in real time, including the shaft angle and displacement data from the position encoder, the load stress and center of gravity offset data from the force sensor, and the pitch angle and roll angle change data from the tilt angle sensor. The monitoring thread sets a millisecond-level periodic scheduling mechanism to poll the sensor sampling port for data at fixed time intervals, binds the sampling results to the system timestamp, and uniformly stores them in the local timing cache, forming a status data stream containing multi-source real-time physical feedback, which is passed as input to the anomaly detection module. After receiving this real-time status monitoring data stream, the system activates the status analysis module to perform structured encapsulation on the sampled data, generating an AMR execution status data packet. This packet is organized in a fixed format and includes 64 bytes of position status data (such as encoder absolute displacement and directional offset), 32 bytes of motion vector data (such as velocity and acceleration components), 16 bytes of load status data (such as upper and lower pressure limits and gradients), and 8 bytes of attitude angle data (such as roll and pitch angles). Each field is timestamped to support subsequent dynamic behavior assessment and anomaly tracing. The system then enters the anomaly detection phase, applying three types of anomaly determination rules to perform dynamic threshold detection and evolution trend tracking on each status data packet.The position error threshold detection module calculates the difference between the current encoder position feedback and the desired execution position. If the difference exceeds a set upper limit (e.g., ±5 cm), a path deviation anomaly is detected. The speed deviation detection module determines the relative error between the target speed and the measured speed. If the difference exceeds a set ratio (e.g., 10%) and persists for more than two sampling cycles, the system triggers a speed out-of-control warning. The load change anomaly detection module analyzes the real-time pressure fluctuation trend of the load sensor. If the load suddenly changes during the static phase, exceeding a preset threshold, or exhibits high-frequency unstable oscillation, the system identifies a possible risk of abnormal collision, overload, platform looseness, or cargo drop. These three detection algorithms are executed separately and each result is labeled. The system aggregates these labeled results into an execution anomaly detection report, which identifies the anomaly type, anomaly parameters, trigger time, data context, and recommended response level. This execution anomaly detection result is uploaded to the Industrial Internet platform via Industrial Ethernet, where a large language model deployed on the platform analyzes the cause of the anomaly and generates a policy. Uploaded anomaly detection data is embedded as input into the large language model and fed into the model's Transformer encoder module for semantic vector mapping. This vector includes elements such as task context, control command sequence, AMR physical state, anomaly type, and parameter structure. The model aligns these embedded vectors with the industrial control fault semantic map established during pre-training. The model then performs an attention-based anomaly causal chain tracking to determine whether the current anomaly is caused by hardware failure, path conflict, control error, or environmental disturbance. The anomaly level is further classified and assessed based on historical anomaly distribution data. Ultimately, an anomaly feedback handling strategy is generated, including compensation strategies, action correction suggestions, task rescheduling instructions, and voice feedback. This strategy includes specific instruction correction methods (such as speed reduction, path replanning, and soft interrupt execution). It is then synchronously transmitted back to the AMR main controller for real-time control logic adjustment. Status feedback is also provided to the operator via a voice broadcast module.

[0111] In a specific embodiment, the execution step inputs the distributed control instruction sequence into the main execution thread of the motion controller in the AMR main controller for instruction parsing and hardware driver control, and simultaneously starts the monitoring thread for status data collection and cache management. The process of obtaining hardware control instructions and real-time status monitoring data streams can specifically include the following steps:

[0112] The distributed control instruction sequence is input into the motion controller of the AMR master controller for instruction format parsing and thread resource allocation, and a multi-thread execution scheduling scheme including a main execution thread task queue and a monitoring thread task queue is obtained;

[0113] Based on the multi-threaded execution scheduling scheme, the action type instructions and execution parameter values ​​in the main execution thread task queue are input into the servo motor driver, steering mechanism controller, and lifting mechanism operator for pulse width modulation signal generation and power amplification processing to obtain hardware control instructions;

[0114] According to the multi-threaded execution scheduling scheme, the monitoring thread is started to perform high-frequency data reading and parameter conversion on the encoder sensor, inertial measurement unit sensor, and tilt angle sensor of the AMR main controller to obtain a real-time status monitoring data stream.

