A device detection management system and method based on the Internet of Things

Abstract

The present invention relates to the technical field of equipment detection and management, and discloses an equipment detection and management system and method based on the Internet of Things. The system includes a multi-source data acquisition unit, which collects heterogeneous time series data and synchronizes time-frequency domain signals; a state feature extraction unit, which uses a deep wavelet neural network to extract features and generate sparse codes; an anomaly detection unit, which constructs a graph attention network model to calculate dynamic anomaly thresholds; a life prediction unit, which predicts the remaining life of the equipment through a multi-task time series convolutional network; and a resource optimization unit, which optimizes resource allocation based on an improved ant colony algorithm. This invention solves the problems of incomplete data collection, inaccurate feature extraction, false positives and omissions in anomaly detection, low life prediction accuracy, and unreasonable resource allocation in traditional equipment detection and management, and achieves accurate monitoring of equipment status, timely detection of anomalies, accurate life prediction, and efficient resource utilization, thereby improving equipment operation reliability and reducing maintenance costs.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment detection management, and in particular to an equipment detection management system and method based on the Internet of Things. Background Art

[0002] In today's era of rapid digitalization and intelligent development, stable equipment operation is crucial for efficient production across various industries. With the widespread adoption of IoT technology and the massive number of devices connected to the network, device management faces new opportunities and challenges. Traditional equipment inspection and management methods are no longer able to meet the demands of modern, complex equipment systems, exposing numerous problems.

[0003] From a data acquisition perspective, early equipment monitoring relied on a small number of single-type sensors, resulting in limited data dimensions and an inability to fully reflect the equipment's operating status. Furthermore, the data collected by different sensors lacked synchronization in time and frequency, leading to biases in subsequent data analysis. In industrial production, a large piece of machinery may experience simultaneous changes in multiple parameters, such as temperature, vibration, and current. If these parameters cannot be precisely and synchronously collected, it is difficult to accurately determine the correlation between equipment failures and these parameters, delaying fault diagnosis and resolution.

[0004] Traditional methods for extracting device status features are based on manual experience and simple signal processing techniques. These methods often only capture superficial characteristics of the device's operating status and struggle to effectively identify complex internal failure modes and potential performance degradation trends. When a device experiences a minor fault or gradually degrades, manual feature extraction methods fail to detect the problem in a timely manner, causing minor faults to escalate into major failures, increasing repair costs and downtime. Electronic devices have complex internal circuits and diverse signal characteristics. Traditional feature extraction methods are unable to deeply explore subtle changes in the signals, making it difficult to predict device failures in advance.

[0005] In the anomaly detection phase, previous methods typically set fixed thresholds to determine whether a device is abnormal. However, in actual operation, the operating status of a device is affected by various factors, such as ambient temperature and load fluctuations. Fixed threshold detection methods are prone to false positives or missed negatives. In smart grids, the operating status of power equipment changes with fluctuations in power load. Fixed anomaly detection thresholds cannot adapt to these dynamic changes and may misinterpret normal load fluctuations as device anomalies, or fail to detect real device failures in a timely manner.

[0006] Equipment life prediction has always been a challenge in equipment management. In the past, methods based on statistical analysis of historical data or simple physical models were often used. However, these methods fail to fully consider the dynamic changes and complex environmental factors during equipment operation, resulting in low prediction accuracy. For some critical equipment, inaccurate life predictions can lead to premature replacement of equipment, resulting in a waste of resources; or failure to replace equipment in a timely manner can cause sudden equipment failures, impacting production continuity and safety. In the field of aircraft engines, the operating environment of the engine is extremely complex, and traditional life prediction methods cannot meet the high-precision prediction requirements.

[0007] With the increasing number of devices and the diversification of testing tasks, resource allocation becomes a major challenge. How to rationally allocate testing resources while ensuring accuracy and reducing energy consumption becomes a major challenge. Traditional resource allocation methods are mostly static, lacking dynamic adjustments based on the real-time status of devices and the priority of testing tasks, resulting in inefficient resource utilization. In data centers, a large number of servers require regular testing. Using static resource allocation methods can result in excessive testing resources being allocated to some servers even when they are performing well, while insufficient resources are allocated to servers that urgently need testing. Summary of the Invention

[0008] The purpose of the present invention is to provide an Internet of Things-based device detection management system and method to solve the problems raised in the above background technology.

