Edge device fault diagnosis method and system in port complex environment

Abstract

The invention discloses an edge device fault diagnosis method and system in a port complex environment, and relates to the technical field of fault detection, and the method comprises the steps: continuously collecting temperature, vibration and noise multi-source heterogeneous data, and carrying out the lightweight preprocessing and dimension reduction processing to obtain a dimension reduction data set; in combination with an equipment damage mechanism, a three-level coefficient is obtained through multi-dimensional fault correlation analysis; and constructing a lightweight deep learning model based on ensemble learning, fusing the multi-source dimension reduction data and the three-level coefficient to carry out localized real-time reasoning, and outputting a fault mode. Through anti-interference data acquisition equipment and robustness algorithm optimization, the influence of strong electromagnetic noise and vibration interference on a diagnosis result is effectively resisted, and the diagnosis accuracy in a complex environment is improved; a multi-source heterogeneous data embedded feature fusion method is constructed based on an equipment damage mechanism, complementary information of temperature, vibration and noise data is fully utilized, and comprehensiveness and accuracy of fault diagnosis are improved through three-level fault correlation analysis.

Description

Technical Field

[0001] This invention relates to the field of fault detection technology, and more specifically to a method and system for diagnosing faults in edge devices in complex port environments. Background Technology

[0002] Port edge equipment is a core infrastructure for port logistics operations, and its operational status directly affects port operational efficiency and safety. The port environment is characterized by complex features such as strong electromagnetic noise, severe vibration interference, and large temperature and humidity variations, leading to diverse, complex, and sudden failures of edge equipment.

[0003] Existing edge device fault diagnosis technologies mainly suffer from the following problems: Relying on single sensor data or simple parameter monitoring fails to fully utilize the complementary information from multi-source heterogeneous data, resulting in low diagnostic accuracy. The algorithm lacks robustness and is susceptible to data interference in scenarios with strong electromagnetic noise and vibration in ports, leading to high false alarm and false negative rates. Existing deep learning models have high computational complexity, making them unsuitable for the limited computing resources of edge devices and hindering localized real-time inference. Furthermore, the lack of fault correlation analysis targeting the damage mechanisms of port edge devices results in a low degree of matching between fault diagnosis and the actual operating status of the equipment.

[0004] Therefore, how to propose a fault diagnosis method and system for edge devices in complex port environments and overcome the shortcomings of existing technologies is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a method and system for fault diagnosis of edge devices in complex port environments, solving the problems of low accuracy, insufficient robustness, and inability to perform localized real-time inference in existing technologies for edge device fault diagnosis in complex port environments. To achieve the above objectives, the present invention adopts the following technical solution: A method for fault diagnosis of edge devices in complex port environments includes: Collect temperature, vibration, and current information from edge devices to obtain temperature, vibration, and current information sets. The temperature information set, vibration information set, and current information set are respectively subjected to dimensionality reduction processing to obtain the dimensionality-reduced temperature information set, dimensionality-reduced vibration information set, and dimensionality-reduced current information set; Correlation analysis was performed on each pairwise set of reduced-dimensional temperature information, reduced-dimensional vibration information, and reduced-dimensional current information to obtain multiple correlation coefficients. A lightweight deep learning model is constructed, and the reduced-dimensional temperature information set, reduced-dimensional vibration information set, and reduced-dimensional current information set are respectively input into the corresponding fault identification branch in the model. The inference weight of each fault identification branch is determined based on multiple correlation analysis coefficients. Multi-source data fusion inference is performed through embedded feature fusion method to obtain fault diagnosis results.

[0006] Optionally, the temperature information collected by the edge device includes: When the port edge equipment is started, temperature information at key locations on the edge equipment is collected at multiple time points to obtain a temperature array set; To address the abnormal temperature data caused by strong electromagnetic interference at the port, outlier values ​​were removed and data was calibrated for the temperature array at each time point. Based on the calibrated temperature array, temperature interpolation is performed on all critical areas of the edge device to obtain a set of temperature information.

[0007] Optionally, the dimensionality reduction process includes: The baseline core temperature is calculated based on the statistical characteristics of the temperature information set. Based on the temperature difference between other key temperatures and the benchmark core temperature in the temperature information set, and combined with the temperature change pattern of equipment in the port environment, the distribution probability is calculated to obtain the key probability distribution. A lightweight dimensionality reduction algorithm adapted to edge devices is used to reduce the dimensionality of the temperature information set according to the key probability distribution, thereby obtaining a dimensionality-reduced temperature information set. The time nodes of multiple key temperature information in the reduced-dimensional temperature information set are obtained, and multi-source data of the corresponding time nodes in the vibration information set and current information set are extracted. Features strongly correlated with temperature changes are extracted through embedded feature fusion method to obtain the reduced-dimensional vibration information set and the reduced-dimensional current information set.

