Industrial internet edge computing environment detection method and system

Abstract

The invention relates to the technical field of computers, discloses an industrial internet edge computing environment detection method and system, and aims to solve the problems of feature extraction failure, poor model adaptability and difficulty in balancing resources and precision under strong time-varying industrial noise interference in the prior art. According to the method, a main sensor group and an environment sensor group synchronously acquire an equipment state and an environment noise signal, and a calculation mode is dynamically decided based on signal energy and spectral entropy; in a high-noise environment, a double-flow heterogeneous architecture is adopted to extract main state and noise features, noise interference is decoupled through a cross-modal attention mechanism, and pure state features are generated; and time sequence modeling and classification diagnosis are carried out in combination with a gating circulation unit network. According to the method, the detection accuracy and stability in a low signal-to-noise ratio scene are remarkably improved, and the false alarm rate is reduced.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, specifically relating to a method and system for detecting edge computing environments in the industrial internet. Background Technology

[0002] As a key carrier for the deep integration of next-generation information technology and manufacturing, the Industrial Internet's core lies in achieving real-time perception and intelligent decision-making of equipment status through edge computing. In typical industrial scenarios such as wind power, automobile manufacturing, and chemicals, edge nodes need to perform high-frequency acquisition and anomaly detection of environmental signals such as vibration, temperature, and electromagnetic fields to support predictive maintenance and safety control. However, these industrial sites are generally plagued by high-intensity noise interference, such as motor harmonics and inverter electromagnetic radiation, which severely contaminates the raw sensor data, posing a dual challenge to the reliability and stability of environmental anomaly detection.

[0003] Existing industrial internet edge computing environment detection mainly relies on two technical approaches: one is a threshold-triggered detection mechanism, which directly determines anomalies based on preset static thresholds, such as vibration acceleration exceeding 0.5 mm / s²; the other uses lightweight artificial intelligence models, such as support vector machines or shallow neural networks, deployed on resource-constrained edge devices, attempting to improve discrimination capabilities through data-driven approaches. Although the above methods have certain effectiveness under ideal operating conditions, their performance degrades significantly in high-noise dynamic environments.

[0004] Existing technologies suffer from the following drawbacks: First, noise processing mechanisms are crude, generally treating interference as filterable components and relying on hardware low-pass filtering for frequency domain truncation. However, industrial noise spectra often highly overlap with real fault characteristics, leading to the accidental deletion of critical diagnostic information during the preprocessing stage. Second, the system lacks dynamic adaptability; neither fixed thresholds nor CNN convolution kernel sizes can adaptively adjust to sudden changes in noise intensity. When interference energy increases sharply, the feature extraction module quickly fails, and there is no feedback loop for parameter reconstruction. Therefore, there is an urgent need for an innovative technical approach that achieves precise decoupling between noise and anomalies under strict resource constraints. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide an industrial internet edge computing environment detection method and system that can effectively suppress strong time-varying industrial noise interference under the condition of limited edge computing resources, and realize high-precision and low false alarm detection of industrial equipment status. This overcomes the defects of existing technologies, such as feature extraction failure caused by noise and abnormal signal spectrum aliasing, poor model dynamic adaptability, and difficulty in balancing computing resources and detection accuracy.

[0006] To address the aforementioned technical issues, this invention provides an industrial internet edge computing environment detection method. The core of this method lies in constructing a dual-stream heterogeneous processing architecture that integrates noise perception and anomaly diagnosis. By actively collecting and modeling environmental noise characteristics, utilizing a cross-modal attention mechanism to dynamically decouple the coupling relationship between noise and device status signals, and combining a phased computing resource scheduling strategy, it achieves accurate identification of device anomalies while meeting the low-power constraints of industrial edge nodes.