[0009] To achieve the above objectives, the present invention provides the following technical solution: an IoT-based device detection and management system, the system comprising:

[0010] A multi-source data acquisition unit is used to collect heterogeneous time series data of the device through a multimodal sensor network and synchronize the temperature, vibration, and current signals in the time and frequency domains based on an adaptive weighted fusion algorithm, wherein the heterogeneous time series data includes high-frequency sampling data and event-triggered data;

[0011] A state feature extraction unit, configured to perform multi-scale decomposition on the heterogeneous time series data using a deep wavelet neural network, extract a joint time-frequency feature matrix, and generate a sparse coding representation of the device operating state using a deformable convolution kernel;

[0012] An anomaly detection unit is used to build an anomaly scoring model based on a graph attention network, calculate a dynamic anomaly threshold based on the topological relationship between device nodes and the sparse coding representation, and output a device anomaly probability distribution map;

[0013] A life prediction unit, configured to perform degradation mode decoupling on the time-frequency joint feature matrix through a multi-task temporal convolutional network, and generate a confidence interval prediction of the remaining life of the device in combination with Weibull distribution parameter estimation;

[0014] The resource optimization unit is used to dynamically optimize the resource allocation matrix of multi-device detection tasks based on the pheromone gradient update strategy of the improved ant colony algorithm to generate the minimum energy consumption scheduling path.

[0015] Preferably, the execution step of the state feature extraction unit further includes:

[0016] Inputting the heterogeneous time series data into a continuous wavelet transform layer to generate a time-frequency scale energy distribution map;

[0017] Capturing the non-stationary features in the energy distribution map by using a deformable convolution kernel to construct a multi-resolution feature pyramid;

[0018] A sparse autoencoder is used to reduce the dimension of the feature pyramid and output the sparse coding representation.

[0019] Preferably, the execution steps of the anomaly detection unit include:

[0020] Construct a device node topology graph, where nodes represent device sensors and edges represent signal correlation strength;

[0021] Aggregate the feature information of neighboring nodes through the multi-head graph attention mechanism to generate node embedding vectors;

[0022] A Gaussian mixture model is used to perform density estimation on the embedding vector, and the Mahalanobis distance is calculated as the anomaly score threshold.

[0023] Preferably, the execution step of the life prediction unit further includes:

[0024] Using causal dilation convolution to extract long-range dependencies in the time-frequency joint feature matrix;

[0025] The degradation trend and sudden failure probability are predicted separately through a parallel branch structure, and the KL divergence is used to constrain the multi-task loss function;

[0026] The shape parameter and scale parameter of the Weibull distribution are sampled a posteriori based on Bayesian inference to generate the confidence interval prediction.

[0027] Preferably, the execution step of the resource optimization unit further includes:

[0028] Construct the energy consumption priority cost function of the device detection task and initialize the pheromone concentration matrix;

[0029] Adjusting the contribution weight of the pheromone gradient in path selection through an adaptive volatility factor;

[0030] The optimal path set is updated using an elite retention strategy, and the minimum energy consumption scheduling path is output.

[0031] Preferably, the state feature extraction unit further includes:

[0032] An adversarial training mechanism is introduced, in which the generator is used to synthesize pseudo features under noise interference, and the discriminator is used to distinguish between real and synthesized features;

[0033] Optimizing the robustness of the feature extractor through gradient reversal layers.

[0034] Preferably, the execution step of the abnormality detection unit further includes:

[0035] Design a spatiotemporal graph convolution module to perform time series modeling on the historical anomaly probability distribution of device nodes;

[0036] The current node embedding vector and historical modeling results are input into the gated recurrent unit to generate a dynamically updated anomaly threshold.

[0037] Preferably, the execution step of the life prediction unit further includes:

[0038] A degradation-failure joint probability model is constructed, and the expression of the degradation-failure joint probability model is:

[0039] in, is the joint probability at time t, which is the probability assessment result of the comprehensive equipment degradation trend and sudden failure risk, is the degradation trend prediction value based on convolutional neural network, which represents the performance degradation degree of the device at time t. is the failure risk function based on survival analysis, which represents the probability of sudden failure of the equipment at time t. is the dynamic weight coefficient;

[0040] The parameters of the joint probability model are iteratively optimized using an expectation maximization algorithm.

[0041] Preferably, the execution step of the resource optimization unit further includes:

[0042] Pareto frontier analysis is introduced to perform multi-objective trade-offs among energy consumption, delay, and detection accuracy.

[0043] A Pareto optimal solution set of the resource allocation matrix is ​​generated by a non-dominated sorting genetic algorithm.

[0044] Preferably, the present invention also includes a device detection and management method based on the Internet of Things, the method comprising the following steps:

[0045] Step 1: Utilize a multi-source data acquisition unit to collect heterogeneous time series data from the device through a multimodal sensor network. This heterogeneous time series data includes high-frequency sampling data and event-triggered data. Temperature, vibration, and current signals are synchronized in the time and frequency domains based on an adaptive weighted fusion algorithm.

