Search

CN-122001912-A - Semantic driving cloud edge collaborative monitoring method based on narrow-band network

CN122001912ACN 122001912 ACN122001912 ACN 122001912ACN-122001912-A

Abstract

The invention belongs to the technical field of network intelligent management, and particularly relates to a semantic driving cloud edge collaborative monitoring method based on a narrow-band network; the cloud edge collaborative monitoring system comprises a cloud semantic decision layer, a network transmission layer and an edge perception layer, wherein the cloud semantic decision layer receives natural language instructions and builds a semantic reasoning environment, analyzes the natural language instructions, packages analysis results into scheduling instruction packages and sends the scheduling instruction packages to edge nodes in the edge perception layer, the edge nodes analyze the scheduling instruction packages and complete hardware parameter configuration according to the analysis results, the edge nodes acquire sensor industrial signals according to the parameter configuration results, process the sensor industrial signals by adopting StarNet _LA network to obtain fault diagnosis results of the edge nodes, and upload the fault diagnosis results to the cloud semantic decision layer.

Inventors

  • ZHANG YAN
  • DAI LIBIN
  • HAN YAN
  • HUANG QINGQING
  • WEI MIN

Assignees

  • 重庆邮电大学

Dates

Publication Date
20260508
Application Date
20260206

Claims (8)

  1. 1. A semantic driving cloud edge collaborative monitoring method based on a narrow-band network is characterized by comprising the following steps: s1, constructing a cloud edge collaborative monitoring system, wherein the cloud edge collaborative monitoring system comprises a cloud semantic decision layer, a network transmission layer and an edge perception layer; s2, the cloud semantic decision layer receives natural language instructions and builds a semantic reasoning environment based on prompt word engineering; S3, the cloud semantic decision layer analyzes the natural language instruction, encapsulates the analysis result into a scheduling instruction packet, and sends the scheduling instruction packet to an edge node in the edge perception layer through the network transmission layer; S4, the edge node analyzes the scheduling instruction packet and completes hardware parameter configuration according to the analyzed atomization parameters; S5, the edge node collects sensor industrial signals according to the parameter configuration result, and StarNet _LA network is adopted to process the sensor industrial signals so as to obtain a fault diagnosis result of the edge node; and S6, uploading the fault diagnosis result to a cloud semantic decision layer by the edge node through a network transmission layer.
  2. 2. The semantic driving type cloud edge collaborative monitoring method based on the narrow-band network according to claim 1, wherein the construction of the semantic reasoning environment based on prompt word engineering comprises the steps of deploying a general large language model in a cloud semantic decision layer, and dynamically injecting three types of structured priori information through system prompt words, wherein the three types of structured priori information comprise a semantic-physical entity mapping table, hardware parameter space and time sequence constraint and a scene-strategy mapping rule base.
  3. 3. The semantic driving type cloud edge collaborative monitoring method based on the narrow-band network according to claim 1, wherein the process of analyzing the natural language instruction comprises the following steps: s31, scope analysis, namely analyzing an instruction subject, and locking an effective object of the instruction by inquiring an entity mapping table; s32, time sequence logic analysis, namely starting a time stamp and duration according to a time-like language analysis task in the instruction; and S33, analyzing scene parameters, namely identifying business intention keywords in the instruction, and matching corresponding duty ratio strategies in a rule base or generating a stop instruction.
  4. 4. The semantic driving cloud edge collaborative monitoring method based on a narrowband network according to claim 1, wherein the StarNet _la network comprises a plurality of stacked computing units, and a data processing process of each computing unit comprises: processing the input features through a Laplace convolution layer to obtain edge features of the signals; after the edge features of the signals are processed by the layer normalization layer, the local time sequence features are fused by the depth separable convolution layer to obtain fusion features; processing the fusion characteristics by adopting a star-shaped gate control interaction layer to obtain interaction characteristics; processing the interaction characteristics by adopting a one-dimensional efficient channel attention module to obtain channel weights; And projecting the channel attention characteristic back to the original dimension through the full connection layer, and carrying out residual addition on the channel attention characteristic and the input characteristic of the computing unit to obtain the output characteristic of the computing unit.
  5. 5. The semantic driving type cloud edge collaborative monitoring method based on the narrow-band network according to claim 4, wherein the edge characteristics of the obtained signals are expressed as: ; Wherein, the Representing the first obtained after Laplace convolution operation The edge feature output values at each instant in time, Indicating that the input raw sensor industry signal is at the first The sampled values at the individual moments in time, Representing the discrete Laplace convolution kernel at the first Weight coefficient for each location.
  6. 6. The semantic driving type cloud edge collaborative monitoring method based on the narrow-band network according to claim 4, wherein the processing of the fusion features by the star-shaped gating interaction layer is represented as: ; Wherein, the The characteristics of the interaction are represented as such, The characteristics of the fusion are represented and, 、 Representing the first and second leachable projection matrices respectively, The product of the Hadamard is represented, 、 Representing the first and second offsets, respectively.
  7. 7. The semantic driving type cloud edge collaborative monitoring method based on the narrow-band network according to claim 4, wherein the obtained channel weight is expressed as: ; Wherein, the The channel weight is represented as a function of the channel weight, Representing the Sigmoid activation function, Indicating that the convolution kernel is of size Is used for the one-dimensional convolution operation of (a), Representing pairs of interaction features And performing global average pooling operation.
  8. 8. The semantic driving type cloud edge collaborative monitoring method based on the narrowband network according to claim 7, wherein the convolution kernel size is as follows The calculation formula of (2) is expressed as: ; Wherein, the The number of channels is indicated and the number of channels is indicated, The scaling factor is represented as such, The bias factor representing the control map is used, Representing the nearest neighbor odd rounding operation.

