An edge-computing-based communication construction equipment state monitoring method

Abstract

The application relates to the technical field of intelligent monitoring of communication equipment and edge Internet of Things, and discloses a communication construction equipment state monitoring method based on edge computing. The method comprises the following steps: deploying an edge computing node at a communication construction equipment or a construction site; synchronously collecting equipment operation data through a multi-sensor synchronous data collection device arranged at the communication construction equipment; inputting an improved lightweight convolutional neural network model to determine fault feature matching degree and deviation data; calculating equipment health degree according to the fault feature matching degree and the deviation data; and determining an equipment state label according to the equipment health degree and the fault feature matching degree by the edge computing node, and returning the equipment state label to the cloud. Real-time state monitoring and fault early warning of the communication construction equipment in a weak network environment in the wild are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring of communication equipment and edge Internet of Things technology, and in particular to a method for monitoring the status of communication construction equipment based on edge computing. Background Technology

[0002] With the rapid advancement of 5G / 6G communication network construction, communication engineering construction sites face a wide variety of equipment and complex deployment environments, posing significant challenges to equipment status monitoring. Traditional solutions often employ a centralized cloud-based monitoring architecture, which suffers from the following technical problems: 1. High data transmission latency: Centralized cloud-based monitoring requires uploading all raw operating data of the equipment to a cloud server for processing. In weak network environments in the field, data transmission latency can reach 500-1000ms, failing to meet the real-time requirements of fault early warning. 2. High bandwidth consumption: Continuous collection of high-frequency data such as vibration, noise, power consumption, and temperature from multiple sensors generates massive amounts of raw data. Uploading all of this data puts enormous pressure on limited bandwidth in the field, easily leading to network congestion and data packet loss. 3. Poor real-time fault diagnosis: Cloud-based diagnosis relies on complete data upload before execution, resulting in significant delays in diagnostic result feedback, making it difficult to promptly detect potential equipment faults. 4. High redundancy of monitoring data: During normal equipment operation, the vast majority of raw data is invalid and redundant; uploading all of it wastes bandwidth resources. Summary of the Invention

[0003] Therefore, the purpose of this invention is to propose a method for monitoring the status of communication construction equipment based on edge computing, so as to solve at least one of the technical problems mentioned in the background art.

[0004] To address the aforementioned technical problems, the first aspect of this invention proposes a method for monitoring the status of communication construction equipment based on edge computing, comprising: Edge computing nodes are deployed on communication construction equipment or at the construction site, and the deployment location of the edge computing nodes covers the working area of ​​the communication construction equipment; Data on equipment operation is collected synchronously by multiple sensors deployed on communication construction equipment; Edge computing nodes acquire device operation data and input it into an improved lightweight convolutional neural network model to determine fault feature matching degree and deviation data; Calculate equipment health based on fault characteristic matching degree and deviation data; Edge computing nodes determine device status labels based on device health and fault characteristic matching degree, and then transmit them back to the cloud.

[0005] In one possible implementation, multiple sensors deployed on the communication construction equipment synchronously collect equipment operation data based on an adaptive sampling frequency; wherein, the adaptive sampling frequency is dynamically switched according to the equipment health status, the initial set equipment health status is used in the first monitoring cycle, and the equipment health status output in the previous monitoring cycle is used in subsequent monitoring cycles.

[0006] In one possible implementation, the equipment operation data includes vibration data, noise data, power consumption data, operating temperature data, and start / stop status data.

[0007] In one possible implementation, edge computing nodes acquire device operation data and input it into an improved lightweight convolutional neural network model to determine fault feature matching degree and deviation data, including: The equipment operation data is input into an improved lightweight convolutional neural network model, which outputs feature vectors. Calculate the cosine similarity between the feature vector and the preset fault feature library, take the maximum value and multiply by 100 to obtain the fault feature matching degree; Each type of sub-data in the equipment operation data is subjected to a deviation formula and normalized to obtain deviation data; the deviation data includes vibration deviation, noise deviation, power consumption deviation and temperature deviation.

[0008] In one possible implementation, the equipment health is calculated based on fault characteristic matching degree and deviation data, including: ; in, For vibration deviation; This is for noise bias; This refers to power consumption deviation; Temperature deviation; For fault feature matching degree; , , , , The weighting coefficient has a value range of 0-1 and satisfies the following conditions: .

[0009] In one possible implementation, before the edge computing node inputs device operation data into the improved lightweight convolutional neural network model, it also includes fusing and denoising the device operation data, including: The equipment operation data is processed by min-max normalization, converting data of different physical units into dimensionless values ​​in the range of 0-1; Calculate the fused noise reduction value: ; in, , , , The fusion weights corresponding to each data source and satisfying the following conditions: The value range is 0.1-0.4; , , , These are the dimensionless values ​​after normalization of vibration data, noise data, power consumption data, and operating temperature data, respectively. The environmental background noise estimation correction term is obtained by calibrating the signal fluctuation under the no-load state of the equipment. , , , , The standard deviations of the normalized vibration data, noise data, power consumption data, and temperature data are calculated based on a sliding window of the most recent M frames. , , , The corresponding weighting coefficients range from 0 to 0.05 and satisfy the following conditions: .

[0010] In one possible implementation, the edge computing node determines the device status label based on the device health and fault characteristic matching degree, including: When the equipment health level is greater than or equal to the preset first health level threshold and the fault feature matching degree is less than the preset first fault feature matching threshold, the equipment status label is normal operation; When the equipment health is greater than or equal to the preset second health threshold and less than the preset first health threshold, and the fault feature matching degree is greater than or equal to the preset first fault feature matching threshold and less than the preset second fault feature matching threshold, the equipment status label is slightly deteriorated. When the equipment health level is greater than or equal to the preset third health level threshold and less than the preset second health level threshold, and the fault feature matching degree is greater than or equal to the preset second fault feature matching degree threshold and less than the preset third fault feature matching degree threshold, the equipment status label is abnormal operation; When the equipment health is less than the preset third health threshold and the fault feature matching degree is greater than or equal to the preset third fault feature matching degree threshold, the equipment status label is fault operation.

[0011] In one possible implementation, edge computing nodes perform hierarchical data transmission back to the cloud based on a dynamic compression strategy matched to device health status, including: When the device health level is greater than or equal to the preset first health threshold, the first compression level is executed; When the device health level is greater than or equal to a preset second health threshold but less than a preset first health threshold, the second compression level is executed. When the device health level is greater than or equal to the preset third health threshold but less than the preset second health threshold, the third compression level is executed. When the device health level is lower than the preset third health level threshold, the fourth compression level is executed; The compression ratios of the first, second, third, and fourth compression levels decrease progressively.