[0046] Step 2: Using a state feature extraction unit, a deep wavelet neural network is used to perform multi-scale decomposition on the collected heterogeneous time series data, extracting the time-frequency joint feature matrix. A sparse coding representation of the device operating status is then generated using a deformable convolution kernel.

[0047] Step 3: With the help of the anomaly detection unit, an anomaly scoring model based on the graph attention network is constructed. The dynamic anomaly threshold is calculated based on the topological relationship between device nodes and the generated sparse coding representation, and the device anomaly probability distribution map is output;

[0048] Step 4: The service life prediction unit uses a multi-task temporal convolutional network to decouple the degradation mode of the time-frequency joint feature matrix and generate a confidence interval prediction of the remaining service life of the equipment based on the Weibull distribution parameter estimation;

[0049] Step 5: Use the resource optimization unit to dynamically optimize the resource allocation matrix of the multi-device detection task based on the pheromone gradient update strategy of the improved ant colony algorithm to generate the minimum energy consumption scheduling path.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] The IoT-based equipment detection management system and method proposed in the present invention have significant advantages over traditional technologies, improving the level of equipment detection management at multiple levels and providing strong guarantees for the efficient operation of equipment in various industries. In the data acquisition link, the multi-source data acquisition unit collects heterogeneous time series data of the equipment through a multimodal sensor network, covering high-frequency sampling data and event-triggered data, and can obtain the operating information of the equipment in an all-round and real-time manner. In addition, the temperature, vibration, and current signals are synchronized in the time and frequency domains based on the adaptive weighted fusion algorithm to ensure the accuracy and consistency of the data, providing a solid and reliable data foundation for subsequent analysis. This makes the monitoring of the equipment's operating status more comprehensive and accurate, and can capture subtle changes in the equipment's operation in a timely manner, providing key data support for early fault diagnosis.

[0052] The state feature extraction unit uses a deep wavelet neural network to perform multi-scale decomposition on heterogeneous time series data, deeply exploring the joint time-frequency features in the data and extracting a more representative feature matrix. A sparse coding representation of the device's operating status is generated using a deformable convolution kernel, reducing data redundancy and improving feature expressiveness. This advanced feature extraction method can more effectively identify potential failure modes and performance degradation trends in equipment, significantly improving the accuracy and reliability of equipment state assessment compared to traditional feature extraction methods based on manual experience.

[0053] The anomaly detection unit constructs an anomaly scoring model based on a graph attention network. It calculates dynamic anomaly thresholds based on the topological relationships between device nodes and sparse coding representations, and outputs a device anomaly probability distribution map. This method fully considers the interdependencies between device components and dynamically adapts to changes in device operating status, effectively avoiding the false positives and false negatives common with traditional fixed-threshold detection methods. By monitoring the device anomaly probability distribution in real time, it can promptly detect device anomalies and precisely locate their locations, providing clear troubleshooting guidance for equipment maintenance personnel and significantly improving the efficiency and accuracy of fault diagnosis.

[0054] The life prediction unit uses a multi-task temporal convolutional network to decouple the degradation patterns of the joint time-frequency feature matrix and, combined with Weibull distribution parameter estimation, generates confidence interval predictions for the equipment's remaining life. This approach comprehensively considers the equipment's various operating characteristics and degradation patterns, resulting in more accurate predictions of the equipment's remaining life than traditional life prediction methods. Based on these predictions, equipment managers can formulate appropriate maintenance plans and equipment replacement strategies in advance, avoiding production interruptions caused by sudden equipment failures, reducing maintenance costs, and improving production continuity and stability. Furthermore, confidence interval predictions provide more comprehensive information for decision-making, enhancing the scientific nature and reliability of decision-making.

[0055] The resource optimization unit dynamically optimizes the resource allocation matrix for multi-device inspection tasks based on the pheromone gradient update strategy of an improved ant colony algorithm, generating a scheduling path with minimal energy consumption. While ensuring inspection accuracy, this unit achieves a rational allocation of inspection resources and effectively reduces energy consumption during the inspection process. This not only aligns with current trends in energy conservation and emission reduction, but also improves resource utilization efficiency and reduces operating costs for enterprises. By dynamically optimizing resource allocation, it can flexibly adjust inspection resource allocation based on the real-time status of equipment and the priority of inspection tasks, ensuring timely and effective inspection of critical equipment and further improving the overall effectiveness of equipment inspection management.