[0115] Specifically, a distributed control instruction sequence is input into the motion controller module within the AMR main controller, where the instruction format parsing and thread resource allocation process is initiated. During this phase, the motion controller performs semantic deconstruction and field separation on each distributed control instruction, extracting key elements such as the action type, target parameter value, time tag, path identifier, and priority tag. Based on the thread division mechanism in the scheduling strategy, the instructions are divided into two categories: the main execution thread tasks, which are responsible for controlling the servo motor, drive platform, and physical operating mechanism; and the status monitoring thread tasks, which require comparing and verifying the execution instructions with feedback data read from high-frequency sensors. During the parsing process, the system comprehensively models the timing logic, thread resource usage, and expected response cycle between tasks. Each task category is written into the corresponding thread's pending task queue based on priority and execution time window, thus forming a multi-threaded execution scheduling scheme consisting of the main execution thread task queue and the monitoring thread task queue. This scheme is registered and activated in the main controller as a thread-independent scheduling table, which guides the specific operation sequence, instruction invocation, and resource scheduling of each subsequent thread. The action execution module is initiated based on the contents of the main execution thread's task queue within the multi-threaded scheduling scheme. This module determines the corresponding hardware execution path based on the type field (e.g., forward, turn, or lift) in each action instruction. It then maps the associated control parameters (e.g., target position, speed, angular amplitude, or lift height) into the device control instruction semantic space. These parameters are then fed into the underlying execution units, including the servo motor driver, steering mechanism controller, and lift mechanism operator. During control signal generation, the system uses a dedicated PWM (pulse width modulation) module to convert the parameter values ​​into periodic electrical signals with corresponding frequency, duty cycle, and phase control. Specifically, the speed value is encoded as a PWM frequency, the angular deviation is encoded as a duty cycle, and the path correction direction is mapped as a phase adjustment. The system then uses a power amplifier circuit to boost the PWM signal to an energy level sufficient to drive the motor load and synchronously output it to the drive actuator. The servo motor uses the received PWM signal for fine speed and displacement control. The steering mechanism uses the angular signal to adjust smooth curved motion. The lift mechanism controls the height or clamping state of the device based on the lift amplitude. The main execution thread continuously reads instructions from the task queue and, with the participation of real-time feedback, performs instruction advancement and exception interrupt identification, ensuring that each task instruction is accurately and smoothly converted into physical execution actions, establishing a direct response chain at the AMR behavioral level. Simultaneously, the system synchronously launches the monitoring thread based on the contents of the monitoring thread's task queue in the multi-threaded scheduling scheme and allocates it high-priority processing channels and cache space to meet the real-time needs of high-frequency data collection and closed-loop status feedback.The monitoring thread is primarily responsible for continuously sampling core sensors related to task execution, including encoder sensors (for recording wheel or actuator rotation angle and displacement), inertial measurement units (IMUs) (for providing acceleration, angular velocity, and three-axis attitude data), and tilt sensors (for acquiring the platform's pitch and roll attitude). The system uses a timed interrupt mechanism to control the sampling period, triggering a sensor polling reading task every 10ms, ensuring high-frequency, continuous time series acquisition of all key state parameters during action execution. The raw analog data is converted to digital signals by the A / D conversion module and then fed into the parameter conversion module for format standardization. This data is converted into a state vector with consistent units and clear dimensions, such as angles to radians, velocities to standard meters per second, and attitude angles to three-dimensional vectors. Timestamp information is then appended and written to the state monitoring buffer. As state data accumulates, the monitoring thread uses the local buffer to generate a real-time state monitoring data stream. This data stream is output in batches in a structured format for real-time use by the anomaly detection, control correction, or feedback voice broadcast modules.

[0116] In an embodiment of the present invention, the spatiotemporal synchronization of a digital microphone array, lidar, inertial measurement unit, and visual sensor achieves precise association between voice commands and environmental conditions, significantly improving the accuracy of command understanding and environmental adaptability. Synchronous alignment processing using a unified time base resolves the time inconsistency issue of multi-source sensor data, ensuring the real-time and reliability of control decisions. Based on a lightweight industrial speech recognition neural network and hybrid neural network architecture, high-precision recognition of industrial terminology is achieved, effectively addressing the low accuracy of traditional speech recognition in industrial environments. Pre-built industrial speech command classification vocabulary matching enables accurate classification and semantic understanding of complex industrial commands, enhancing the system's ability to process industrial domain expertise. The Transformer architecture of encoders and decoders enables deep reasoning and context-aware processing of industrial domain expertise, significantly enhancing the intelligent level of task planning. Through terminology query and semantic relationship analysis within the industrial knowledge graph, the integrated application of multi-dimensional knowledge, including manufacturing processes, equipment control, and quality management, improves the professionalism and feasibility of execution plans. By improving the distributed consensus algorithm to perform Pareto front search and non-dominated sorting, a multi-objective global optimization of total execution time, energy consumption, and conflict probability is achieved, effectively solving the local optimality problem of traditional task allocation algorithms. Feasibility verification based on load restrictions, power constraints, and safe distance constraints ensures the safety and executability of the task allocation scheme and improves the stability of AMR cluster coordination. The parallel processing architecture of the main execution thread and the monitoring thread enables simultaneous instruction execution and status monitoring, improving the system's real-time responsiveness and execution efficiency. High-frequency sampling and circular storage in a ring buffer enable high-precision real-time status monitoring. Intelligent anomaly analysis and processing based on the large language model of the industrial Internet enables the automatic generation of adaptive control strategies such as trajectory correction, speed adjustment, and path replanning, significantly improving the level of intelligence.