[0012] In one possible implementation, the method also includes: The cloud receives the analysis results, abnormal feature data and incremental model parameters returned by the edge computing nodes, performs global data aggregation and model iterative optimization, and sends the updated parameters to the edge computing nodes through incremental parameter differential encoding. The cloud-based model iterative optimization includes an incremental parameter update mechanism: During the local diagnosis process, when the detected feature vector has a cosine similarity to the most similar vector in the preset fault feature library that is lower than a preset threshold, it is judged as a suspected new fault feature. The new feature and its corresponding model parameter update amount are reported to the cloud as incremental parameters. The preset threshold is determined based on the lowest matching degree of the covered fault samples. Incremental parameters are transmitted only after differential encoding of parameter blocks that have changed relative to the previous version of the model; parameter blocks that have not changed are not transmitted. The incremental parameters reported by each edge computing node in the cloud are weighted and aggregated according to the proportion of sample size contributed by each edge computing node. The aggregated model is first evaluated on the cloud validation set. If the diagnostic accuracy is not lower than that of the model before aggregation, the updated parameters are sent to each edge computing node in the form of differential encoding; otherwise, the original parameters are kept.

[0013] In one possible implementation, the improved lightweight convolutional neural network model has a six-layer structure, as follows: The first layer is a one-dimensional deep convolutional layer, which performs channel-by-channel convolution on multi-source time-series signals to extract local time-series features of each channel. The second layer is a pointwise convolutional layer, which achieves feature fusion between channels through 1×1 convolution; The third layer is a max pooling layer, which performs downsampling and preserves significant features; The fourth layer is a depthwise separable convolutional layer, which extracts high-level abstract features; The fifth layer is a global average pooling layer, which compresses the feature map into a feature vector; The sixth layer is a fully connected output layer that outputs the probability distribution of fault types.

[0014] This invention addresses the problem of monitoring failures in weak network environments at communication construction sites using existing technologies. By deploying edge computing nodes on communication construction equipment or at the construction site, it completes fusion noise reduction, fault diagnosis, and health measurement locally at the edge, improving the efficiency from data acquisition to diagnostic result output and overcoming the shortcomings of cloud-based state assessment and network-constrained real-time performance. Through min-max normalization and fusion noise reduction processing including background noise correction for idle calibration environment, it integrates four heterogeneous physical quantities—vibration, noise, power consumption, and temperature—into a unified dimensionless fused state quantity. This eliminates environmental interference while preserving effective equipment operating signals, improving the accuracy of input signals. Furthermore, it integrates equipment health and fault characteristics... The dual-dimensional state determination of matching degree improves the accuracy of state assessment at the edge. The sampling frequency and compression level are dynamically adjusted by the same device health, forming a closed loop where the lower the health, the denser the sampling, the less compression, and the more complete the backhaul. This avoids the defect of independent settings for sampling and compression, which leads to decoupling from the degree of device degradation. Through a closed-loop co-evolution mechanism of edge discovery of new fault features, incremental parameter differential coding backhaul, cloud-based weighted aggregation iteration, verification rollback protection, and differential distribution, the model is continuously optimized with minimal backhaul traffic. This overcomes the defects of cloud-based unidirectional downhaul updates that lack differential coding, weighted aggregation, and rollback protection. Real-time status monitoring and fault early warning of communication construction equipment are realized.

[0015] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 is a flowchart of a communication construction equipment status monitoring method based on edge computing according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be described below with reference to the accompanying drawings. In the description of this application, unless otherwise stated, " / " indicates that the objects before and after are in an "or" relationship; for example, A / B can represent A or B. "And / or" in this application merely describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone, where A and B can be singular or plural. Furthermore, in the description of this application, unless otherwise stated, "multiple" refers to two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. Additionally, to facilitate a clear description of the technical solutions of the embodiments of this application, the terms "first," "second," etc., are used in the embodiments of this application to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first," "second," etc., do not limit the quantity or order of execution, and that "first," "second," etc., are not necessarily different. Furthermore, in the embodiments of this application, words such as "exemplary" or "for example" are used to indicate that something is being used as an example, illustration, or description. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner for ease of understanding.

[0019] Furthermore, the business scenarios described in the embodiments of this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided in the embodiments of this application. As those skilled in the art will know, with the emergence of new business scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0020] like Figure 1 As shown, this invention proposes a method for monitoring the status of communication construction equipment based on edge computing, including steps S1-S5: S1. Deploy edge computing nodes on communication construction equipment or at the construction site, wherein the deployment location of the edge computing nodes covers the working area of ​​the communication construction equipment; S2. Synchronously collect equipment operation data through multiple sensors deployed on communication construction equipment; S3. Edge computing nodes acquire equipment operation data and input them into an improved lightweight convolutional neural network model to determine fault feature matching degree and deviation data. S4. Calculate the equipment health status based on the fault characteristic matching degree and deviation data; S5: Edge computing nodes determine device status tags based on device health and fault characteristic matching degree, and then transmit them back to the cloud.

[0021] In this embodiment, communication construction equipment refers to the on-site work equipment used in communication engineering projects such as base station construction, pipeline laying, and equipment room construction, such as drilling rigs, generators, and air compressors. These types of equipment operate in the field environment for extended periods, with frequent start-ups and shutdowns and large load fluctuations. Their operating parameters, such as vibration, noise, power consumption, and temperature, can reflect the health status and fault development trends of the equipment.

[0022] In this embodiment, an edge computing node refers to an embedded computing unit deployed on a communication construction equipment terminal or at a construction site, possessing local data processing capabilities. It is used to perform noise reduction, feature extraction, fault diagnosis, and health status calculation locally near the data source. Node initialization includes loading pre-trained parameters of an improved lightweight convolutional neural network model, loading a preset fault feature library, configuring the acquisition parameters of each sensor, and calibrating the baseline value of the environmental background noise correction term.

[0023] In this embodiment, "deployment location covering the communication construction equipment's operating area" means that the signal coverage between the edge computing node and the monitored equipment can cover the entire operating area. In actual deployment, the distance between the node and the monitored equipment is generally controlled within 50 meters.

[0024] In this embodiment, the multiple sensors include an accelerometer, an acoustic sensor, a current transformer, a temperature sensor, and a device start / stop signal interface, which are used to collect vibration data (unit m / s²), noise data (unit dB), power consumption data (unit W), operating temperature data (unit °C), and start / stop status data, respectively. The start / stop status data is a switch signal that represents whether the device is currently in a running or stopped state.

[0025] In this embodiment, the improved lightweight convolutional neural network model refers to a fault diagnosis model deployed locally on an edge computing node. Its input is normalized multi-channel time-series data, and its output is the probability distribution of fault types.