[0056] Furthermore, the various units in this invention collaborate and organically integrate to form a complete equipment detection and management system. From data collection and status analysis to anomaly detection, lifespan prediction, and resource optimization, each link works closely together to achieve intelligent, automated, and efficient equipment detection and management. This systematic solution can be widely applied across multiple industries, providing unified and efficient detection and management services for different types of equipment, and promoting the digital transformation and intelligent development of various industries. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a working principle diagram of the device detection and management system based on the Internet of Things according to the present invention;

[0058] Figure 2 This is the working principle diagram of the anomaly detection unit;

[0059] Figure 3 This is the working principle diagram of the life prediction unit;

[0060] Figure 4 Diagram of how adversarial training works for the state feature extraction unit. DETAILED DESCRIPTION

[0061] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0062] See also Figure 1-Figure 4 The present invention provides an IoT-based device detection and management system, which aims to achieve comprehensive monitoring, status analysis, anomaly detection, life prediction, and resource optimization allocation of devices. It specifically includes the following parts:

[0063] Multi-source data acquisition unit: This unit utilizes a multimodal sensor network to collect heterogeneous time-series data from devices, including high-frequency sampling data and event-triggered data. Temperature, vibration, and current signals are synchronized in the time and frequency domains using an adaptive weighted fusion algorithm. This ensures the collected data accurately reflects the device's operating status and provides a reliable foundation for subsequent analysis.

[0064] State Feature Extraction Unit: This unit uses a deep wavelet neural network to perform multi-scale decomposition on the collected heterogeneous time series data and extract a joint time-frequency feature matrix. Using a deformable convolution kernel, it generates a sparse coded representation of the device's operating status, effectively capturing key characteristic information within the device's operating state and providing strong support for device status assessment and anomaly detection.

[0065] Anomaly detection unit: Builds an anomaly scoring model based on the graph attention network. Based on the topological relationship between device nodes and the sparse coding representation generated by the state feature extraction unit, it calculates the dynamic anomaly threshold and then outputs the device anomaly probability distribution map, realizing real-time anomaly monitoring and precise positioning of the device operating status.

[0066] Life Prediction Unit: This unit uses a multi-task temporal convolutional network to decouple the degradation patterns of the time-frequency joint feature matrix. Combined with Weibull distribution parameter estimation, it generates confidence interval predictions for the remaining life of the equipment. This provides a scientific basis for equipment maintenance and replacement, helping companies plan ahead and reduce losses caused by equipment failures.

[0067] Resource Optimization Unit: Based on the pheromone gradient update strategy of the improved ant colony algorithm, the resource allocation matrix of multi-device detection tasks is dynamically optimized to generate the minimum energy consumption scheduling path. While meeting the device detection requirements, it also minimizes energy consumption during the detection process and improves resource utilization efficiency.

[0068] The present invention will be further described below in conjunction with Examples 1 to 5:

[0069] Example 1:

[0070] In the practical application of the state feature extraction unit, heterogeneous time series data is an important foundation for analyzing device operating status. The collected heterogeneous time series data is input into the continuous wavelet transform layer, which analyzes the data at different time and frequency scales to generate a time-frequency energy distribution map. This map intuitively displays the energy distribution of the data at different frequencies and time points, providing rich information for subsequent capture of device operational characteristics.

[0071] Deformable convolution kernels are used to capture non-stationary features in energy distribution maps. Unlike traditional convolution kernels, the sampling position of a deformable convolution kernel can be adaptively adjusted based on the characteristics of the data. When processing device operating data, the operating status of the device often changes. The deformable convolution kernel can flexibly adjust the sampling position based on these changes, thereby more accurately capturing the non-stationary features of the energy distribution map.

[0072] The purpose of constructing a multi-resolution feature pyramid is to analyze and process features at different resolutions. By performing multiple downsampling and convolution operations on the energy distribution map, feature maps of varying resolutions are generated, which form the multi-resolution feature pyramid. Low-resolution feature maps can reflect the overall trends and macroscopic features of the data, while high-resolution feature maps preserve the data's detailed information. This allows subsequent analysis to select feature maps of appropriate resolutions based on specific needs.

[0073] A sparse autoencoder is used to reduce the dimensionality of the feature pyramid. A sparse autoencoder is a specialized neural network that, while learning data features, uses sparsity constraints to make the encoder's output feature representation more concise and effective. When reducing the dimensionality of the feature pyramid, the sparse autoencoder removes redundant information, retaining the most critical features, and outputs a sparsely coded representation of the device's operating status. This sparsely coded representation not only reduces the burden of data storage and transmission, but also improves the efficiency and accuracy of subsequent analysis.

[0074] To further improve the robustness of the feature extractor, an adversarial training mechanism is introduced. The generator synthesizes pseudo-features under noise interference, simulating the various interference conditions that a device may experience during actual operation. The discriminator is responsible for distinguishing between real features and synthesized pseudo-features. During training, the generator and discriminator compete with each other. The generator continuously optimizes its pseudo-feature generation strategy, making it more difficult for the discriminator to identify; the discriminator continuously improves its recognition ability, accurately distinguishing real features from pseudo-features. Through the gradient reversal layer, the discriminator's gradient is backpropagated to the feature extractor, thereby optimizing the robustness of the feature extractor, enabling it to accurately extract features of the device's operating status even in the presence of noise and interference.