[0007] According to one aspect of the present invention, a method for detecting industrial internet edge computing environments is provided, specifically comprising: Data is collected synchronously by a main sensor group deployed on the target industrial equipment and an environmental sensor group deployed in the vicinity of the equipment, generating a main sensor signal sequence and an environmental accompanying signal sequence with a unified high-precision timestamp. The main sensor group includes a triaxial accelerometer, an equipment surface temperature sensor, and an input line current clamp, which are used to acquire time-domain waveform data characterizing the equipment's mechanical vibration, thermal state, and electrical load, respectively. The environmental sensor group includes a broadband acoustic sensor and a triaxial electromagnetic field sensor, which are used to acquire time-domain waveform data characterizing the background noise and electromagnetic interference in the equipment's operating environment. Synchronous acquisition is achieved through a synchronization clock module based on a hardware-accurate time protocol, ensuring that the timestamp deviation between the main sensor signal sequence and the environmental accompanying signal sequence is less than one millisecond.

[0008] After acquiring the main sensor signal sequence and the accompanying environmental signal sequence, the method further includes a dynamic decision-making step based on signal energy and complexity analysis. Specifically, the steps are as follows: the acquired signal sequence is windowed, with a window length of 256 sampling points; within each analysis window, the power spectral density of the main sensor signal and the accompanying environmental signal is calculated using a Fast Fourier Transform unit; based on the power spectral density, the root mean square energy value and spectral entropy value of the signal are calculated; a first energy threshold, a second energy threshold, a first spectral entropy threshold, and a second spectral entropy threshold are set; when the root mean square energy value of the main sensor signal is lower than the first energy threshold and the spectral entropy value is lower than the first spectral entropy threshold, the system determines it to be in a low-load silent state and activates the baseline threshold monitoring mode; when the root mean square energy value of the accompanying environmental signal is higher than the second energy threshold or the spectral entropy value is higher than the second spectral entropy threshold, the system determines it to be in a high-noise interference environment and activates the full-function dual-stream decoupling diagnostic mode; under other conditions, the system adopts the standard single-stream diagnostic mode.

[0009] When the full-function dual-stream decoupling diagnostic mode is activated, the method includes the following steps: The master sensing signal sequence is input into a master feature encoding network, which consists of three stacked one-dimensional convolutional neural network layers, batch normalization layers, and linear rectified activation function layers to extract multi-scale deep temporal features related to the device's operating state from the master sensing signal sequence and generate a master state feature vector. The convolutional kernel sizes of the three one-dimensional convolutional neural network layers are three, five, and seven, respectively, to capture signal fluctuation patterns at different time scales.

[0010] An environmental accompanying signal sequence is input into an accompanying feature coding network. The accompanying feature coding network has the same network structure and parameter configuration as the main feature coding network. It is used to extract multi-scale temporal features that characterize the current environmental noise in the environmental accompanying signal sequence and generate an environmental noise feature vector.

[0011] The main state feature vector and the environmental noise feature vector are input into a cross-modal attention decoupling module. Inside the module, the environmental noise feature vector is used as the query vector, and the main state feature vector is used as both the key vector and the value vector. By calculating the dot product attention weights of the query vector and the key vector, and normalizing the weights, the value vectors are weighted and summed to generate a clean state feature vector that is dynamically adjusted by the noise context information. This process is mathematically equivalent to a scaled dot product attention operation, which aims to adaptively suppress the artifact components introduced by noise in the main state feature vector by utilizing the feature distribution of environmental noise, while enhancing the feature representation related to the true state of the device.

[0012] The clean state feature vector is concatenated with the original master state feature vector to form an enhanced fusion feature vector. This fusion feature vector contains both the original information of the device state and the clean information after decoupling from the noise information.

[0013] The enhanced fused feature vector sequence is input into a gated recurrent unit network. The gated recurrent unit network models the temporal dependency of the fused feature vector sequence through its internal update gate and reset gate structure, captures the evolution of the device state in the time dimension, and outputs a context state vector containing historical state information.