[0117] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0118] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a speech data interactive feedback control processing device based on a large language model (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program code.

[0119] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A speech data interactive feedback control processing method based on a large language model, characterized in that: include: Collecting industrial voice command data through the AMR master controller, and performing industrial vocabulary matching and voice intent classification on the industrial voice command data to obtain a standardized voice intent vector; Inputting the standardized speech intention vector into the industrial Internet large language model for industrial domain expertise reasoning and context perception processing to obtain an AMR execution plan; Based on the AMR execution scheme, distributed consistency coordination and task allocation are performed on the AMR cluster network to obtain a distributed control instruction sequence, including: splitting the task decomposition operation sequence in the AMR execution scheme into task units and evaluating the workload, and simultaneously synchronizing the state data and updating the information of each AMR master controller in the AMR cluster network through industrial Ethernet to obtain an AMR cluster global state table; based on the AMR cluster global state table, load balancing is performed on the task decomposition operation sequence to obtain a set of task allocation candidate schemes; the set of task allocation candidate schemes is input into a distributed consistency algorithm for multi-objective optimization solution and feasibility verification to obtain an optimal task allocation scheme; and control instruction sequence construction and emergency task priority arbitration are performed according to the optimal task allocation scheme to obtain a distributed control instruction sequence; The distributed control instruction sequence is input into the motion controller of the AMR master controller for multi-threaded execution monitoring and exception processing to obtain an exception feedback processing strategy.

2. The speech data interactive feedback control processing method based on a large language model according to claim 1, characterized in that: The industrial voice command data is collected by the AMR master controller, and the industrial vocabulary matching and voice intent classification are performed on the industrial voice command data to obtain a standardized voice intent vector, including: Collecting industrial voice command data through a voice data acquisition module in the AMR main controller, and inputting the industrial voice command data into an audio digital signal processor for voice activity detection to obtain voice feature data; Synchronously reading data and extracting parameters from the lidar sensor, inertial measurement unit sensor, and vision sensor of the AMR master controller to obtain a multi-source sensor data set; Based on the reception time of the voice feature data, the multi-source sensor data set is timestamped and synchronized to obtain a spatiotemporal synchronized data group; Associating, binding, and encoding the speech feature data with the spatiotemporal synchronization data group to obtain multi-dimensional fused speech data; The multi-dimensional fused voice data is input into the edge reasoning module of the AMR main controller for industrial vocabulary matching and voice intent classification to obtain a standardized voice intent vector.

3. The speech data interactive feedback control processing method based on a large language model according to claim 2, characterized in that: The step of inputting the multi-dimensional fused voice data into the edge reasoning module of the AMR master controller for industrial vocabulary matching and voice intent classification to obtain a standardized voice intent vector includes: Inputting the multi-dimensional fusion voice data into the edge reasoning module in the AMR main controller to extract voice feature vectors and perform data format standardization processing to obtain voice feature vectors; Inputting the speech feature vector into a hybrid neural network architecture for feature learning and speech pattern recognition to obtain an intermediate representation vector; Performing vocabulary matching and semantic similarity calculation on a pre-built industrial voice command classification vocabulary based on the intermediate representation vector to obtain vocabulary matching data; Speech intent classification calculation and confidence evaluation are performed based on the vocabulary matching data to obtain a standardized speech intent vector.

4. The speech data interactive feedback control processing method based on a large language model according to claim 3, characterized in that: The step of inputting the speech feature vector into a hybrid neural network architecture for feature learning and speech pattern recognition to obtain an intermediate representation vector comprises: Inputting the speech feature vector into a convolutional neural network in a hybrid neural network architecture to perform multi-scale feature extraction and local speech pattern recognition to obtain convolution feature data; Inputting the convolution feature data into the long short-term memory network in the hybrid neural network architecture to perform temporal dependency modeling and speech context semantic learning to obtain sequence representation data; Based on the sequence representation data, linear transformation and nonlinear mapping are performed to obtain a high-level semantic feature vector, and attention weight analysis is performed on the high-level semantic feature vector to obtain an intermediate representation vector.