[0075] Example 2:

[0076] In the anomaly detection unit, the first step is to construct a device node topology map. Nodes in the device node topology map represent device sensors, which are distributed across the device and collect various operational data. Edges represent the strength of signal correlation. By calculating the correlation between sensor signals, we can determine their connections. For example, in an industrial motor device, the data collected by temperature sensors, vibration sensors, and current sensors may exhibit certain correlations. A topology map constructed based on these correlations can intuitively display the relationships between the sensors.

[0077] The multi-head graph attention mechanism aggregates the feature information of neighboring nodes. The multi-head graph attention mechanism is an effective method for feature learning on graph-structured data. It uses multiple attention heads to concurrently calculate the attention weights of nodes, thereby aggregating the feature information of neighboring nodes from different perspectives. Each attention head focuses on different features, and the results of multiple attention heads are combined to more comprehensively reflect the characteristics of the node.

[0078] A Gaussian mixture model is used to estimate the density of the embedding vectors, and the Mahalanobis distance is calculated as the anomaly scoring threshold. The Gaussian mixture model is a commonly used probabilistic model that can represent complex probability distributions as a weighted sum of multiple Gaussian distributions. When estimating the density of node embedding vectors, the Gaussian mixture model can well fit the data distribution. The Mahalanobis distance is a distance metric that takes into account data covariance and can more accurately measure the similarity between data points. By calculating the Mahalanobis distance, an anomaly scoring threshold is obtained. When the node's characteristics differ from the normal state by more than this threshold, the node is considered anomaly.

[0079] To better model the historical anomaly probability distribution of device nodes over time, a spatiotemporal graph convolution module was designed. This module combines graph convolution with time series analysis methods to simultaneously consider both the spatial relationships and time series information between device nodes. In practical applications, device anomalies often exhibit a certain degree of temporal correlation. Modeling the historical anomaly probability distribution using the spatiotemporal graph convolution module captures this temporal correlation, enabling more accurate prediction of future anomalies.

[0080] The current node embedding vector and historical modeling results are input into a gated recurrent unit to generate a dynamically updated anomaly threshold. A gated recurrent unit is a special type of recurrent neural network that effectively handles long-term dependencies in time series data. During anomaly detection, the gated recurrent unit dynamically updates the anomaly threshold based on the current node embedding vector and the modeling results of the historical anomaly probability distribution. This allows the anomaly threshold to be adjusted in real time as the device's operating status changes, improving the accuracy and timeliness of anomaly detection.

[0081] Example 3:

[0082] In the life prediction unit's workflow, causal dilation convolution is used to extract long-range dependencies in the joint time-frequency feature matrix. Causal dilation convolution is a special convolution operation that introduces holes in the convolution kernel to increase the receptive field, thereby capturing longer-range temporal dependencies. In equipment operation data, performance degradation is often a long-term process, characterized by long-range dependencies. Causal dilation convolution can effectively extract these long-range dependencies, providing important support for accurately predicting the remaining life of equipment.

[0083] A parallel branching structure is used to predict degradation trends and sudden failure probabilities. This parallel branching structure consists of two independent branches: one for predicting device degradation trends and the other for predicting sudden failure probabilities. The branch that predicts degradation trends uses a convolutional neural network to analyze the joint time-frequency feature matrix to predict the degree of performance degradation at different future points in time. The branch that predicts sudden failure probabilities uses survival analysis, taking into account factors such as device age and operating environment, to predict the probability of sudden failures at different points in time.

[0084] The KL divergence is used to constrain the multi-task loss function. KL divergence is a metric that measures the difference between two probability distributions. In lifespan prediction, minimizing the KL divergence between the predicted results and the ground truth can make the predictions closer to the actual equipment degradation trends and sudden failure probabilities. Incorporating KL divergence into the multi-task loss function can simultaneously optimize both degradation trend prediction and sudden failure probability prediction, improving prediction accuracy.

[0085] Based on Bayesian inference, posterior sampling of the shape and scale parameters of the Weibull distribution is performed to generate confidence interval predictions. The Weibull distribution is a commonly used life distribution model, and its shape and scale parameters determine the shape and characteristics of the distribution. Bayesian inference is a probability-based inference method that combines prior information with observed data to obtain the posterior distribution of parameters. In life prediction, using Bayesian inference to perform posterior sampling of the parameters of the Weibull distribution can obtain parameter uncertainty information, thereby generating confidence interval predictions for the remaining life of the equipment. This confidence interval prediction can more comprehensively reflect the reliability of the prediction results, providing a more scientific basis for equipment maintenance decisions.