[0008] Optionally, the lightweight dimensionality reduction algorithm adapted to edge devices is used to reduce the dimensionality of the temperature information set according to the key probability distribution, and the resulting dimensionality-reduced temperature information set includes: Set the preset dimensionality reduction ratio and preset number of dimensionality reductions to adapt to the computing capabilities of edge devices; According to the preset dimensionality reduction ratio, key temperatures are randomly selected from the temperature information set to obtain the first dimensionality-reduced temperature information set. Based on the temperature difference between multiple first-dimensionality-reduced key temperatures and the benchmark core temperature in the first-dimensionality-reduced temperature information set, the allocation probability is calculated to obtain the first-dimensionality-reduced key probability distribution. Calculate the similarity between the first key probability distribution and the key probability distribution to obtain the first key dimension reduction; Continue to randomly sample and reduce the dimensionality according to the preset dimensionality reduction ratio to obtain the second dimensionality-reduced temperature information set, and process it to obtain the second key dimensionality reduction; Repeat the random sampling dimensionality reduction process until the preset number of dimensionality reductions is reached, and output the set of dimensionality-reduced temperature information that maximizes the key dimensionality reduction and meets the computing latency requirements of edge devices.

[0009] Optionally, the correlation analysis is performed pairwise on the reduced-dimensional temperature information set, the reduced-dimensional vibration information set, and the reduced-dimensional current information set to obtain multiple correlation coefficients, including: Based on the reduced-dimensional temperature information set and the reduced-dimensional current information set, a first-level fault correlation analysis is performed based on the equipment thermal damage mechanism to obtain the first coefficient. Using the reduced-dimensional vibration information set and the reduced-dimensional temperature information set, a secondary fault correlation analysis is performed based on the equipment damage mechanism to obtain the second coefficient; Using the aforementioned reduced-dimensional current information set and reduced-dimensional vibration information set, a three-level fault correlation analysis is performed for strong electromagnetic noise interference scenarios to obtain the third coefficient.

[0010] Optionally, the step of performing a first-level fault correlation analysis based on the reduced-dimensional temperature information set and the reduced-dimensional current information set to obtain the first coefficient includes: Delete the initial unstable stage data in the reduced temperature information set and retain the valid data after the equipment is running stably. Based on the reduced-dimensional current information set, a first-level fault correlation analysis was performed according to the equipment thermal damage mechanism, and the first coefficient was calculated using the Pearson correlation coefficient algorithm: ; Where r is the first coefficient, and m is the number of valid data points in the reduced-dimensional temperature information set and the reduced-dimensional current information set. For the i-th key temperature in the set of reduced-dimensional temperature information, For the deformation parameters of the i-th part within the reduced-dimensional current information set, and These are the mean values ​​of the reduced-dimensional temperature information set and the reduced-dimensional current information set, respectively.

[0011] Optionally, constructing a lightweight deep learning model includes: Historical fault diagnosis data of edge devices in complex port environments are collected to obtain sample reduced-dimensional temperature information set, sample reduced-dimensional vibration information set, sample reduced-dimensional current information set, and corresponding sample fault mode set. The sample information includes labeled data under strong electromagnetic noise and vibration interference scenarios. The sample reduced-dimensional temperature information set, sample reduced-dimensional vibration information set, and sample reduced-dimensional current information set are respectively used as inputs, and the sample fault mode set is used as output. Based on the ensemble learning algorithm, temperature fault identification branch, vibration fault identification branch, and noise fault identification branch are constructed. Each fault identification branch adopts a lightweight network structure, and the quantization compression optimization algorithm improves the model's running efficiency on edge devices.

[0012] Optionally, the fault identification branch includes multiple fault identification paths, each optimized for different types of damage faults; It integrates four fault identification branches and embeds a multi-source heterogeneous data feature fusion module to obtain a lightweight deep learning model that supports local real-time inference on edge devices.

[0013] Optionally, the step of performing multi-source data fusion inference through an embedded feature fusion method to obtain fault diagnosis results includes: Based on the first coefficient, the second coefficient, and the third coefficient, determine the first inference weight, the second inference weight, and the third inference weight; The reduced-dimensional temperature information set, reduced-dimensional vibration information set, and reduced-dimensional current information set are respectively input into the temperature fault identification branch, vibration fault identification branch, and noise fault identification branch. Computational resources are allocated according to the corresponding inference weights to perform fault analysis and output multiple basic fault modes. The embedded feature fusion module integrates all basic fault modes, selects the most frequent basic fault modes that conform to the equipment damage mechanism, and uses them as the final fault modes, including fault type, damage degree, fault location and maintenance suggestions.

[0014] Optionally, an edge device fault diagnosis system for complex port environments includes: Acquisition module: Used to acquire temperature, vibration and current information from edge devices, and obtain sets of temperature, vibration and current information. Dimensionality reduction module: used to perform dimensionality reduction processing on the temperature information set, vibration information set, and current information set respectively, to obtain dimensionality-reduced temperature information set, dimensionality-reduced vibration information set, and dimensionality-reduced current information set; The correlation analysis module is used to perform pairwise correlation analysis on the reduced-dimensional temperature information set, the reduced-dimensional vibration information set, and the reduced-dimensional current information set to obtain multiple correlation analysis coefficients. Inference module: Used to build lightweight deep learning models. It inputs the reduced temperature information set, reduced vibration information set, and reduced current information set into the corresponding fault identification branches in the model, and determines the inference weights of each fault identification branch based on multiple correlation analysis coefficients. Fault diagnosis module: Used to perform multi-source data fusion inference through embedded feature fusion method to obtain fault diagnosis results.