[0026] In this embodiment, determining the fault feature matching degree and deviation data means that after the model outputs a feature vector, it compares the similarity with the feature vector in the preset fault feature library to obtain the fault feature matching degree, which is used to measure the degree of conformity between the current operating state and the known fault mode; the deviation data reflects the degree of deviation of each operating parameter relative to the equipment health benchmark. The fault feature matching degree is given by the model and the deviation data is given by actual measurement calculation, serving as two input evidences for health calculation.

[0027] In this embodiment, the preset fault feature library refers to the set of fault feature vectors stored locally on the edge computing node. The feature vectors are extracted from various fault samples during the model training phase, and each record is associated with a fault type label.

[0028] In this embodiment, the equipment health score refers to the quantitative score obtained by weighted fusion of fault feature matching degree and deviation data of each path. The score ranges from 0 to 100. The higher the score, the better the equipment's operating status, providing a unified comparison benchmark between different devices and different monitoring cycles.

[0029] In this embodiment, the device status label refers to the status conclusion given by the edge computing node based on two dimensions: device health and fault feature matching degree, such as normal operation, slight degradation, abnormal operation, and fault operation. The status label is used to drive subsequent sampling frequency switching and data backhaul strategy.

[0030] The beneficial effects of the above technologies are as follows: edge computing nodes acquire equipment operation data and input it into an improved lightweight convolutional neural network model to determine fault feature matching degree and deviation data; equipment health is calculated based on fault feature matching degree and deviation data; and equipment status labels are determined through equipment health and fault feature matching degree, which facilitates improved accuracy of equipment status monitoring and enables accurate early warning.

[0031] In one possible implementation, multiple sensors deployed on the communication construction equipment synchronously collect equipment operation data based on an adaptive sampling frequency; wherein, the adaptive sampling frequency is dynamically switched according to the equipment health status, the initial set equipment health status is used in the first monitoring cycle, and the equipment health status output in the previous monitoring cycle is used in subsequent monitoring cycles.

[0032] In this embodiment, the adaptive sampling frequency refers to a data acquisition mechanism where the sampling frequency changes according to the device's health status. A set of exemplary sampling levels is provided: 10Hz sampling is used when the device is operating normally (device health H ≥ 90) to reduce energy consumption and data processing burden; 20Hz is increased during mild degradation (70 ≤ H < 90) to capture early degradation characteristics; 35Hz is increased during abnormal operation (50 ≤ H < 70) to capture abnormal operation details; and 50Hz is increased during fault operation (H < 50) to capture transient fault characteristics. It should be noted that the above sampling level values ​​are exemplary and can be adjusted according to the device type and site conditions during actual deployment.

[0033] In this embodiment, the initial setting of device health in the first monitoring cycle means that when the system has no health output, the initial setting of device health is set to the value corresponding to normal operation (e.g., H=100 for example). In subsequent monitoring cycles, the sampling frequency of the current cycle is determined by the device health output of the previous cycle, forming an adaptive adjustment that rolls cycle by cycle.

[0034] In this embodiment, to ensure data continuity before and after the sampling frequency switch, the edge computing node maintains a circular buffer. When a change in device health is detected and triggers the sampling frequency switch, the previous N frames of historical data in the buffer are spliced ​​with the newly collected data after the switch. N is the preset number of buffer frames. In this embodiment, 50 frames are used to avoid timing breaks caused by the sampling frequency switch.

[0035] The beneficial effects of the above technologies are as follows: the sampling frequency is automatically adjusted according to the health status of the equipment. When the equipment is in good condition, it operates at a low frequency to save energy and computing power. When the equipment deteriorates, it automatically encrypts the sampling to capture fault details, realizing adaptive monitoring with higher risk and more detailed monitoring. The rolling connection and switching buffer mechanism of the health status of the previous and subsequent cycles ensures the smoothness of the adaptive process and the integrity of the data.

[0036] In one possible implementation, the equipment operation data includes vibration data, noise data, power consumption data, operating temperature data, and start / stop status data.

[0037] In this embodiment, vibration data is collected by an accelerometer installed on the equipment housing or transmission part, reflecting the mechanical state of the rotating and transmission components. Faults such as bearing wear, loose fasteners, and load imbalance are usually the first to appear in the vibration signal.

[0038] In this embodiment, noise data is collected by an acoustic sensor to help identify acoustic signs such as mechanical friction and abnormal airflow. Power consumption data is obtained by collecting the operating current of the equipment through a current transformer and converting it. Electrical faults such as motor overheating, circuit short circuit, and abnormal load will cause characteristic changes in the power consumption curve. Operating temperature data is collected from key components such as motor windings and bearing housings using temperature sensors. Start-stop status data is collected through the device start-stop signal interface and is used to identify whether the device is currently running or stopped.

[0039] The beneficial effects of the above technologies are: they can characterize the equipment status from five aspects: mechanical, acoustic, electrical, thermal and operating conditions, and can cover more comprehensive fault monitoring compared to single-type sensor data.

[0040] In one possible implementation, edge computing nodes acquire device operation data and input it into an improved lightweight convolutional neural network model to determine fault feature matching degree and deviation data, including: The equipment operation data is input into an improved lightweight convolutional neural network model, which outputs feature vectors. Calculate the cosine similarity between the feature vector and the preset fault feature library, take the maximum value and multiply by 100 to obtain the fault feature matching degree; Each type of sub-data in the equipment operation data is subjected to a deviation formula and normalized to obtain deviation data; the deviation data includes vibration deviation, noise deviation, power consumption deviation and temperature deviation.

[0041] In this embodiment, the feature vector refers to the fixed-length vector representation formed before the output layer after the model extracts features from the input multi-channel time-series data layer by layer, representing the key mode information of the device's operating status during the current monitoring period.

[0042] In this embodiment, the formula for calculating cosine similarity is: ; in, The cosine similarity between the feature vector and the feature vector of the preset fault feature library; The feature vector output by the model. The model features are derived from a pre-defined fault feature library. The features are activated using ReLU to ensure non-negativity. The value range is 0-1. The fault feature matching degree is obtained by calculating the maximum value of each feature vector in the feature library and multiplying it by 100 to map it to the 0-100 range. : ; The larger the value, the higher the degree of agreement between the current operating state and the known fault mode.

[0043] In this embodiment, the deviation formula takes vibration deviation as an example: ; in, These are measured vibration values. This refers to the vibration baseline value, which is the health baseline calibrated after the equipment has passed acceptance testing or during normal operation. This refers to the vibration range. The calculation methods for noise deviation, power consumption deviation, and temperature deviation are similar.