[0014] Finally, the context state vector is input into a fully connected classification layer. The fully connected classification layer calculates the posterior probability of different equipment state categories through a flexible maximum function and outputs the final equipment state diagnosis result. The equipment state categories include normal operation, bearing wear, rotor imbalance, electrical fault, and other predefined fault types.

[0015] As one embodiment of the present invention, the baseline threshold monitoring mode specifically involves comparing the root mean square value, peak-to-peak value, and kurtosis index of the main sensor signal sequence with the preset statistical baseline threshold under normal operating conditions of the device. When any index exceeds 20 percent of its corresponding threshold range, an early warning signal is generated, and the system automatically switches to the full-function dual-stream decoupling diagnostic mode for in-depth analysis.

[0016] As one embodiment of the present invention, the standard single-stream diagnostic mode is as follows: when the system determines that the environment is not a high-noise interference environment in the dynamic decision-making step of the calculation mode, the system will disable the environmental sensor group and the accompanying feature coding network, and only input the main sensing signal sequence into the main feature coding network, and directly input the generated main state feature vector into the gated cyclic unit network for subsequent timing state diagnosis, thereby reducing the overall computing load and power consumption of the system in this mode while ensuring detection accuracy.

[0017] According to another aspect of the present invention, an industrial internet edge computing environment detection system is provided, the system being integrated onto a low-power embedded system-on-a-chip, the system comprising: A dual-channel synchronous data acquisition module integrates a main sensor group, an environmental sensor group, and a synchronization clock module based on a hardware precision time protocol. The dual-channel synchronous data acquisition module is configured to perform the synchronous data acquisition steps in the aforementioned method.

[0018] A dynamic computing resource allocation module is electrically connected to a dual-channel synchronous data acquisition module and receives the signal sequence output by the module. The dynamic computing resource allocation module has a built-in hardware acceleration circuit for calculating the root mean square energy value and spectral entropy value of the signal, and includes a state machine logic unit configured to execute the dynamic decision-making steps of the computing mode in the aforementioned method, and generate system mode control instructions based on the decision results.

[0019] A signal decoupling and feature extraction module is electrically connected to a dual-channel synchronous data acquisition module and a dynamic computing resource allocation module. The signal decoupling and feature extraction module includes a main feature encoding unit, an adjoint feature encoding unit, and a cross-modal attention decoupling unit. The main feature encoding unit and the adjoint feature encoding unit implement the computational logic of the main feature encoding network and the adjoint feature encoding network, respectively. The cross-modal attention decoupling unit implements the computational logic of the cross-modal attention decoupling module. The signal decoupling and feature extraction module selectively activates all or some of its internal units according to the system mode control instructions issued by the dynamic computing resource allocation module.

[0020] A timing state diagnostic module is electrically connected to the signal decoupling and feature extraction module, and receives the enhanced fusion feature vector sequence output by the module. The timing state diagnostic module integrates a gated recurrent unit network processing core and a fully connected layer computing array, and is configured to perform the timing modeling and classification diagnostic steps in the aforementioned method, and output the final equipment state diagnostic result.

[0021] All modules exchange data through a high-speed on-chip system bus. The overall operation of the system is coordinated and scheduled by the dynamic computing resource allocation module to achieve the optimal balance between computing accuracy and power consumption under different operating conditions.

[0022] In summary, this application includes at least one of the following beneficial technical effects: (1): By introducing an environmental sensor group that runs parallel to the main sensor and constructing a dual-stream feature coding network, this invention fundamentally changes the technical paradigm of treating noise as passive interference and filtering it out. Instead, it transforms into actively sensing, modeling, and using noise features to decouple the signal. This effectively solves the technical problem of high coupling and difficulty in separating noise and abnormal equipment signals in the frequency and time domains in industrial sites, and significantly improves the detection accuracy in low signal-to-noise ratio environments.