[0015] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method and system for diagnosing edge equipment faults in complex port environments, which has the following beneficial effects: This invention proposes a fault diagnosis method for edge equipment in complex port environments, comprising: collecting temperature, vibration, and current information from the edge equipment to obtain temperature, vibration, and current information sets; performing dimensionality reduction on the temperature, vibration, and current information sets respectively to obtain dimensionality-reduced temperature, vibration, and current information sets; performing pairwise correlation analysis on each of the dimensionality-reduced temperature, vibration, and current information sets to obtain multiple correlation coefficients; constructing a lightweight deep learning model, inputting the dimensionality-reduced temperature, vibration, and current information sets into the corresponding fault identification branches of the model, determining the inference weights of each fault identification branch based on the multiple correlation coefficients, and performing multi-source data fusion inference through an embedded feature fusion method to obtain the fault diagnosis result.

[0016] This invention continuously collects multi-source heterogeneous data on temperature, vibration, and noise after the edge device is started. After lightweight preprocessing and dimensionality reduction, a dimensionality-reduced dataset is obtained. Combining this with the device damage mechanism, a three-level coefficient is obtained through multi-dimensional fault correlation analysis. A lightweight deep learning model is constructed based on ensemble learning, fusing the multi-source dimensionality-reduced data and the three-level coefficients for localized real-time inference, outputting the fault mode. This invention achieves localized real-time fault diagnosis for edge devices by optimizing algorithm robustness and using embedded feature fusion methods, improving the accuracy and efficiency of fault detection in complex environments. By optimizing anti-interference data acquisition equipment and robust algorithms, the system effectively resists the impact of strong electromagnetic noise and vibration interference on diagnostic results, improving diagnostic accuracy in complex environments. Lightweight preprocessing algorithms, dimensionality reduction techniques, and model optimization methods are employed to construct a deep learning model adapted to edge devices, enabling localized real-time inference with an inference latency of less than 100ms, meeting the real-time monitoring needs of port equipment. Based on equipment damage mechanisms, a multi-source heterogeneous data embedded feature fusion method is constructed, fully utilizing the complementary information from temperature, vibration, and noise data. Three-level fault correlation analysis enhances the comprehensiveness and accuracy of fault diagnosis. Fault diagnosis results include specific fault types, damage levels, locations, and maintenance recommendations, directly guiding port equipment maintenance, reducing downtime, and improving port operational efficiency and safety. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This invention provides a schematic flowchart of a fault diagnosis method for edge devices in a complex port environment. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This invention discloses a method for diagnosing faults in edge devices in complex port environments, such as... Figure 1 As shown, it includes: Collect temperature, vibration, and current information from edge devices to obtain temperature, vibration, and current information sets. The temperature information set, vibration information set, and current information set are respectively subjected to dimensionality reduction processing to obtain the dimensionality-reduced temperature information set, dimensionality-reduced vibration information set, and dimensionality-reduced current information set; Correlation analysis was performed on each pairwise set of reduced-dimensional temperature information, reduced-dimensional vibration information, and reduced-dimensional current information to obtain multiple correlation coefficients. A lightweight deep learning model is constructed, and the reduced-dimensional temperature information set, reduced-dimensional vibration information set, and reduced-dimensional current information set are respectively input into the corresponding fault identification branch in the model. The inference weight of each fault identification branch is determined based on multiple correlation analysis coefficients. Multi-source data fusion inference is performed through embedded feature fusion method to obtain fault diagnosis results.

[0021] Furthermore, the temperature information acquired by the edge device includes: When the port edge equipment is started, temperature information at key locations on the edge equipment is collected at multiple time points to obtain a temperature array set; To address the abnormal temperature data caused by strong electromagnetic interference at the port, outlier values ​​were removed and data was calibrated for the temperature array at each time point. Based on the calibrated temperature array, temperature interpolation is performed on all critical areas of the edge device to obtain a set of temperature information.

[0022] Furthermore, the dimensionality reduction process includes: The baseline core temperature is calculated based on the statistical characteristics of the temperature information set. Based on the temperature difference between other key temperatures and the benchmark core temperature in the temperature information set, and combined with the temperature change pattern of equipment in the port environment, the distribution probability is calculated to obtain the key probability distribution. A lightweight dimensionality reduction algorithm adapted to edge devices is used to reduce the dimensionality of the temperature information set according to the key probability distribution, thereby obtaining a dimensionality-reduced temperature information set. The time nodes of multiple key temperature information in the reduced-dimensional temperature information set are obtained, and multi-source data of the corresponding time nodes in the vibration information set and current information set are extracted. Features strongly correlated with temperature changes are extracted through embedded feature fusion method to obtain the reduced-dimensional vibration information set and the reduced-dimensional current information set.

[0023] Furthermore, the lightweight dimensionality reduction algorithm adapted to edge devices performs dimensionality reduction processing on the temperature information set based on key probability distributions to obtain the dimensionality-reduced temperature information set, including: Set the preset dimensionality reduction ratio and preset number of dimensionality reductions to adapt to the computing capabilities of edge devices; According to the preset dimensionality reduction ratio, key temperatures are randomly selected from the temperature information set to obtain the first dimensionality-reduced temperature information set. Based on the temperature difference between multiple first-dimensionality-reduced key temperatures and the benchmark core temperature in the first-dimensionality-reduced temperature information set, the allocation probability is calculated to obtain the first-dimensionality-reduced key probability distribution. Calculate the similarity between the first key probability distribution and the key probability distribution to obtain the first key dimension reduction; Continue to randomly sample and reduce the dimensionality according to the preset dimensionality reduction ratio to obtain the second dimensionality-reduced temperature information set, and process it to obtain the second key dimensionality reduction; Repeat the random sampling dimensionality reduction process until the preset number of dimensionality reductions is reached, and output the set of dimensionality-reduced temperature information that maximizes the key dimensionality reduction and meets the computing latency requirements of edge devices.