[0044] An exemplary approach to setting a baseline is to collect at least 7 days of normal operating data, take the statistical median of each parameter, and recalibrate after equipment maintenance or component replacement. When the range calibration is abnormal (e.g., the range is 0), the factory range of that sensor is used as a fallback to avoid calculation overflow that could lead to abnormal health calculation.

[0045] In this embodiment, normalization means unifying the deviation values ​​of each path to the range of 0-100, so that the four deviations with different physical units can participate in the equipment health calculation on the same numerical scale as the fault characteristic matching degree.

[0046] The beneficial effects of the above technologies are as follows: the fault feature matching degree reflects the confidence of the deep model in identifying the fault mode, the deviation data reflects the measured deviation of the operating parameters relative to the health benchmark, and the two sources of evidence complement and fuse. Compared with the method of simply relying on model output or simply relying on threshold judgment, the risk of misjudgment and missed judgment is significantly reduced.

[0047] In one possible implementation, the equipment health is calculated based on fault characteristic matching degree and deviation data, including: ; in, For vibration deviation; This is for noise bias; This refers to power consumption deviation; Temperature deviation; For fault feature matching degree; , , , , The weighting coefficient has a value range of 0-1 and satisfies the following conditions: .

[0048] In this embodiment, when the matching degree between each deviation and the fault characteristic is taken as the maximum value of 100, That is, the worst state corresponds to a device health level of 0, and the best state corresponds to a device health level of 100. Device health level has become a unified indicator that can be directly compared across devices and across cycles.

[0049] In this embodiment, the method for calibrating the weighting coefficients is as follows: For the target equipment type, multi-source data under various fault conditions are collected, and the accuracy rate for fault identification using only a single deviation parameter or fault feature matching degree is calculated. This accuracy rate is used as the contribution of the corresponding parameter, and the contribution of each parameter is normalized and weights are allocated proportionally. Different equipment types have different fault mechanisms, and therefore different weight configurations. For example, the exemplary weight configuration for a drilling rig is... , , , , An example weight configuration for the generator is as follows: , , , , .

[0050] For example, if a drilling rig measures ΔV=20, ΔN=15, ΔP=30, and ΔT=10 within a certain monitoring period, and the fault characteristic matching degree S=40, and the weight is taken as the above drilling rig configuration, then H = 100 -(0.30×20 + 0.20×15 + 0.20×30 +0.15×10 + 0.15×40) = 100 - 22.5 = 77.5, which falls into the slightly deteriorated range. The system switches the sampling frequency and feedback strategy accordingly.

[0051] The beneficial effects of the above technologies are as follows: the health formula compresses the matching degree between the four measured deviations and the fault characteristics output by the model into a single comparable index, which has a small computational load and intuitive physical meaning, and is suitable for real-time execution on resource-constrained edge nodes on a cycle basis; the weights are calibrated separately according to the equipment type, so that the health can adapt to the differences in fault mechanisms of different equipment such as drilling rigs and generators, and provides a unified quantitative basis for sampling frequency switching and compression strategy matching.

[0052] In one possible implementation, before the edge computing node inputs device operation data into the improved lightweight convolutional neural network model, it also includes fusing and denoising the device operation data, including: The equipment operation data is processed by min-max normalization, converting data of different physical units into dimensionless values ​​in the range of 0-1; Calculate the fused noise reduction value: ; in, , , , The fusion weights corresponding to each data source and satisfying the following conditions: The value range is 0.1-0.4; , , , These are the dimensionless values ​​after normalization of vibration data, noise data, power consumption data, and operating temperature data, respectively. The environmental background noise estimation correction term is obtained by calibrating the signal fluctuation under the no-load state of the equipment. , , , , The standard deviations of the normalized vibration data, noise data, power consumption data, and temperature data are calculated based on a sliding window of the most recent M frames. , , , The corresponding weighting coefficients range from 0 to 0.05 and satisfy the following conditions: .

[0053] In this embodiment, the formula for calculating min-max normalization is: ; in, The original data, and These are the historical minimum and maximum values ​​for this data path, respectively. This is the dimensionless value after normalization.

[0054] In this embodiment, the fusion weight avoids the loss of contribution in the fusion due to the low weight of a certain data path, and also avoids the dominance of a single data path in the fusion result; the specific allocation is determined according to the significance of each data path in the symptoms of different fault types. For example, vibration signals are usually the most direct symptoms of faults in rotating and transmission components, and their fusion weight is taken as a higher value in most equipment types.

[0055] In this embodiment, the environmental background noise estimation correction term This is used solely to eliminate ambient background noise, not to suppress valid signals from the device. The calibration method is as follows: After the device is powered on, it is run without load for a predetermined time (5 minutes for example). Each signal is continuously collected, and its standard deviation is calculated as the background noise level of that signal. This standard deviation is then multiplied by the corresponding weighting coefficient to obtain the final signal level. The calibration reference value is used; when the equipment is actually running, the estimated background noise value is subtracted from the fused signal to retain the effective operating signal characteristics of the equipment. To adapt to the slow drift of background noise at the construction site (such as day-night temperature difference, changes in surrounding construction machinery, etc.), the edge computing node automatically recalibrates and updates the reference value during the idle period of the first power-on each day.

[0056] In this embodiment, the sliding window is updated frame by frame as new data frames arrive. In this embodiment, M is 100 frames. The weighting coefficient is limited to a range of 0-0.05 and the sum does not exceed 0.2 to ensure that the background noise correction amplitude does not exceed a reasonable proportion of the fusion value and to avoid excessive subtraction that could damage the effective signal of the device.

[0057] In this embodiment, due to To reduce items, Theoretically, it could be negative under low signal and high background noise conditions, therefore Before participating in subsequent calculations, the interval is truncated to [0, 1], i.e., the value is taken as follows: This truncation is only used as a numerical stability guarantee and does not change the definition of the fusion denoising formula itself.

[0058] In this embodiment, the multi-source data fusion denoising formula and the health formula are applied synergistically. The multi-source data fusion denoising formula first normalizes data from different physical units into dimensionless values ​​and merges them into a unified standardized output, eliminating the problem of inconsistent dimensions. Similarly, the deviation parameters in the health formula also use a normalization range of 0-100, ensuring consistency between the two at the data standardization level. This results in no scale conversion loss in the entire data stream from denoising to health calculation. Furthermore, This method only eliminates environmental background noise rather than suppressing effective signals, ensuring the quality of the input signal for health calculation and thus enabling the health score to more accurately reflect the true state of the device. Using the noise reduction formula alone without the health formula, or using the health formula alone with unreduced data input, will not achieve the aforementioned synergistic effect.