[0023] (1): The cross-modal attention decoupling module designed in this invention can dynamically and nonlinearly adjust the attention weight of the main sensor signal features according to the real-time acquired environmental noise characteristics, thus achieving a high degree of adaptability to time-varying noise environments. Compared with methods using fixed parameter filtering or static models, this invention can better cope with the complex situation of sudden changes in noise intensity and spectral characteristics in industrial sites, ensuring the stability of the diagnostic model.

[0024] (1): This invention proposes a dynamic decision-making mechanism for computing modes based on signal energy and complexity analysis, and constructs a multi-level processing architecture that includes three modes: baseline monitoring, standard diagnosis, and dual-stream decoupling. This mechanism enables the system to intelligently schedule computing resources according to the signal-to-noise ratio of the real-time operating environment, operate with extremely low power consumption when the environment is quiet, and only call all computing resources for deep decoupling analysis when high-intensity interference is detected. This fundamentally solves the contradiction between the limited computing resources of edge computing devices and the high-precision diagnostic requirements, and achieves synergistic optimization of power consumption and performance, meeting the stringent deployment requirements of industrial IoT edge nodes.

[0025] (1): By using a gated cyclic unit network in the diagnostic backend, the present invention can effectively capture the dynamic law of the evolution of equipment status characteristics over time, and combine instantaneous features with historical context information for comprehensive judgment. Compared with the static classification model that only relies on instantaneous data points, it can more accurately distinguish between occasional signal spikes and continuously developing fault precursors, further reducing the false alarm rate and false alarm rate of the system. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the principle framework for parallel noise perception and anomaly diagnosis in this invention; Figure 3 This is a logical flow diagram of the present invention based on signal energy and complexity analysis. Detailed Implementation

[0027] This invention provides a method and system for detecting industrial internet edge computing environments, aiming to address the technical shortcomings of existing technologies, such as high false alarm rates, poor dynamic adaptability of models, and difficulty in balancing resources and accuracy under strong time-varying industrial noise interference. By constructing a dual-stream heterogeneous processing architecture that integrates noise perception and anomaly diagnosis, high-precision, low-false-alarm detection of device status is achieved while meeting the low-power constraints of edge nodes.

[0028] See attached document Figure 1 To be continued Figure 3 The present invention discloses an industrial internet edge computing environment detection method, which is described below with reference to the accompanying drawings.

[0029] First, a synchronous data acquisition step is performed. Data is synchronously acquired by the main sensor group deployed on the target industrial equipment and the environmental sensor group deployed in the vicinity of the equipment under a unified high-precision timestamp, generating a main sensor signal sequence and an accompanying environmental signal sequence.

[0030] The main sensor group includes a triaxial accelerometer, a device surface temperature sensor, and an input line current clamp, which are used to acquire time-domain waveform data characterizing the device's mechanical vibration, thermal state, and electrical load, respectively.

[0031] The environmental sensor group includes a broadband acoustic sensor and a triaxial electromagnetic field sensor, used to acquire time-domain waveform data characterizing background noise and electromagnetic interference in the operating environment of the equipment.

[0032] Synchronous acquisition is achieved through a synchronization clock module based on a hardware-precise time protocol, ensuring that the timestamp deviation between the main sensor signal sequence and the accompanying environmental signal sequence is less than 1 millisecond. This synchronization mechanism eliminates data misalignment caused by clock drift between sensors by sending a synchronization trigger pulse to all sensors before the start of each sampling period and calibrating the local oscillator using a hardware phase-locked loop.

[0033] After completing synchronous data acquisition, the system enters the dynamic decision-making step based on signal energy and complexity analysis. This step performs windowing processing on the acquired signal sequence, with a window length of 256 sampling points, a sampling frequency of 1 kHz, and a corresponding analysis window duration of 256 milliseconds.

[0034] Within each analysis window, the system calculates the power spectral density of the main sensing signal and the accompanying environmental signal using a Fast Fourier Transform (FFT) unit. Based on the power spectral density, the system calculates the root-mean-square (RMS) energy and spectral entropy of the signal. The RMS energy is defined as the square root of the power spectral density integral over the entire frequency band, used to measure the overall energy intensity of the signal; the spectral entropy is defined as the Shannon entropy normalized to the power spectral density, used to measure the complexity and randomness of the signal spectrum.