[0024] Furthermore, the correlation analysis is performed pairwise on the reduced-dimensional temperature information set, the reduced-dimensional vibration information set, and the reduced-dimensional current information set to obtain multiple correlation coefficients, including: Based on the reduced-dimensional temperature information set and the reduced-dimensional current information set, a first-level fault correlation analysis is performed based on the equipment thermal damage mechanism to obtain the first coefficient. Using the reduced-dimensional vibration information set and the reduced-dimensional temperature information set, a secondary fault correlation analysis is performed based on the equipment damage mechanism to obtain the second coefficient; Using the aforementioned reduced-dimensional current information set and reduced-dimensional vibration information set, a three-level fault correlation analysis is performed for strong electromagnetic noise interference scenarios to obtain the third coefficient.

[0025] Furthermore, the step of performing a first-level fault correlation analysis based on the reduced-dimensional temperature information set and the reduced-dimensional current information set to obtain the first coefficient includes: Delete the initial unstable stage data in the reduced temperature information set and retain the valid data after the equipment is running stably. Based on the reduced-dimensional current information set, a first-level fault correlation analysis was performed according to the equipment thermal damage mechanism, and the first coefficient was calculated using the Pearson correlation coefficient algorithm: ; Where r is the first coefficient, and m is the number of valid data points in the reduced-dimensional temperature information set and the reduced-dimensional current information set. For the i-th key temperature in the set of reduced-dimensional temperature information, For the deformation parameters of the i-th part within the reduced-dimensional current information set, and These are the mean values ​​of the reduced-dimensional temperature information set and the reduced-dimensional current information set, respectively.

[0026] Furthermore, the construction of the lightweight deep learning model includes: Historical fault diagnosis data of edge devices in complex port environments are collected to obtain sample reduced-dimensional temperature information set, sample reduced-dimensional vibration information set, sample reduced-dimensional current information set, and corresponding sample fault mode set. The sample information includes labeled data under strong electromagnetic noise and vibration interference scenarios. The sample reduced-dimensional temperature information set, sample reduced-dimensional vibration information set, and sample reduced-dimensional current information set are respectively used as inputs, and the sample fault mode set is used as output. Based on the ensemble learning algorithm, temperature fault identification branch, vibration fault identification branch, and noise fault identification branch are constructed. Each fault identification branch adopts a lightweight network structure, and the quantization compression optimization algorithm improves the model's running efficiency on edge devices.

[0027] Furthermore, the fault identification branch includes multiple fault identification paths, each optimized for different types of damage faults; It integrates four fault identification branches and embeds a multi-source heterogeneous data feature fusion module to obtain a lightweight deep learning model that supports local real-time inference on edge devices.

[0028] Furthermore, the fault diagnosis results obtained by performing multi-source data fusion inference through the embedded feature fusion method include: Based on the first coefficient, the second coefficient, and the third coefficient, determine the first inference weight, the second inference weight, and the third inference weight; The reduced-dimensional temperature information set, reduced-dimensional vibration information set, and reduced-dimensional current information set are respectively input into the temperature fault identification branch, vibration fault identification branch, and noise fault identification branch. Computational resources are allocated according to the corresponding inference weights to perform fault analysis and output multiple basic fault modes. The embedded feature fusion module integrates all basic fault modes, selects the most frequent basic fault modes that conform to the equipment damage mechanism, and uses them as the final fault modes, including fault type, damage degree, fault location and maintenance suggestions.

[0029] In a specific implementation, an edge device fault diagnosis system for complex port environments includes: Acquisition module: Used to acquire temperature, vibration and current information from edge devices, and obtain sets of temperature, vibration and current information. Dimensionality reduction module: used to perform dimensionality reduction processing on the temperature information set, vibration information set, and current information set respectively, to obtain dimensionality-reduced temperature information set, dimensionality-reduced vibration information set, and dimensionality-reduced current information set; The correlation analysis module is used to perform pairwise correlation analysis on the reduced-dimensional temperature information set, the reduced-dimensional vibration information set, and the reduced-dimensional current information set to obtain multiple correlation analysis coefficients. Inference module: Used to build lightweight deep learning models. It inputs the reduced temperature information set, reduced vibration information set, and reduced current information set into the corresponding fault identification branches in the model, and determines the inference weights of each fault identification branch based on multiple correlation analysis coefficients. Fault diagnosis module: Used to perform multi-source data fusion inference through embedded feature fusion method to obtain fault diagnosis results.