[0059] In this embodiment, the output results are compressed and transmitted via a compression module; based on the output results, corresponding alarm prompts are issued via a local alarm module. The device operation data consists of multi-source raw data.

[0060] The beneficial effects of the above technologies are as follows: Based on the cascaded processing of normalization, weighted fusion and background noise correction of no-load calibration, four heterogeneous physical quantities are integrated into a unified dimensionless fused state quantity, eliminating the problem of dimension inconsistency. By eliminating only ambient noise without suppressing effective signals, this approach preserves effective fluctuations caused by changes in equipment status, compared to methods that directly use the standard deviation of the operating signal for bandpass suppression. This improves the accuracy of data input. Furthermore, the dual-constraint design of fusion weights and noise correction coefficients balances fusion sensitivity and numerical stability.

[0061] In one possible implementation, the edge computing node determines the device status label based on the device health and fault characteristic matching degree, including: When the equipment health level is greater than or equal to the preset first health level threshold and the fault feature matching degree is less than the preset first fault feature matching threshold, the equipment status label is normal operation; When the equipment health is greater than or equal to the preset second health threshold and less than the preset first health threshold, and the fault feature matching degree is greater than or equal to the preset first fault feature matching threshold and less than the preset second fault feature matching threshold, the equipment status label is slightly deteriorated. When the equipment health level is greater than or equal to the preset third health level threshold and less than the preset second health level threshold, and the fault feature matching degree is greater than or equal to the preset second fault feature matching degree threshold and less than the preset third fault feature matching degree threshold, the equipment status label is abnormal operation; When the equipment health is less than the preset third health threshold and the fault feature matching degree is greater than or equal to the preset third fault feature matching degree threshold, the equipment status label is fault operation.

[0062] In this embodiment, a set of exemplary values ​​are given for the preset health threshold and the preset fault feature matching threshold: the preset first, second, and third health thresholds are 90, 70, and 50, respectively, and the preset first, second, and third fault feature matching thresholds are 30, 60, and 80, respectively. As shown in Table 1, the edge computing node determines the device status label; Table 1

[0063] H≥90 S<30 Normal operation Level 1 70≤H<90 30≤S<60 Mild degradation Level 2 50≤H<70 60≤S<80 Abnormal operation Level 3 H<50 S≥80 Fault operation Level 4: Emergency Fault

[0064] In this embodiment, when the equipment health and fault characteristics do not fall into the same level range, for example, H=85 (belonging to the mild deterioration range) and S=70 (belonging to the abnormal operation range), the worse level is taken as the final judgment result by the conservative principle, that is, it is judged as abnormal operation.

[0065] The beneficial effects of the above technologies are as follows: equipment health reflects the cumulative deviation of operating parameters relative to the health benchmark, fault feature matching degree reflects the matching confidence of the deep model to the specific fault mode, and the cross-validation of the two reduces the probability of misjudgment; the conservative principle ensures that potential faults are not missed when the conclusions of the two dimensions differ, thus improving the accuracy of fault judgment.

[0066] In one possible implementation, edge computing nodes perform hierarchical data transmission back to the cloud based on a dynamic compression strategy matched to device health status, including: When the device health level is greater than or equal to the preset first health threshold, the first compression level is executed; When the device health level is greater than or equal to a preset second health threshold but less than a preset first health threshold, the second compression level is executed. When the device health level is greater than or equal to the preset third health threshold but less than the preset second health threshold, the third compression level is executed. When the device health level is lower than the preset third health level threshold, the fourth compression level is executed; The compression ratios of the first, second, third, and fourth compression levels decrease progressively.

[0067] In this embodiment, the exemplary configurations of each compression level are as follows: The first compression level corresponds to an ultra-high compression ratio, i.e., 80%~90% compression ratio, and only transmits health scores, device numbers, runtime, and status logs; the second compression level corresponds to a medium compression ratio, i.e., 50%~70% compression ratio, and transmits simplified feature data and regular operation logs; the third compression level corresponds to a low compression ratio, i.e., 20%~40% compression ratio, and uploads core fault feature data, abnormal deviations, and alarm records; the fourth compression level corresponds to uncompressed full upload, i.e., uncompressed high-fidelity transmission, while triggering the highest level fault alarm and audible and visual alarm linkage. The compression ratio mentioned here refers to the percentage of data reduction; i.e., an 80%~90% compression ratio corresponds to transmitting 10%~20% of the original collected data, a 50%~70% compression ratio corresponds to transmitting 30%~50% of the data, and a 20%~40% compression ratio corresponds to transmitting 60%~80% of the data. The cloud also refers to a cloud-based management platform.

[0068] In this embodiment, the compression methods at each level are selected according to the characteristics of the returned content: structured short data such as health scores and device numbers are directly returned after conventional lossless compression; simplified feature data are low-dimensional feature vectors extracted from edge nodes, and lossless compression is used to ensure feature integrity; core fault feature data under abnormal and fault conditions and original time series data prioritize data integrity rather than compression ratio.

[0069] In this embodiment, the four-level compression strategy is linked with the adaptive sampling frequency: the decrease in device health not only increases the sampling frequency step by step to obtain more intensive monitoring data, but also gradually relaxes the compression rate and expands the returned content from extremely simple information to full data, forming a closed loop where the lower the health, the more intensive the sampling, the less compression, and the more complete the returned data.

[0070] The beneficial effects of the above technologies are as follows: the compression intensity is matched according to the health status of the equipment; only minimal information is transmitted back when the equipment is running normally, which greatly reduces the amount of redundant data transmission; when the equipment is in failure, the full data is uploaded to ensure the data integrity of the post-analysis, and the bandwidth resources are allocated on demand in the time and equipment dimensions; the closed-loop linkage with the adaptive sampling frequency enables the system to operate at low cost during low-risk periods and automatically switch to fine monitoring mode during high-risk periods, so that the adaptive adjustment of the monitoring intensity can be completed without manual intervention.

[0071] In one possible implementation, the method also includes: The cloud receives the analysis results, abnormal feature data and incremental model parameters returned by the edge computing nodes, performs global data aggregation and model iterative optimization, and sends the updated parameters to the edge computing nodes through incremental parameter differential encoding. The cloud-based model iterative optimization includes an incremental parameter update mechanism: During the local diagnosis process, when the detected feature vector has a cosine similarity to the most similar vector in the preset fault feature library that is lower than a preset threshold, it is judged as a suspected new fault feature. The new feature and its corresponding model parameter update amount are reported to the cloud as incremental parameters. The preset threshold is determined based on the lowest matching degree of the covered fault samples. Incremental parameters are transmitted only after differential encoding of parameter blocks that have changed relative to the previous version of the model; parameter blocks that have not changed are not transmitted. The incremental parameters reported by each edge computing node in the cloud are weighted and aggregated according to the proportion of sample size contributed by each edge computing node. The aggregated model is first evaluated on the cloud validation set. If the diagnostic accuracy is not lower than that of the model before aggregation, the updated parameters are sent to each edge computing node in the form of differential encoding; otherwise, the original parameters are kept.