[0035] The system presets the first energy threshold to be 0.05 volts RMS, the second energy threshold to be 0.2 volts RMS, the first spectral entropy threshold to be 1.2, and the second spectral entropy threshold to be 2.5.

[0036] When the root mean square energy value of the main sensor signal is lower than the first energy threshold and the spectral entropy value is lower than the first spectral entropy threshold, the system determines that the current device is in a low-load silent state and enables the baseline threshold monitoring mode; when the root mean square energy value of the environmental accompanying signal is higher than the second energy threshold or the spectral entropy value is higher than the second spectral entropy threshold, the system determines that the current environment is a high-noise interference environment and activates the full-function dual-stream decoupling diagnostic mode; under other conditions, the system adopts the standard single-stream diagnostic mode.

[0037] In baseline threshold monitoring mode, the system compares the root mean square (RMS), peak-to-peak (peak-to-peak), and kurtosis indices of the main sensor signal sequence with preset statistical baseline thresholds under normal equipment operation. The statistical baseline thresholds are calculated from data collected during 72 hours of continuous trouble-free operation of the equipment during the factory testing phase, with the RMS baseline threshold at 0.03 volts, the peak-to-peak baseline threshold at 0.15 volts, and the kurtosis baseline threshold at 3.0.

[0038] When any indicator exceeds its corresponding threshold range by 20%, the system generates an early warning signal and automatically switches to the full-function dual-stream decoupled diagnostic mode for in-depth analysis. In this mode, the environmental sensor group stops working, the main feature encoding network runs at the lowest frequency, and only performs basic feature extraction and threshold comparison, keeping the overall system power consumption below 1.5W.

[0039] In standard single-stream diagnostic mode, the system disables the environmental sensor group and the accompanying feature coding network, and only inputs the main sensor signal sequence to the main feature coding network.

[0040] The main feature encoding network consists of a stacked 3-dimensional convolutional neural network layer, a batch normalization layer, and a linear rectified activation function layer. The first convolutional layer has a kernel size of 3, a stride of 1, and 16 output channels; the second convolutional layer has a kernel size of 5, a stride of 1, and 32 output channels; and the third convolutional layer has a kernel size of 7, a stride of 1, and 64 output channels. Each convolutional operation is followed by a batch normalization layer and a linear rectified activation function layer to accelerate convergence and enhance nonlinear expressive power.

[0041] This network extracts multi-scale deep temporal features related to the device's operating state from the master sensor signal sequence, generating a 64-dimensional master state feature vector. This master state feature vector is then directly input into a gated recurrent unit network for subsequent temporal state diagnosis. This mode is suitable for conventional operating conditions with a signal-to-noise ratio greater than 20 dB, and the system power consumption is approximately 3.8 W, meeting the national standard that the standby power consumption of industrial edge nodes should not exceed 5 W.

[0042] When the system determines that the environment is subject to high noise interference, the full-function dual-stream decoupling diagnostic mode is activated. In this mode, the main sensor signal sequence is input to the main feature encoding network, and the accompanying environmental signal sequence is input to the accompanying feature encoding network. The accompanying feature encoding network has the same network structure and parameter configuration as the main feature encoding network, including three layers of one-dimensional convolutional neural networks, batch normalization layers, and linear rectified activation function layers, with convolutional kernel sizes of 3, 5, and 7, and output channel numbers of 16, 32, and 64, respectively. The two networks operate independently. The main feature encoding network outputs a 64-dimensional main state feature vector, and the accompanying feature encoding network outputs a 64-dimensional environmental noise feature vector.