[0030] In a specific implementation, a method for diagnosing faults in edge devices in a complex port environment includes the following steps: S1: When the port edge equipment is started, temperature information from multiple key locations on the edge equipment is continuously collected to obtain a set of temperature information; S2: Continuously collect vibration and current information of the edge device under strong electromagnetic noise and vibration interference environment to obtain vibration information set and current information set; S3: Perform lightweight preprocessing on the temperature information set, extract core temperature data to obtain the temperature information set, use a lightweight dimensionality reduction algorithm to perform dimensionality reduction on the temperature information set to obtain a dimensionality-reduced temperature information set, and perform embedded feature extraction and dimensionality reduction extraction on the vibration information and current information of multiple key temperature time nodes to obtain a dimensionality-reduced vibration information set and a dimensionality-reduced current information set. S4: Based on the reduced-dimensional temperature information set and the reduced-dimensional current information set, and combined with the equipment thermal damage mechanism, perform a first-level fault correlation analysis to obtain the first coefficient; S5: Using the reduced-dimensional vibration information set and the reduced-dimensional temperature information set, perform secondary fault correlation analysis based on the equipment damage mechanism to obtain the second coefficient; Using the aforementioned reduced-dimensional current information set and reduced-dimensional vibration information set, a three-level fault correlation analysis is performed for a strong electromagnetic noise interference scenario to obtain the third coefficient. S6: Based on ensemble learning, a lightweight deep learning model is constructed. Multi-source dimensionality-reduced data is input into the corresponding fault identification branch. The inference weight of each branch is determined according to the three-level coefficient. Multi-source data fusion inference is performed through embedded feature fusion method to obtain the fault mode as the fault diagnosis result.

[0031] In a specific embodiment, a fault diagnosis method for edge devices in a complex port environment is applied to port edge devices, and specifically includes the following steps: S1: When the port edge equipment is started, temperature information from multiple key locations on the edge equipment is continuously collected to obtain a set of temperature information.

[0032] Specifically, key locations on port edge equipment include core component areas (such as motor rotors and stators), power transmission areas (such as bearings and gearboxes), and heat dissipation areas (such as cooling fans and heat sinks). Electromagnetic interference-resistant temperature sensors (such as K-type thermocouples and isolated RTDs) are used to collect temperature data at each key location at multiple time points, with the sampling frequency set to 1-10Hz based on the equipment's operating characteristics. To address temperature data anomalies caused by strong electromagnetic interference in the port, a sliding window filtering method is used for outlier removal and data calibration. Based on the calibrated temperature array at each time point, linear interpolation or spline interpolation methods are used to calculate the temperature values ​​of all key areas of the equipment, forming a temperature information set. This set contains the temperature information of all key locations on the equipment at each time point.

[0033] S2: Continuously collect vibration and current information of the edge device under strong electromagnetic noise and vibration interference environment to obtain vibration information set and current information set.

[0034] Specifically, high-sensitivity vibration sensors (such as piezoelectric accelerometers) are installed at vibration-sensitive locations such as the equipment housing and bearing housings to collect vibration signals. Electromagnetic interference noise is removed through filtering circuits to form a vibration information set. An anti-noise microphone array is deployed around the equipment to collect operating noise signals, which are then processed through noise reduction algorithms to form a current information set. All data acquisition equipment is designed to be vibration-proof, dustproof, and resistant to electromagnetic interference, making it suitable for the complex environment of ports.

[0035] S3: Perform lightweight preprocessing on the temperature information set, extract core temperature data to obtain the temperature information set, use a lightweight dimensionality reduction algorithm to perform dimensionality reduction processing on the temperature information set to obtain a dimensionality-reduced temperature information set, and perform embedded feature extraction and dimensionality reduction extraction on the vibration information and current information of multiple key temperature time nodes to obtain a dimensionality-reduced vibration information set and a dimensionality-reduced current information set.

[0036] Specifically, temperature data from core components and equipment damage-sensitive areas are extracted from the temperature information set to form a temperature information set. The average temperature after 30 minutes of stable operation following equipment startup is selected as the baseline core temperature. The difference between each temperature in the temperature information set and the baseline temperature is calculated, and the probability distribution is assigned using a normal distribution statistical method to obtain the key probability distribution. A lightweight dimensionality reduction algorithm using improved principal component analysis (PCA) is employed to reduce the data dimensionality while preserving key features. Dimensionality reduction is validated through multiple random sampling, outputting the dimensionality-reduced temperature information set with the largest key dimensionality reduction that meets the computational latency requirements of edge devices. Based on the time nodes of the key temperatures for dimensionality reduction, data corresponding to the time nodes are extracted from the vibration and current information sets. An embedded feature fusion method is used to extract the correlation features between temperature and noise, temperature and vibration, and vibration and noise, performing simultaneous dimensionality reduction processing to obtain a dimensionality-matched multi-source data set for efficient computation by edge devices.

[0037] S4: Based on the set of images of the reduced-dimensional equipment and the set of information on the reduced-dimensional current, and combined with the thermal damage mechanism of the equipment, a first-level fault correlation analysis is performed to obtain the first coefficient.

[0038] Specifically, the initial unstable stage data in the reduced-dimensional temperature information set is deleted, the first coefficient of temperature and deformation parameters is calculated, and the correlation between thermal damage and part deformation is quantified.

[0039] S5: Using the reduced-dimensional vibration information set and the reduced-dimensional temperature information set, perform secondary fault correlation analysis based on the equipment damage mechanism to obtain the second coefficient; Using the aforementioned reduced-dimensional current information set and reduced-dimensional temperature information set, a three-level fault correlation analysis is performed for strong electromagnetic noise interference scenarios to obtain the third coefficient.