[0072] In this embodiment, a suspected new fault feature refers to a feature in which the cosine similarity between the feature vector output by the model and the most similar vector in the preset fault feature library is lower than a preset threshold during local diagnosis of the edge node. This indicates that the feature does not belong to a known fault mode. The preset threshold is determined based on the minimum matching degree of the fault samples already covered on the test set, and is 0.6 in this embodiment (exemplary value).

[0073] In this embodiment, the model parameter update amount refers to the change in model parameters relative to the previous version after the edge node performs small-sample fine-tuning locally for new fault features. This change is reported to the cloud along with the new features as incremental parameters.

[0074] In this embodiment, the differential coding adopts a sparse differential method, which only encodes and transmits parameter blocks whose parameters have changed relative to the previous version of the model. Unchanged parameter blocks are not transmitted. The updated parameters are also sent out in differential coding form, minimizing the transmission traffic of parameter synchronization.

[0075] In this embodiment, weighted aggregation means that when the incremental parameters reported by each edge computing node are aggregated in the cloud, the weight of each node's incremental parameters is the proportion of the sample size contributed by that node. Nodes with larger sample sizes have a greater impact on the global model. Rollback protection means that the aggregated model is first evaluated on the cloud validation set, and updates are only issued when the diagnostic accuracy is not lower than that of the model before aggregation. Otherwise, the original parameters are maintained to prevent the degraded parameters caused by data noise in individual nodes from spreading to the entire network.

[0076] This embodiment implements the synergistic application of a hierarchical compression strategy and cloud-based model iteration: the edge only transmits high-value data and incremental parameters, while the cloud performs global model iteration and optimization based on the incremental parameters and issues updated parameters, forming a closed-loop co-evolution mechanism of edge diagnosis, incremental reporting, cloud iteration, and parameter issuance. This mechanism ensures continuous model optimization while minimizing bandwidth consumption, thus improving bandwidth efficiency and model accuracy.

[0077] The beneficial effects of the above technologies are as follows: the incremental parameter update mechanism enables new fault features discovered locally by edge nodes to feed back into the global model; after cloud aggregation and iteration, all edge nodes share the benefits of model evolution, forming a closed-loop collaborative evolution of edge diagnosis, incremental reporting, cloud iteration, and parameter distribution; sparse differential coding minimizes update traffic, and the rollback protection mechanism ensures that each version of the model distributed to edge nodes is verified by the validation set, and the model quality will not degrade during the iteration process.

[0078] In one possible implementation, the improved lightweight convolutional neural network model has a six-layer structure, as follows: The first layer is a one-dimensional deep convolutional layer, which performs channel-by-channel convolution on multi-source time-series signals to extract local time-series features of each channel. The second layer is a pointwise convolutional layer, which achieves feature fusion between channels through 1×1 convolution; The third layer is a max pooling layer, which performs downsampling and preserves significant features; The fourth layer is a depthwise separable convolutional layer, which extracts high-level abstract features; The fifth layer is a global average pooling layer, which compresses the feature map into a feature vector; The sixth layer is a fully connected output layer that outputs the probability distribution of fault types.

[0079] In this embodiment, the one-dimensional depth convolutional layer independently convolves the input multi-channel temporal signal channel by channel to extract the local temporal features of each channel without information mixing between channels; the pointwise convolutional layer linearly combines the feature values ​​of each channel at the same time through 1×1 convolution to achieve feature fusion between channels, and after being concatenated with the one-dimensional depth convolution, it forms a depthwise separable convolution.

[0080] In this embodiment, the max pooling layer downsamples the fused features, retaining only the maximum activation value within each pooling window, thus compressing the feature length while preserving significant fault features; the depthwise separable convolutional layer extracts high-level features across channels and time periods at a higher level of abstraction; the global average pooling layer averages each feature channel over the entire time dimension, compressing the feature map into a fixed-length feature vector, replacing the traditional large-scale fully connected layer, thus reducing the number of parameters while suppressing overfitting; the fully connected output layer activates the output fault type probability distribution with softmax, and takes the category with the highest probability as the fault type identification result.

[0081] In this embodiment, the fault types provide a set of 12 exemplary fault types, namely bearing wear, motor overheating, mechanical jamming, abnormal power consumption, insufficient lubrication, loose fasteners, short circuit, insulation aging, load imbalance, cooling failure, excessive bearing clearance, and seal failure.

[0082] In this embodiment, the number of model parameters is compressed to 15%~20% of that of a traditional convolutional neural network model, and the inference latency is controlled within 20ms. Under test conditions, the single inference time in this embodiment is 8~18ms. Compression is achieved through structured channel pruning: the L1 norm of the weights of each convolutional channel is used as the channel contribution, and the channel with the lowest contribution is pruned according to a preset pruning ratio. Then, the pruned model is fine-tuned with the retained training data to restore accuracy. The model can be deployed only if the diagnostic accuracy on the test set after fine-tuning is not less than 95% of that before pruning. The training dataset consists of multi-source time-series signals of vibration, noise, power consumption, and temperature collected under the actual operating conditions of communication construction equipment. Each fault type includes normal samples and corresponding fault samples, divided into training set, validation set, and test set, with an exemplary ratio of 7:2:1.

[0083] The beneficial effects of the above technologies are as follows: with depthwise separable convolution as the core, while maintaining the ability to classify multiple types of faults, the number of parameters is compressed to 15%~20% of that of traditional convolutional neural network models, and the inference latency is controlled within 20ms, enabling the fault diagnosis model to run in real time on low-power edge computing nodes in the field; global average pooling improves the robustness of the model to the sliding of the sampling window while compressing parameters, and the compression method combining channel pruning and fine-tuning ensures that the diagnostic accuracy is controllable during the lightweight process.