[0043] Subsequently, the main state feature vector and the environmental noise feature vector are input into the cross-modal attention decoupling module. Within this module, the environmental noise feature vector is used as the query vector Q, and the main state feature vector serves as both the key vector K and the value vector V. The system calculates the dot product attention weight between the query vector and the key vector, with the mathematical expression as follows: in, The feature vector dimension is 64. This formula calculates the attention weights of environmental noise features on each dimension of the main state features using a scaled dot product attention mechanism. After normalization, the weights are weighted and summed to generate a clean state feature vector dynamically adjusted by noise context information. This process utilizes the feature distribution of environmental noise to adaptively suppress artifacts introduced by noise in the main state feature vector, while enhancing the feature representation related to the true state of the device.

[0044] The generated clean state feature vector is concatenated with the original master state feature vector to form an enhanced fusion feature vector with a dimension of 128. This fusion feature vector retains the original information of the device state while incorporating the cleaned-up information decoupled from noise, providing a more stable input for subsequent time series modeling.

[0045] The enhanced fused feature vector sequence is input into a gated recurrent unit (GRU) network. The GRU network consists of two stacked layers, each with 128 hidden units. The GRU models the temporal dependencies of the fused feature vector sequence through its internal update and reset gates. The update gate determines how much information in the current hidden state comes from the previous hidden state, while the reset gate controls the influence of the previous hidden state on the current candidate hidden state. This mechanism effectively captures the evolution of the device state over time, outputting a context state vector containing historical state information.

[0046] Finally, the context state vector is input to the fully connected classification layer. The fully connected classification layer contains two fully connected sub-layers. The first sub-layer has an output dimension of 64, and the second sub-layer has an output dimension of 5, corresponding to five equipment state categories: normal operation, bearing wear, rotor imbalance, electrical fault, and other predefined fault types. The second sub-layer is followed by a flexible maximum function to calculate the posterior probability of each category. When the posterior probability of any fault category exceeds a preset threshold of 0.9, the system determines that the equipment is in that fault state and generates a corresponding alarm; if the probabilities of all categories are below this threshold, it is determined to be in normal operation.

[0047] The execution of the aforementioned industrial internet edge computing environment detection method in this application relies on an integrated industrial internet edge computing environment detection system. This system is integrated onto a low-power embedded system-on-a-chip and includes a dual-channel synchronous data acquisition module, a dynamic computing resource allocation module, a signal decoupling and feature extraction module, and a timing status diagnosis module.

[0048] The dual-channel synchronous data acquisition module integrates a main sensor group, an environmental sensor group, and a synchronization clock module based on a hardware-based precise time protocol. This module amplifies, filters, and performs analog-to-digital conversion on the raw sensor signals using a dedicated analog front-end circuit, achieving a 16-bit conversion accuracy and a sampling rate of 1000 Hz. The synchronization clock module is implemented in hardware using the IEEE 1588 precise time protocol, ensuring time alignment accuracy of the dual-channel data is better than 0.5 milliseconds through a physical layer timestamp mechanism.

[0049] The dynamic computing resource allocation module is electrically connected to the dual-channel synchronous data acquisition module, receiving its output digital signal sequence. Internally, this module incorporates hardware acceleration circuitry for calculating the root-mean-square energy and spectral entropy of the signals, including a dedicated Fast Fourier Transform coprocessor and entropy calculation logic unit. The module also includes a state machine logic unit that generates system mode control commands based on preset threshold comparison results, controlling the power supply status and operating frequency of other modules. For example, in baseline threshold monitoring mode, this module shuts off the power supply to the environmental sensors and reduces the operating frequency of the main feature encoding network to 10 MHz; in full-function dual-stream decoupling diagnostic mode, all modules operate at the highest frequency.

[0050] The signal decoupling and feature extraction module is electrically connected to the dual-channel synchronous data acquisition module and the dynamic computing resource allocation module. This module includes a main feature encoding unit, an adjoint feature encoding unit, and a cross-modal attention decoupling unit. The main and adjoint feature encoding units are implemented using dedicated neural network accelerators, supporting hardware pipelined processing of one-dimensional convolution, batch normalization, and activation functions. The cross-modal attention decoupling unit consists of a matrix multiplication array and a softmax computation unit, dedicated to performing scaled dot product attention operations. Based on system mode control commands issued by the dynamic computing resource allocation module, this module selectively activates all or some of its internal units, achieving on-demand allocation of computing resources.