[0086] Construct the degradation-failure joint probability model, which is expressed as:

[0087] in, is the joint probability at time t, which combines the probability assessment results of equipment degradation trend and sudden failure risk; is the degradation trend prediction value based on convolutional neural network, which represents the performance degradation degree of the device at time t; is the failure risk function based on survival analysis, which represents the probability of sudden failure of the equipment at time t; is a dynamic weight coefficient used to adjust the relative importance of degradation trend and sudden failure risk in the joint probability. The joint probability model is iteratively optimized using the expectation maximization algorithm, continuously adjusting model parameters to better fit the actual operation of the equipment and improve the accuracy of life prediction.

[0088] Example 4:

[0089] During the operation of the resource optimization unit, an energy priority cost function for the equipment inspection tasks is first constructed, and a pheromone concentration matrix is ​​initialized. This energy priority cost function comprehensively considers factors such as the importance of the equipment inspection task and the required energy consumption. Inspection tasks for key equipment are given a higher priority and a larger weight in the cost function. For tasks with higher energy consumption but lower importance, the cost function is adjusted accordingly. The pheromone concentration matrix is ​​used to simulate the pheromones released by ants during path finding. Initially, the pheromone concentration matrix is ​​set to a small constant to provide a basis for subsequent path selection.

[0090] The adaptive volatility factor adjusts the contribution of the pheromone gradient to path selection. This factor dynamically adjusts the pheromone's evaporation rate based on the current task situation and search progress. In the early stages of a search, the volatility factor is high, allowing the pheromone to evaporate quickly, encouraging ants to explore more paths and preventing premature algorithm convergence. As the search progresses, the volatility factor gradually decreases, slowing the pheromone's evaporation rate. This makes ants more likely to choose paths with previously accumulated pheromones, accelerating the algorithm's convergence.

[0091] An elitist strategy is used to update the optimal path set. This strategy involves retaining the currently found optimal path during each iteration and passing it directly to the next iteration. This ensures that the algorithm does not lose the current optimal solution due to randomness during the search process, improving search efficiency and convergence speed. By continuously updating the optimal path set, the system ultimately outputs the minimum energy consumption scheduling path, optimizing resource allocation for multi-device detection tasks.

[0092] Pareto front analysis is introduced to perform multi-objective trade-offs between energy consumption, latency, and detection accuracy. Pareto front analysis is a commonly used method for multi-objective optimization problems. It can find a set of non-dominated solutions that achieve a balance between different objectives. In device detection tasks, energy consumption, latency, and detection accuracy are three interrelated objectives. Reducing energy consumption may increase detection latency or decrease detection accuracy, while improving detection accuracy may increase energy consumption and latency. Through Pareto front analysis, a series of resource allocation solutions that meet different requirements can be found, allowing the most appropriate solution to be selected in different application scenarios.

[0093] The non-dominated sorting genetic algorithm (NSA) generates a Pareto-optimal solution set for the resource allocation matrix. This algorithm is a multi-objective optimization algorithm based on a genetic algorithm. It optimizes the resource allocation matrix by simulating the selection, crossover, and mutation operations in biological evolution. The algorithm first performs a non-dominated sort on the initial population, classifying individuals into different levels based on dominance relationships. Individuals are then selected based on their level and congestion, giving higher-performing individuals a greater chance of being selected for crossover and mutation operations. Through continuous iteration, a Pareto-optimal solution set for the resource allocation matrix is ​​ultimately generated, providing a rich set of options for resource allocation in multi-device inspection tasks.

[0094] Example 5:

[0095] In the actual operation of the entire IoT-based equipment detection and management system, various units collaborate closely. The multi-source data acquisition unit continuously and stably collects heterogeneous time-series data from the equipment. This includes not only high-frequency sampling data, such as temperature and vibration data from key equipment parts, which are collected at a high frequency to promptly capture subtle changes in the equipment's operating status, but also event-triggered data, such as data collected by key events such as equipment startup and shutdown, to ensure a comprehensive record of the equipment's operating behavior. An adaptive weighted fusion algorithm is used to synchronize the temperature, vibration, and current signals in the time and frequency domains. During this process, the algorithm dynamically adjusts the weights of different signals based on their importance in reflecting the equipment's operating status, ensuring that the fused data more accurately reflects the equipment's true status.