[0084] In another embodiment of the invention, it further includes: For each device operation data, the normalized data is processed by the 3σ principle to remove outliers. The removed positions are filled by linear interpolation of the adjacent valid frames. When the continuous failure time of a certain sensor exceeds the preset time limit, the weight corresponding to the deviation data of that channel is temporarily set to zero in the device health calculation, and the remaining weights are normalized proportionally. The output device health is accompanied by a reduced accuracy flag and sent back to the cloud through the hierarchical feedback mechanism. Edge computing nodes maintain a local ring buffer. When a backhaul fails, the data to be backhauled is written into the ring buffer. After the network is restored, the data is resumed from the breakpoint according to the priority of the status level. When the buffer overflows, the data with the lowest level and the earliest time is evicted first. Each sensor and edge computing node synchronizes time through a time synchronization protocol. Each frame of data is attached with a timestamp. The operating data of multiple source devices are aligned according to the timestamps before fusion and noise reduction are performed. Frames with timestamp deviations exceeding the tolerance are discarded. During the interruption of the link between the edge computing node and the cloud, data acquisition, fusion and noise reduction, fault diagnosis, device health calculation and hierarchical caching are performed autonomously on the edge computing node. After the link is restored, cached data is automatically resumed and model parameters are synchronized.

[0085] In this embodiment, the 3σ principle outlier removal means that for each normalized data, sampling points that deviate from the mean of the sliding window of the nearest M frames of that path by more than 3 times the standard deviation are judged as outliers and removed. The removed positions are filled by linear interpolation of the adjacent valid frames before and after, avoiding outlier pollution in the fusion noise reduction and health calculation.

[0086] In this embodiment, when the continuous failure time of a certain sensor exceeds the preset time limit (10 seconds in this embodiment), the edge node records the failure event in the local log, and temporarily sets the weight corresponding to the deviation data of that path to zero in the device health calculation, and normalizes the remaining weights proportionally to keep the weight sum to 1. The device health output in this state is accompanied by a reduced accuracy flag, which is sent back to the cloud through the hierarchical feedback mechanism to prompt the maintenance personnel to inspect the sensor.

[0087] In this embodiment, the capacity of the ring buffer is designed to store high-value backhaul data from the most recent 24 hours (exemplary value); when backhaul fails, the data to be backhauled is written into the ring buffer, and after the network recovers, the data is resumed from the breakpoint according to the priority of the status level (fault operation is higher than abnormal operation, abnormal operation is higher than slight degradation, and slight degradation is higher than normal operation) to ensure that emergency fault data arrives at the cloud first; when the buffer overflows, the lowest level and earliest data are prioritized for elimination to retain data from high-risk periods to the maximum extent.

[0088] In this embodiment, the time synchronization protocol adopts the network time protocol, and the precise time protocol is adopted when the local area network conditions at the construction site allow. After each sensor synchronizes with the edge computing node, a timestamp is attached to each frame of data. Multi-source data are aligned according to the timestamps and then fusion and noise reduction are performed. Frames with timestamp deviations greater than 50ms (exemplary value) are discarded and counted. The discard count is used as a data quality indicator and is transmitted back with the operation log.

[0089] In this embodiment, during the interruption of the link between the edge computing node and the cloud, data acquisition, fusion and noise reduction, fault diagnosis, device health calculation and hierarchical caching are carried out autonomously locally, and the monitoring and alarm functions are not interrupted due to the link failure. When the interruption time exceeds the preset time limit (72 hours in this embodiment) and the cache is about to be full, the earliest data is overwritten locally according to the priority of the status level. After the link is restored, the cached data is automatically resumed and the model parameters are synchronized.

[0090] The beneficial effects of the above technologies are that they enable usable monitoring results to be obtained even in abnormal situations such as single-channel sensor failure, weak network backhaul failure, time asynchrony of multi-source data, and disconnection between the edge and the cloud.

[0091] In another embodiment of the invention, it further includes: When the device is in a shutdown state, the edge computing node suspends fault diagnosis inference and temporarily reduces the data acquisition frequency to the lowest level. After the device is restarted, the data acquisition frequency and diagnosis process are restored according to the current device health status. Combine noise reduction values Enter the sliding window trend monitor, when When the deviation from the normal operating baseline reaches a preset number of frames, a diagnostic review is triggered, and the trend event is recorded in the operation log and sent back to the cloud along with the returned data.

[0092] In this embodiment, when the device is in a shutdown state (start-stop state data is shutdown), the edge computing node suspends fault diagnosis inference and temporarily reduces the data acquisition frequency to the lowest level (5Hz in this embodiment, an exemplary value) to retain only environmental monitoring, so as to avoid environmental noise and residual vibration during the shutdown period being misjudged as abnormal signals of the device; after the device restarts, the data acquisition frequency and diagnostic process matched to the current device health status are restored.

[0093] In this embodiment, vibration and power consumption data collected during the transient period after the device starts up (exemplarily the first 60 seconds after startup) are separately marked as start-up and shutdown transient samples. If the vibration peak value during this period exceeds three times the normal steady-state vibration peak value, and the time required for the vibration to reach a steady state exceeds twice the normal startup time, it is judged as a suspected mechanical jamming or load imbalance feature, and the diagnostic review priority of the device in the next monitoring cycle is increased, so that the fault information contained in the start-up and shutdown process can be utilized rather than simply filtered out.

[0094] In this embodiment, the sliding window trend monitor maintains the fused noise reduction value. The time-series cache has a window length W equal to a preset number of frames (300 frames in this embodiment), and calculations are based on this window. The real-time average and fluctuation band form a normal operation benchmark band; when When the deviation from the normal operating baseline reaches a preset number of frames (30 consecutive frames in this embodiment), regardless of whether the deviation of a single channel exceeds the limit, the edge computing node will trigger a diagnostic review and record the trend event in the operation log, and transmit it back to the cloud with the data according to the current compression level.

[0095] The beneficial effects of the above technologies are as follows: shutdown gating avoids misjudgment of shutdown noise and reduces the data acquisition and calculation overhead during shutdown; start-stop transient feature mining utilizes the windows where faults such as mechanical jamming and load imbalance are most easily exposed, transforming the originally discarded transient data into diagnostic evidence. Trend monitoring provides early warnings when the deviation of a single channel has not yet exceeded the limit, but the multi-source fusion status has already experienced a continuous shift. This forms a two-level warning system that complements the health assessment, improving the timeliness and accuracy of fault warnings.

[0096] This invention proposes a status monitoring method for communication construction equipment based on edge computing. Through local real-time diagnosis at the edge, the end-to-end latency from data acquisition to diagnostic result output is significantly reduced. Differential data compression and hierarchical backhaul significantly reduce data transmission bandwidth consumption. Multi-source data fusion noise reduction and an improved lightweight convolutional neural network model significantly improve fault diagnosis accuracy. Adaptive sampling frequency adjustment and hierarchical compression strategies effectively solve the data packet loss problem in weak network environments, thus better adapting to weak network conditions in the field.

[0097] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from the spirit and scope of this invention.