[0051] The timing status diagnostic module is electrically connected to the signal decoupling and feature extraction module, receiving its output enhanced fusion feature vector sequence. Internally, this module integrates a gated recurrent unit network processing core and a fully connected layer computing array. The gated recurrent unit processing core employs an architecture combining loop unrolling and pipelined processing, supporting parallel computation across two layers of 128 units. The fully connected layer computing array consists of a weight memory, multiply-accumulate units, and activation function units, supporting hardware implementation of a flexible maximum function. This module ultimately outputs the device status diagnostic results and uploads them to a cloud monitoring platform via an industrial Ethernet interface or wireless communication module.

[0052] All modules exchange data via a high-speed on-chip system bus with a bandwidth of 8 gigabits per second, ensuring low-latency transmission of feature data between modules. The overall system operation is coordinated and scheduled by a dynamic computing resource allocation module, which intelligently switches operating modes based on the signal-to-noise ratio of the real-time operating environment. It operates with extremely low power consumption when the environment is quiet, and calls upon all computing resources for deep decoupling analysis when there is high-intensity interference, thereby fundamentally resolving the contradiction between the limited computing resources of edge computing devices and the high-precision diagnostic requirements.

[0053] In practical deployments, this system can be installed near critical equipment such as wind turbine gearboxes, industrial motor bearing housings, or high-voltage switchgear. After startup, the system first undergoes a 24-hour self-learning phase, during which data is continuously collected and a statistical baseline of the equipment's normal operating status is established. After the self-learning phase, the system enters online monitoring mode and performs real-time detection according to the method and process described in this application. When a potential fault is detected, the system not only generates an alarm but also packages and stores the original signal fragments, feature vectors, and diagnostic data for subsequent fault tracing and analysis.

[0054] The method and system of this application effectively overcome the performance bottleneck of traditional solutions under noise interference by actively sensing environmental noise, dynamically decoupling signal coupling, and scheduling computing resources in stages.

[0055] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0056] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An industrial internet edge computing environment detection method, characterized in that, The method comprises the following steps: Synchronously collecting data through a main sensor group deployed on a target industrial equipment body and an environment sensor group deployed in a space adjacent to the equipment, to generate a main sensing signal sequence and an environment accompanying signal sequence with uniform high-precision timestamps; Performing windowing processing on the collected main sensing signal sequence and environment accompanying signal sequence, calculating the power spectral density of the main sensing signal and the environment accompanying signal in each analysis window through a fast Fourier transform unit, and calculating the root mean square energy value and the spectral entropy value of the signal based on the stability power spectral density; setting a first energy threshold, a second energy threshold, a first spectral entropy threshold, and a second spectral entropy threshold; when the root mean square energy value of the main sensing signal is lower than the first energy threshold and the spectral entropy value is lower than the first spectral entropy threshold, enabling a baseline threshold monitoring mode; when the root mean square energy value of the environment accompanying signal is higher than the second energy threshold or the spectral entropy value is higher than the second spectral entropy threshold, activating a full-function dual-flow decoupling diagnosis mode; otherwise, adopting a standard single-flow diagnosis mode; Under the condition that the stability full-function dual-flow decoupling diagnosis mode is activated, inputting the stability main sensing signal sequence into a main feature coding network to extract multi-scale deep time sequence features and generate a main state feature vector; inputting the stability environment accompanying signal sequence into an accompanying feature coding network to extract environment noise features and generate an environment noise feature vector; Inputting the stability main state feature vector and the stability environment noise feature vector into a cross-modal attention decoupling module to generate a pure state feature vector; Concatenating the stability pure state feature vector and the original main state feature vector to form an enhanced fusion feature vector; Inputting the stability enhanced fusion feature vector sequence into a gated recurrent unit network to model time dependence and output a context state vector; Inputting the stability context state vector into a fully connected classification layer to calculate the posterior probability of different equipment state categories through a flexible maximum value function and output an equipment state diagnosis result. 2.The industrial internet edge computing environment detection method of claim 1, wherein, The baseline threshold monitoring mode specifically comprises: comparing the root mean square value, the peak-to-peak value, and the kurtosis index of the main sensing signal sequence with the statistical baseline threshold preset in the normal running state of the equipment; when any index exceeds 20% of the corresponding threshold range, a warning signal is generated, and the full-function dual-flow decoupling diagnosis mode is automatically switched to for in-depth analysis. 3.The industrial internet edge computing environment detection method of claim 1, wherein, The standard single-flow diagnosis mode specifically comprises: disabling the environment sensor group and the accompanying feature coding network, inputting only the main sensing signal sequence into the main feature coding network, and directly inputting the generated main state feature vector into the gated recurrent unit network for subsequent time sequence state diagnosis.