[0040] Specifically, the secondary fault correlation analysis, based on the equipment vibration damage mechanism, uses a mutual information algorithm to calculate the correlation between the reduced-dimensional vibration information set and the reduced-dimensional temperature information set, obtaining a second coefficient that reflects the degree of correlation between vibration impact and temperature change. The tertiary fault correlation analysis, targeting the characteristics of strong electromagnetic noise interference in ports, uses wavelet packet transform to extract the characteristic frequency components of noise and vibration signals, calculates their correlation coefficient, and obtains a third coefficient, thus eliminating the influence of electromagnetic noise interference on fault diagnosis.

[0041] S6: Based on ensemble learning, a lightweight deep learning model is constructed. Multi-source dimensionality-reduced data is input into the corresponding fault identification branch. The inference weight of each branch is determined according to the three-level coefficient. Multi-source data fusion inference is performed through embedded feature fusion method to obtain the fault mode as the fault diagnosis result.

[0042] Specifically, historical fault data of edge equipment in the complex environment of ports is collected, including normal operation data, thermal damage fault data, vibration and shock fault data, and abnormal data caused by electromagnetic interference. Fault type and damage degree information are labeled to construct a sample dataset. Based on ensemble learning algorithms, three fault identification branches are constructed for temperature, vibration, and noise. Each branch uses a pruned lightweight CNN or Transformer model, optimized through quantization compression and knowledge distillation to ensure that the inference latency on edge equipment is less than 100ms. Dimensionally reduced multi-source data are input into the corresponding branches, and inference weights are assigned according to a three-level distribution; the higher the correlation coefficient, the greater the weight of the inference result of the corresponding branch. An embedded feature fusion module fuses the basic fault modes output from each branch, selecting the fault modes with the highest frequency that conform to the equipment damage mechanism. The resulting fault type (e.g., rotor thermal deformation, bearing wear, abnormal electromagnetic interference), damage degree, fault location, and maintenance suggestions are output to achieve localized real-time diagnosis of edge equipment.

[0043] In a specific implementation, a method for diagnosing faults in edge devices in a complex port environment includes the following diagnostic steps: Historical fault data of edge equipment in complex port environments are collected, covering multiple scenarios such as normal operation data, thermal damage fault data, vibration and shock fault data, and abnormal data caused by electromagnetic interference. At the same time, key information such as fault type, damage degree, and operating conditions are labeled. A sample dataset containing all categories of source domain data and normal category data of target domain data is constructed. Among them, the source domain data covers various fault modes, and the target domain data focuses on the normal operation status of equipment, adapting to the differences in data distribution characteristics in complex environments.

[0044] Based on ensemble learning algorithms and domain adversarial neural network concepts, three fault identification branches—temperature, vibration, and noise—are constructed. Each branch adopts a pruned lightweight CNN model with a simplified structure containing convolutional layers, batch normalization layers, pooling layers, and activation layers. Through quantization compression, knowledge distillation optimization, and gradient inversion layer embedding, the computational complexity of the model is reduced while retaining feature extraction capabilities, ensuring that the inference latency of the model on edge devices is less than 100ms, thus adapting to the limited computing resources of edge devices.

[0045] Dimensionally reduced multi-source data, including dimensionality-reduced temperature, vibration, and noise information sets, are input into the corresponding fault identification branches. The inference weight of each branch is dynamically determined based on the three-level correlation coefficient, including the first coefficient, the second coefficient, and the third coefficient. The higher the correlation coefficient, the higher the proportion of computing resources allocated to the corresponding branch and the stronger the inference priority.

[0046] The basic fault modes output by each branch are fused by an embedded feature fusion module. During the fusion process, a domain-adaptive loss optimization strategy is introduced to reduce the impact of the difference in data distribution between the source domain and the target domain. The most frequent basic fault modes that conform to the equipment damage mechanism are selected, and finally accurate fault diagnosis results are output, including fault type, damage degree, fault location and targeted maintenance suggestions, so as to realize local real-time diagnosis of edge equipment in the complex environment of the port.

[0047] The specific steps include: (I) Optimization of Sample Dataset Construction The historical fault data of port edge equipment is divided into a source domain and a target domain. The source domain includes all categories of data, such as normal operation data, thermal damage fault data, vibration and shock fault data, and electromagnetic interference anomaly data, while the target domain only contains normal category data.

[0048] A sliding window strategy is adopted, with a window width of 10 and a sliding step of 1, to convert one-dimensional sensor data into two-dimensional data. The data is then processed using a normalization formula to eliminate the influence of dimensions. The normalization formula is as follows: ; in, For raw sensor data, For normalized data, and These are the mean and variance of the original data, respectively.

[0049] For categories where fault data is scarce in the source domain, data augmentation techniques are used to generate similar samples to ensure a balanced distribution of data for each fault type.

[0050] (II) Optimization Design of Lightweight Deep Learning Models

[0051] Network structure composition: The model is based on the domain adversarial neural network framework, integrating three fault identification branches: temperature, vibration, and noise. Each branch contains a feature extractor and a category classifier, and a gradient inversion layer is embedded between the feature extractor and the domain classifier.

[0052] Feature extractor structure: It adopts a two-layer convolutional structure, each of which includes a convolutional layer, a batch normalization layer (BN layer), a pooling layer, and an activation layer.