Claims

1. A method for monitoring the status of communication construction equipment based on edge computing, characterized in that, include: Edge computing nodes are deployed on communication construction equipment or at the construction site, and the deployment location of the edge computing nodes covers the working area of ​​the communication construction equipment; Data on equipment operation is collected synchronously by multiple sensors deployed on communication construction equipment; Edge computing nodes acquire device operation data and input it into an improved lightweight convolutional neural network model to determine fault feature matching degree and deviation data; Calculate equipment health based on fault characteristic matching degree and deviation data; Edge computing nodes determine device status labels based on device health and fault characteristic matching degree, and then transmit them back to the cloud.

2. The edge computing-based communication construction equipment status monitoring method as described in claim 1, characterized in that, Multiple sensors deployed on communication construction equipment synchronously collect equipment operation data based on an adaptive sampling frequency. The adaptive sampling frequency dynamically switches according to the equipment health status. The initial equipment health status is used in the first monitoring cycle, and the equipment health status output in the previous monitoring cycle is used in subsequent monitoring cycles.

3. The edge computing-based communication construction equipment status monitoring method as described in claim 1, characterized in that, Equipment operation data includes vibration data, noise data, power consumption data, operating temperature data, and start / stop status data.

4. The edge computing-based communication construction equipment status monitoring method as described in claim 3, characterized in that, Edge computing nodes acquire device operation data and input it into an improved lightweight convolutional neural network model to determine fault feature matching degree and deviation data, including: The equipment operation data is input into an improved lightweight convolutional neural network model, which outputs feature vectors. Calculate the cosine similarity between the feature vector and the preset fault feature library, take the maximum value and multiply by 100 to obtain the fault feature matching degree; Each type of sub-data in the equipment operation data is subjected to a deviation formula and normalized to obtain deviation data; the deviation data includes vibration deviation, noise deviation, power consumption deviation and temperature deviation.

5. The edge computing-based communication construction equipment status monitoring method as described in claim 4, characterized in that, Equipment health is calculated based on fault characteristic matching degree and deviation data, including: ； in, For vibration deviation; This is for noise bias; This refers to power consumption deviation; Temperature deviation; For fault feature matching degree; , , , , The weighting coefficient has a value range of 0-1 and satisfies the following conditions: .

6. The edge computing-based communication construction equipment status monitoring method as described in claim 3, characterized in that, Before edge computing nodes input device operation data into the improved lightweight convolutional neural network model, the data undergoes fusion and noise reduction processing, including: The equipment operation data is processed by min-max normalization, converting data of different physical units into dimensionless values ​​in the range of 0-1; Calculate the fused noise reduction value: ； in, , , , The fusion weights corresponding to each data source and satisfying the following conditions: The value range is 0.1-0.4; , , , These are the dimensionless values ​​after normalization of vibration data, noise data, power consumption data, and operating temperature data, respectively. The environmental background noise estimation correction term is obtained by calibrating the signal fluctuation under the no-load state of the equipment. , , , , The standard deviations of the normalized vibration data, noise data, power consumption data, and temperature data are calculated based on a sliding window of the most recent M frames. , , , The corresponding weighting coefficients range from 0 to 0.05 and satisfy the following conditions: .

7. The edge computing-based communication construction equipment status monitoring method as described in claim 1, characterized in that, Edge computing nodes determine device status labels based on device health and fault characteristic matching, including: When the equipment health level is greater than or equal to the preset first health level threshold and the fault feature matching degree is less than the preset first fault feature matching threshold, the equipment status label is normal operation; When the equipment health is greater than or equal to the preset second health threshold and less than the preset first health threshold, and the fault feature matching degree is greater than or equal to the preset first fault feature matching threshold and less than the preset second fault feature matching threshold, the equipment status label is slightly deteriorated. When the equipment health level is greater than or equal to the preset third health level threshold and less than the preset second health level threshold, and the fault feature matching degree is greater than or equal to the preset second fault feature matching degree threshold and less than the preset third fault feature matching degree threshold, the equipment status label is abnormal operation; When the equipment health is less than the preset third health threshold and the fault feature matching degree is greater than or equal to the preset third fault feature matching degree threshold, the equipment status label is fault operation.

8. The edge computing-based communication construction equipment status monitoring method as described in claim 1, characterized in that, Edge computing nodes use dynamic compression strategies based on device health to perform tiered data transmission back to the cloud, including: When the device health level is greater than or equal to the preset first health threshold, the first compression level is executed; When the device health level is greater than or equal to a preset second health threshold but less than a preset first health threshold, the second compression level is executed. When the device health level is greater than or equal to the preset third health threshold but less than the preset second health threshold, the third compression level is executed. When the device health level is lower than the preset third health level threshold, the fourth compression level is executed; The compression ratios of the first, second, third, and fourth compression levels decrease progressively.

9. The edge computing-based communication construction equipment status monitoring method as described in claim 4, characterized in that, The method also includes: The cloud receives the analysis results, abnormal feature data and incremental model parameters returned by the edge computing nodes, performs global data aggregation and model iterative optimization, and sends the updated parameters to the edge computing nodes through incremental parameter differential encoding. The cloud-based model iterative optimization includes an incremental parameter update mechanism: During the local diagnosis process, when the detected feature vector has a cosine similarity to the most similar vector in the preset fault feature library that is lower than a preset threshold, it is judged as a suspected new fault feature. The new feature and its corresponding model parameter update amount are reported to the cloud as incremental parameters. The preset threshold is determined based on the lowest matching degree of the covered fault samples. Incremental parameters are transmitted only after differential encoding of parameter blocks that have changed relative to the previous version of the model; parameter blocks that have not changed are not transmitted. The incremental parameters reported by each edge computing node in the cloud are weighted and aggregated according to the proportion of sample size contributed by each edge computing node. The aggregated model is first evaluated on the cloud validation set. If the diagnostic accuracy is not lower than that of the model before aggregation, the updated parameters are sent to each edge computing node in the form of differential encoding; otherwise, the original parameters are kept.

10. The edge computing-based communication construction equipment status monitoring method as described in claim 1, characterized in that, The improved lightweight convolutional neural network model has a six-layer structure, as follows: The first layer is a one-dimensional deep convolutional layer, which performs channel-by-channel convolution on multi-source time-series signals to extract local time-series features of each channel. The second layer is a pointwise convolutional layer, which achieves feature fusion between channels through 1×1 convolution; The third layer is a max pooling layer, which performs downsampling and preserves significant features; The fourth layer is a depthwise separable convolutional layer, which extracts high-level abstract features; The fifth layer is a global average pooling layer, which compresses the feature map into a feature vector; The sixth layer is a fully connected output layer that outputs the probability distribution of fault types.

Images