5. The speech data interactive feedback control processing method based on a large language model according to claim 1, characterized in that: The step of inputting the standardized speech intention vector into the industrial Internet large language model for industrial domain expertise reasoning and context perception processing to obtain an AMR execution plan includes: Inputting the standardized speech intention vector into the Transformer encoder of the industrial Internet large language model for mapping processing to obtain industrial speech intention encoding data; Based on the industrial speech intention coding data, professional terminology query and semantic relationship analysis are performed on the pre-built industrial knowledge graph to obtain industrial domain professional knowledge reasoning results; The industrial domain expertise reasoning results are integrated with the current production line status data, AMR cluster location distribution information, and task queue status information through attention mechanism fusion and context perception to obtain comprehensive semantic reasoning data; The comprehensive semantic reasoning data is input into the Transformer decoder of the industrial Internet large language model to generate task decomposition sequences, construct parameter configuration tables, and sort execution priorities to obtain the AMR execution plan.

6. The speech data interactive feedback control processing method based on a large language model according to claim 5, characterized in that: Inputting the standardized speech intention vector into the Transformer encoder of the industrial Internet large language model for mapping processing to obtain industrial speech intention encoded data includes: Performing dimension expansion and numerical normalization processing on the standardized speech intention vector to obtain vector dimension conversion data; Inputting the vector dimension conversion data into the input embedding layer of the Transformer encoder for linear transformation and weight matrix mapping to obtain input embedding vector data; Sine-cosine position coding calculation and vector fusion are performed based on the sequence position of the input embedded vector data to obtain a position coding fusion vector, and multi-head attention calculation is performed on the position coding fusion vector to obtain industrial speech intention coding data.

7. The speech data interactive feedback control processing method based on a large language model according to claim 1, characterized in that: The step of inputting the set of candidate task allocation solutions into a distributed consensus algorithm for multi-objective optimization and feasibility verification to obtain an optimal task allocation solution includes: Performing total execution time calculation, energy consumption evaluation, and path conflict probability statistics on each candidate task allocation scheme in the set of task allocation candidate schemes to obtain a multi-objective function numerical matrix; Based on the multi-objective function numerical matrix, the set of task allocation candidate solutions is subjected to load limit boundary check, power consumption threshold verification and safety distance interval verification processing to obtain a subset of feasible candidate solutions; The subset of feasible candidate solutions is input into a distributed consensus algorithm for optimization and solution to obtain an optimal task allocation solution.

8. The speech data interactive feedback control processing method based on a large language model according to claim 1, characterized in that: The step of inputting the distributed control instruction sequence into the motion controller of the AMR master controller for multi-threaded execution monitoring and exception handling to obtain an exception feedback handling strategy includes: The distributed control instruction sequence is input into the main execution thread of the motion controller in the AMR main controller for instruction parsing and hardware drive control, and the monitoring thread is started to collect status data and manage cache to obtain hardware control instructions and real-time status monitoring data stream; Based on the real-time status monitoring data stream, encoder position feedback, force sensor load feedback, and tilt sensor posture feedback are periodically sampled and time series data cached to obtain an AMR execution status data packet; Performing position error threshold detection, speed deviation threshold detection, and load change anomaly detection based on the AMR execution status data packet to obtain an execution anomaly detection result; According to the execution anomaly detection result, data is uploaded and analyzed by the industrial Internet large language model through industrial Ethernet to obtain an abnormal feedback processing strategy.

9. The speech data interactive feedback control processing method based on a large language model according to claim 8, characterized in that: The distributed control instruction sequence is input into the main execution thread of the motion controller in the AMR main controller for instruction parsing and hardware drive control, and the monitoring thread is started to collect status data and manage cache to obtain hardware control instructions and real-time status monitoring data streams, including: Inputting the distributed control instruction sequence into the motion controller of the AMR master controller for instruction format parsing and thread resource allocation, thereby obtaining a multi-thread execution scheduling scheme including a main execution thread task queue and a monitoring thread task queue; Based on the multi-threaded execution scheduling scheme, the action type instruction and execution parameter values ​​in the main execution thread task queue are input into the servo motor driver, the steering mechanism controller, and the lifting mechanism operator to generate pulse width modulation signals and perform power amplification processing to obtain hardware control instructions; According to the multi-threaded execution scheduling scheme, the monitoring thread is started to perform high-frequency data reading and parameter conversion on the encoder sensor, inertial measurement unit sensor, and tilt angle sensor of the AMR main controller to obtain a real-time status monitoring data stream.

Images