[0053] The first convolutional layer has a 3×3 kernel, 1 input channel, 10 output channels, and a stride of 1. A batch normalization (BN) layer is used to accelerate convergence and prevent overfitting. The pooling layer uses 2×2 max pooling. The activation function is ReLU, with the following formula: .

[0054] The second convolutional layer has a kernel size of 3×3, 10 input channels, 40 output channels, and a stride of 1. Subsequent BN, pooling, and activation layers are configured the same as the first layer.

[0055] Classifier structure: Both the category classifier and the domain classifier adopt a two-layer perceptron structure.

[0056] Category classifier: The first fully connected layer has 640 input nodes, adapts to the output dimension of the feature extractor, and has 50 output nodes; the subsequent layers are a one-dimensional BN layer, a Dropout layer with a deactivation probability of 0.1, a ReLU activation layer, and the last fully connected layer outputs the number of fault categories.

[0057] Domain classifier: Its structure is the same as the category classifier, with the last fully connected layer having 2 output nodes. It is used to distinguish between the source and target domains.

[0058] Model optimization strategy: Each branch adopts a pruned lightweight CNN structure, combined with quantization compression and knowledge distillation optimization; the gradient inversion layer coefficient λ is dynamically adjusted, where λ changes with the number of iterations, and P is the ratio of the current iteration number to the total number of iterations, to improve the model's domain adaptability.

[0059] (III) Inference weight allocation and feature fusion

[0060] Based on the third-order correlation coefficient (first coefficient) Second coefficient Third coefficient The normalization method is used to determine the inference weights of each branch, and the formula is as follows: ; ； ; in, , , These are the inference weights for the temperature, vibration, and noise branches, respectively.

[0061] By incorporating an embedded feature fusion module, a domain-adaptive loss optimization strategy is introduced, with the objective function being: ; The first term is the source domain data sample label prediction loss, the second term is the target domain data domain label prediction loss, and the third term is the source domain normal category data domain label prediction loss. The loss function is the sample label. For the domain label loss function, , , These are the parameters for the feature extractor, category classifier, and domain classifier, respectively. , They are the source domain and the target domain, respectively. Fault category, For domain category.

[0062] The system integrates the basic fault modes output from each branch, selects the most frequent modes that conform to the equipment damage mechanism, and outputs the final diagnostic results.

[0063] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0064] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for fault diagnosis of edge equipment in a complex port environment, characterized in that, include: Collect temperature, vibration, and current information from edge devices to obtain temperature, vibration, and current information sets. The temperature information set, vibration information set, and current information set are respectively subjected to dimensionality reduction processing to obtain the dimensionality-reduced temperature information set, dimensionality-reduced vibration information set, and dimensionality-reduced current information set; Correlation analysis was performed on each pairwise set of reduced-dimensional temperature information, reduced-dimensional vibration information, and reduced-dimensional current information to obtain multiple correlation coefficients. A lightweight deep learning model is constructed, and the reduced-dimensional temperature information set, reduced-dimensional vibration information set, and reduced-dimensional current information set are respectively input into the corresponding fault identification branch in the model. The inference weight of each fault identification branch is determined based on multiple correlation analysis coefficients. Multi-source data fusion inference is performed through embedded feature fusion method to obtain fault diagnosis results.

2. The method for fault diagnosis of edge equipment in a complex port environment according to claim 1, characterized in that, The temperature information acquired by the edge acquisition device includes: When the port edge equipment is started, temperature information at key locations on the edge equipment is collected at multiple time points to obtain a temperature array set; To address the abnormal temperature data caused by strong electromagnetic interference at the port, outlier values ​​were removed and data was calibrated for the temperature array at each time point. Based on the calibrated temperature array, temperature interpolation is performed on all critical areas of the edge device to obtain a set of temperature information.

3. The method for fault diagnosis of edge equipment in a complex port environment according to claim 1, characterized in that, The dimensionality reduction process includes: The baseline core temperature is calculated based on the statistical characteristics of the temperature information set. Based on the temperature difference between other key temperatures and the benchmark core temperature in the temperature information set, and combined with the temperature change pattern of equipment in the port environment, the distribution probability is calculated to obtain the key probability distribution. A lightweight dimensionality reduction algorithm adapted to edge devices is used to reduce the dimensionality of the temperature information set according to the key probability distribution, thereby obtaining a dimensionality-reduced temperature information set. The time nodes of multiple key temperature information in the reduced-dimensional temperature information set are obtained, and multi-source data of the corresponding time nodes in the vibration information set and current information set are extracted. Features strongly correlated with temperature changes are extracted through embedded feature fusion method to obtain the reduced-dimensional vibration information set and the reduced-dimensional current information set.

4. The method for fault diagnosis of edge equipment in a complex port environment according to claim 3, characterized in that, The lightweight dimensionality reduction algorithm adapted to edge devices performs dimensionality reduction processing on the temperature information set based on key probability distributions, resulting in a dimensionality-reduced temperature information set including: Set the preset dimensionality reduction ratio and preset number of dimensionality reductions to adapt to the computing capabilities of edge devices; According to the preset dimensionality reduction ratio, key temperatures are randomly selected from the temperature information set to obtain the first dimensionality-reduced temperature information set. Based on the temperature difference between multiple first-dimensionality-reduced key temperatures and the benchmark core temperature in the first-dimensionality-reduced temperature information set, the allocation probability is calculated to obtain the first-dimensionality-reduced key probability distribution. Calculate the similarity between the first key probability distribution and the key probability distribution to obtain the first key dimension reduction; Continue to randomly sample and reduce the dimensionality according to the preset dimensionality reduction ratio to obtain the second dimensionality-reduced temperature information set, and process it to obtain the second key dimensionality reduction; Repeat the random sampling dimensionality reduction process until the preset number of dimensionality reductions is reached, and output the set of dimensionality-reduced temperature information that maximizes the key dimensionality reduction and meets the computing latency requirements of edge devices.