Description

Semantic driving cloud edge collaborative monitoring method based on narrow-band network Technical Field The invention belongs to the technical field of network intelligent management, and particularly relates to a semantic driving cloud edge collaborative monitoring method based on a narrow-band network. Background The industrial Internet of things (IIoT) technology is widely applied to the fields of petrochemical industry, electric power, rail transit, precision manufacturing and the like. As a core component of IIoT perception layers, wireless sensor networks are used for carrying important tasks of collecting physical data such as equipment vibration, temperature, pressure and the like. By analyzing the data, predictive maintenance of the equipment is realized, unplanned shutdown can be effectively avoided, and operation and maintenance cost is reduced. In particular in fault diagnosis of rotating machinery such as motors, pumps, fans, gearboxes, analysis based on dither signals is considered to be the most effective means of capturing early weak faults such as bearing spalling, gear cracking. However, in practical industrial applications, wireless sensor nodes are typically battery powered and often deployed in dangerous or enclosed areas that are difficult for personnel to reach, the replacement of batteries is extremely costly. Thus, power consumption control is a primary constraint for WSN design. According to shannon's sampling theorem, in order to capture high frequency fault signatures without distortion, the sampling frequency must be at least 2 times the highest frequency of the signal. In an industrial scenario this often means that high frequency sampling on the kHz scale is required, which will produce massive amounts of data and consume large amounts of radio frequency transmission energy. To extend network life, existing mainstream schemes typically employ a "fixed low frequency polling" or "sparse periodic sampling" strategy, such as waking up one acquisition per hour for a few seconds. The stiff acquisition mode has the remarkable defects that faults of industrial equipment tend to have burstiness and transience, a fixed sampling period is extremely easy to miss a gold window period of the faults, so that a pile of useless data is acquired, a key moment is in a dormant embarrassing situation, when a production line is in high-load operation or abnormal symptoms occur, the system cannot automatically improve monitoring intensity, otherwise, when the equipment is stopped or unloaded, the node still performs meaningless routine acquisition, and energy waste is caused. On the edge computing side, in order to reduce network congestion caused by insufficient energy consumption and bandwidth due to data backhaul, the trend is to sink a fault diagnosis algorithm to the sensor nodes. However, the conventional edge intelligent algorithm has two main disadvantages that a general lightweight deep learning model for direct transplanting is usually designed based on image data, and a physical manifold structure special for an industrial time sequence signal such as periodic impact and modulation sidebands are ignored. The weak pulse feature extraction capability of the models under the strong noise background is weak, the traditional signal processing method is high in calculation complexity and difficult to run on the MCU with limited resources in real time, and the simplified statistical indexes are simple in calculation, but complex composite fault modes cannot be distinguished. In summary, there is a lack of a cloud edge cooperative system with physical perception capability in the prior art, which not only can understand the natural language scheduling intention of human beings, but also can automatically configure nodes according to semantic instructions. How to utilize the emerging Large Language Model (LLM) technology to break through the semantic control flow and combine the edge network structure optimized for the industrial signals, breaks the energy efficiency and the rigidity of the precision, and is a key technical problem to be solved in the field of the current industrial Internet of things. Disclosure of Invention Aiming at the defects existing in the prior art, the invention provides a semantic driving cloud edge collaborative monitoring method based on a narrow-band network, which comprises the following steps: s1, constructing a cloud edge collaborative monitoring system, wherein the cloud edge collaborative monitoring system comprises a cloud semantic decision layer, a network transmission layer and an edge perception layer; s2, the cloud semantic decision layer receives natural language instructions and builds a semantic reasoning environment based on prompt word engineering; S3, the cloud semantic decision layer analyzes the natural language instruction, encapsulates the analysis result into a scheduling instruction packet, and sends the scheduling instruction packet to an edge