4. The industrial internet edge computing environment detection method of claim 1, wherein, In the main feature coding network, the convolution kernel size of the first one-dimensional convolutional neural network layer is 3, the output channel number is 16, the convolution kernel size of the second layer is 5, the output channel number is 32, the convolution kernel size of the third layer is 7, and the output channel number is 64; after each convolution operation, a batch normalization layer and a linear rectifier activation function layer are connected.

5. The industrial internet edge computing environment detection method of claim 4, wherein, The accompanying feature coding network has the same network structure and parameter configuration as the main feature coding network, and outputs an environment noise feature vector with a dimension of 64.

6. The industrial internet edge computing environment detection method of claim 1, wherein, In the cross-modal attention decoupling module, the calculation formula of scaled dot-product attention is as follows: Wherein, Q is an environmental noise feature vector, K and V are main state feature vectors, is a feature vector dimension, and the value is 64.

7. The industrial internet edge computing environment detection method of claim 1, wherein, The dimension of the enhanced fusion feature vector is 128, which is spliced by the pure state feature vector with a dimension of 64 and the main state feature vector with a dimension of 64.

8. The industrial internet edge computing environment detection method of claim 1, wherein, The gated recurrent unit network comprises a two-layer stacked structure, each layer having 128 hidden units, for modeling the time dependence of the fusion feature vector sequence and outputting a context state vector containing historical state information. 9.An industrial internet edge computing environment detection system, characterized in that, The method comprises the following steps: a dual-channel synchronous data acquisition module is used to synchronously acquire data through a main sensor group deployed on the target industrial equipment body and an environmental sensor group deployed in the space adjacent to the equipment, to generate a main sensing signal sequence and an environmental accompanying signal sequence with uniform high-precision timestamps; the main sensor group comprises a three-axis acceleration sensor, a device surface temperature sensor and an input line current clamp; the environmental sensor group comprises a wideband acoustic sensor and a three-axis electromagnetic field sensor; synchronous acquisition is realized through a hardware-based precision time protocol-based synchronous clock module, ensuring that the timestamp deviation is less than one millisecond; a dynamic computing resource allocation module is used to perform window processing on the main sensing signal sequence and the environmental accompanying signal sequence and calculate the root mean square energy value and the spectral entropy value, to determine the system operation mode according to a preset threshold, and to generate a system mode control instruction; a signal decoupling and feature extraction module is used to extract a main state feature vector and an environmental noise feature vector through a main feature encoding unit and an accompanying feature encoding unit respectively in a full-function dual-flow decoupling diagnosis mode, and to generate a pure state feature vector through a cross-modal attention decoupling unit, thereby forming an enhanced fusion feature vector; a time series state diagnosis module is used to input the enhanced fusion feature vector sequence into a gated recurrent unit network for time series modeling, and to output a device state diagnosis result through a fully connected classification layer.

Images