5. The method for fault diagnosis of edge equipment in a complex port environment according to claim 1, characterized in that, The correlation analysis was performed on each pairwise set of reduced-dimensional temperature information, reduced-dimensional vibration information, and reduced-dimensional current information to obtain several correlation coefficients, including: Based on the reduced-dimensional temperature information set and the reduced-dimensional current information set, a first-level fault correlation analysis is performed based on the equipment thermal damage mechanism to obtain the first coefficient. Using the reduced-dimensional vibration information set and the reduced-dimensional temperature information set, a secondary fault correlation analysis is performed based on the equipment damage mechanism to obtain the second coefficient; Using the aforementioned reduced-dimensional current information set and reduced-dimensional vibration information set, a three-level fault correlation analysis is performed for strong electromagnetic noise interference scenarios to obtain the third coefficient.

6. The method for fault diagnosis of edge equipment in a complex port environment according to claim 5, characterized in that, The step of performing a first-level fault correlation analysis based on the reduced-dimensional temperature information set and the reduced-dimensional current information set to obtain the first coefficient includes: Delete the initial unstable stage data in the reduced temperature information set and retain the valid data after the equipment is running stably. Based on the reduced-dimensional current information set, a first-level fault correlation analysis was performed according to the equipment thermal damage mechanism, and the first coefficient was calculated using the Pearson correlation coefficient algorithm: ； Where r is the first coefficient, and m is the number of valid data points in the reduced-dimensional temperature information set and the reduced-dimensional current information set. For the i-th key temperature in the set of reduced-dimensional temperature information, For the deformation parameters of the i-th part within the reduced-dimensional current information set, and These are the mean values ​​of the reduced-dimensional temperature information set and the reduced-dimensional current information set, respectively.

7. The method for fault diagnosis of edge equipment in a complex port environment according to claim 5, characterized in that, The construction of the lightweight deep learning model includes: Historical fault diagnosis data of edge devices in complex port environments are collected to obtain sample reduced-dimensional temperature information set, sample reduced-dimensional vibration information set, sample reduced-dimensional current information set, and corresponding sample fault mode set. The sample information includes labeled data under strong electromagnetic noise and vibration interference scenarios. The sample reduced-dimensional temperature information set, sample reduced-dimensional vibration information set, and sample reduced-dimensional current information set are respectively used as inputs, and the sample fault mode set is used as output. Based on the ensemble learning algorithm, temperature fault identification branch, vibration fault identification branch, and noise fault identification branch are constructed. Each fault identification branch adopts a lightweight network structure, and the quantization compression optimization algorithm improves the model's running efficiency on edge devices.

8. The method for fault diagnosis of edge equipment in a complex port environment according to claim 7, characterized in that, The fault identification branch includes multiple fault identification paths, each optimized for different types of damage and faults; It integrates four fault identification branches and embeds a multi-source heterogeneous data feature fusion module to obtain a lightweight deep learning model that supports local real-time inference on edge devices.

9. The method for fault diagnosis of edge equipment in a complex port environment according to claim 8, characterized in that, The fault diagnosis results obtained by performing multi-source data fusion inference through the embedded feature fusion method include: Based on the first coefficient, the second coefficient, and the third coefficient, determine the first inference weight, the second inference weight, and the third inference weight; The reduced-dimensional temperature information set, reduced-dimensional vibration information set, and reduced-dimensional current information set are respectively input into the temperature fault identification branch, vibration fault identification branch, and noise fault identification branch. Computational resources are allocated according to the corresponding inference weights to perform fault analysis and output multiple basic fault modes. The embedded feature fusion module integrates all basic fault modes, selects the most frequent basic fault modes that conform to the equipment damage mechanism, and uses them as the final fault modes, including fault type, damage degree, fault location and maintenance suggestions.

10. A fault diagnosis system for edge devices in a complex port environment, characterized in that, include: Acquisition module: Used to acquire temperature, vibration and current information from edge devices, and obtain sets of temperature, vibration and current information. Dimensionality reduction module: used to perform dimensionality reduction processing on the temperature information set, vibration information set, and current information set respectively, to obtain dimensionality-reduced temperature information set, dimensionality-reduced vibration information set, and dimensionality-reduced current information set; The correlation analysis module is used to perform pairwise correlation analysis on the reduced-dimensional temperature information set, the reduced-dimensional vibration information set, and the reduced-dimensional current information set to obtain multiple correlation analysis coefficients. Inference module: Used to build lightweight deep learning models. It inputs the reduced temperature information set, reduced vibration information set, and reduced current information set into the corresponding fault identification branches in the model, and determines the inference weights of each fault identification branch based on multiple correlation analysis coefficients. Fault diagnosis module: Used to perform multi-source data fusion inference through embedded feature fusion method to obtain fault diagnosis results.

Images