[0096] After receiving data from the multi-source data acquisition unit, the state feature extraction unit uses a deep wavelet neural network. Its multi-scale decomposition function analyzes data at different time and frequency scales, extracting a rich joint time-frequency feature matrix. When constructing the multi-resolution feature pyramid, the downsampling and convolution operations at each layer are carefully designed to preserve the key features of the data. The deformable convolution kernel continuously adjusts the sampling position based on data changes when capturing non-stationary features, laying the foundation for generating accurate sparse coding representations. Furthermore, the introduction of an adversarial training mechanism further enhances the robustness of the feature extractor, enabling the system to stably extract device operating characteristics even in complex operating environments.

[0097] The anomaly detection unit builds an anomaly scoring model based on a graph attention network. The device node topology map accurately reflects the signal correlations between device sensors, providing a reliable structure for subsequent feature aggregation. The multi-head graph attention mechanism aggregates feature information of neighboring nodes from multiple perspectives, making node embedding vectors more representative. The combination of the Gaussian mixture model and Mahalanobis distance provides a scientific basis for calculating anomaly scoring thresholds. The collaborative work of the spatiotemporal graph convolution module and the gated recurrent unit enables dynamic updating of anomaly thresholds, enabling timely detection of anomalies in device operation.

[0098] The life prediction unit utilizes a multi-task temporal convolutional network to decouple degradation patterns from the joint time-frequency feature matrix. Causal dilation convolution effectively extracts long-range dependencies, and a parallel branching structure accurately predicts degradation trends and sudden failure probabilities. A KL-divergence-constrained multi-task loss function ensures the coordinated optimization of the two prediction tasks. Combining Bayesian inference with Weibull distribution parameter estimation, the generated confidence interval predictions provide a scientific basis for equipment maintenance. The construction of a joint degradation-failure probability model and the application of the expectation-maximization algorithm further improve the accuracy of life prediction.

[0099] The resource optimization unit dynamically optimizes the resource allocation matrix for multi-device detection tasks based on an improved ant colony algorithm. The energy priority cost function is constructed to fully consider the actual requirements and energy consumption of the device detection tasks. The initialization of the pheromone concentration matrix and the adjustment of the adaptive volatility factor enable ants to more efficiently find the minimum energy consumption scheduling path during path search. An elite retention strategy ensures that the algorithm continuously optimizes the path. The application of Pareto frontier analysis and a non-dominated sorting genetic algorithm achieves a multi-objective trade-off between energy consumption, latency, and detection accuracy, providing diverse resource allocation solutions for different application scenarios.

[0100] In practical application scenarios, such as equipment management in large factories, the system can monitor the operating status of a large number of devices in real time, promptly detect abnormalities, and accurately predict the remaining lifespan of equipment. It also optimizes resource allocation for inspection tasks, reduces energy consumption, and improves production efficiency and equipment management. In data center equipment management, the system can comprehensively monitor and manage servers and network equipment, ensuring stable data center operation and reducing losses caused by equipment failures.

[0101] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0102] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An equipment detection and management system based on the Internet of Things, characterized in that: include: A multi-source data acquisition unit is used to collect heterogeneous time series data of the device through a multimodal sensor network and synchronize the temperature, vibration, and current signals in the time and frequency domains based on an adaptive weighted fusion algorithm, wherein the heterogeneous time series data includes high-frequency sampling data and event-triggered data; A state feature extraction unit, configured to perform multi-scale decomposition on the heterogeneous time series data using a deep wavelet neural network, extract a joint time-frequency feature matrix, and generate a sparse coding representation of the device operating state using a deformable convolution kernel; An anomaly detection unit is used to build an anomaly scoring model based on a graph attention network, calculate a dynamic anomaly threshold based on the topological relationship between device nodes and the sparse coding representation, and output a device anomaly probability distribution map; A life prediction unit, configured to perform degradation mode decoupling on the time-frequency joint feature matrix through a multi-task temporal convolutional network, and generate a confidence interval prediction of the remaining life of the device in combination with Weibull distribution parameter estimation; The resource optimization unit is used to dynamically optimize the resource allocation matrix of multi-device detection tasks based on the pheromone gradient update strategy of the improved ant colony algorithm to generate the minimum energy consumption scheduling path; The execution step of the state feature extraction unit further includes: Inputting the heterogeneous time series data into a continuous wavelet transform layer to generate a time-frequency scale energy distribution map; Capturing the non-stationary features in the energy distribution map by using a deformable convolution kernel to construct a multi-resolution feature pyramid; Using a sparse autoencoder to reduce the dimension of the feature pyramid and output the sparse coding representation; The execution steps of the life prediction unit also include: Using causal dilation convolution to extract long-range dependencies in the time-frequency joint feature matrix; The degradation trend and sudden failure probability are predicted separately through a parallel branch structure, and the KL divergence is used to constrain the multi-task loss function; The shape parameter and scale parameter of the Weibull distribution are sampled a posteriori based on Bayesian inference to generate the confidence interval prediction.

2. The device detection and management system based on the Internet of Things according to claim 1, characterized in that: The execution steps of the abnormality detection unit include: Construct a device node topology graph, where nodes represent device sensors and edges represent signal correlation strength; Aggregate the feature information of neighboring nodes through the multi-head graph attention mechanism to generate node embedding vectors; A Gaussian mixture model is used to perform density estimation on the embedded vector, and a Mahalanobis distance is calculated as the dynamic anomaly threshold.

3. The device detection and management system based on the Internet of Things according to claim 1, characterized in that: The execution steps of the resource optimization unit also include: Construct the energy consumption priority cost function of the device detection task and initialize the pheromone concentration matrix; Adjusting the contribution weight of the pheromone gradient in path selection through an adaptive volatility factor; The optimal path set is updated using an elite retention strategy, and the minimum energy consumption scheduling path is output.

4. The device detection and management system based on the Internet of Things according to claim 1, characterized in that: The state feature extraction unit further includes: An adversarial training mechanism is introduced, in which the generator is used to synthesize pseudo features under noise interference, and the discriminator is used to distinguish between real and synthesized features; Optimizing the robustness of the feature extractor through gradient reversal layers.

5. The device detection and management system based on the Internet of Things according to claim 2, characterized in that: The execution step of the abnormality detection unit further includes: Design a spatiotemporal graph convolution module to perform time series modeling on the historical anomaly probability distribution of device nodes; The current node embedding vector and historical modeling results are input into the gated recurrent unit to generate a dynamically updated anomaly threshold.

6. The device detection and management system based on the Internet of Things according to claim 1, characterized in that: The execution steps of the life prediction unit also include: A degradation-failure joint probability model is constructed, and the expression of the degradation-failure joint probability model is: in, is the joint probability at time t, which is the probability assessment result of the comprehensive equipment degradation trend and sudden failure risk, is the degradation trend prediction value based on convolutional neural network, which represents the performance degradation degree of the device at time t. is the failure risk function based on survival analysis, which represents the probability of sudden failure of the equipment at time t. is the dynamic weight coefficient; The parameters of the joint probability model are iteratively optimized using an expectation maximization algorithm.

7. The device detection and management system based on the Internet of Things according to claim 3, characterized in that: The execution steps of the resource optimization unit also include: Pareto frontier analysis is introduced to perform multi-objective trade-offs among energy consumption, delay, and detection accuracy. A Pareto optimal solution set of the resource allocation matrix is ​​generated by a non-dominated sorting genetic algorithm.

8. A device detection and management method based on the Internet of Things, characterized in that: The following steps are involved: Step 1: Utilize a multi-source data acquisition unit to collect heterogeneous time series data from the device through a multimodal sensor network. This heterogeneous time series data includes high-frequency sampling data and event-triggered data. Temperature, vibration, and current signals are synchronized in the time and frequency domains based on an adaptive weighted fusion algorithm. Step 2: Using a state feature extraction unit, a deep wavelet neural network is used to perform multi-scale decomposition on the collected heterogeneous time series data, extracting the time-frequency joint feature matrix. A sparse coding representation of the device operating status is then generated using a deformable convolution kernel. Step 3: With the help of the anomaly detection unit, an anomaly scoring model based on the graph attention network is constructed. The dynamic anomaly threshold is calculated based on the topological relationship between device nodes and the generated sparse coding representation, and the device anomaly probability distribution map is output; Step 4: The service life prediction unit uses a multi-task temporal convolutional network to decouple the degradation mode of the time-frequency joint feature matrix and generate a confidence interval prediction of the remaining service life of the equipment based on the Weibull distribution parameter estimation; Step 5: Use the resource optimization unit to dynamically optimize the resource allocation matrix of the multi-device detection task based on the pheromone gradient update strategy of the improved ant colony algorithm to generate the minimum energy consumption scheduling path; The execution step of the state feature extraction unit further includes: Inputting the heterogeneous time series data into a continuous wavelet transform layer to generate a time-frequency scale energy distribution map; Capturing the non-stationary features in the energy distribution map by using a deformable convolution kernel to construct a multi-resolution feature pyramid; Using a sparse autoencoder to reduce the dimension of the feature pyramid and output the sparse coding representation; The execution steps of the life prediction unit also include: Using causal dilation convolution to extract long-range dependencies in the time-frequency joint feature matrix; The degradation trend and sudden failure probability are predicted separately through a parallel branch structure, and the KL divergence is used to constrain the multi-task loss function; The shape parameter and scale parameter of the Weibull distribution are sampled a posteriori based on Bayesian inference to generate the confidence interval prediction.

Images