A real-time monitoring method and system for the operating state of electromechanical equipment of a highway

Abstract

The application provides a real-time monitoring method and system for the operation state of electromechanical equipment of a highway, and belongs to the technical field of monitoring of electromechanical equipment of a highway. The method captures mechanical vibration energy generated by the driving of a vehicle on a road surface and converts the mechanical vibration energy into a wake-up electrical signal. According to the wake-up electrical signal, the waveform characteristics of a current vibration event are extracted to determine the vibration source type that generates the vibration. According to the vibration source type, it is determined whether it is a preset effective ferry carrier to collect real-time operation state data of electromechanical equipment. The collected real-time operation state data is subjected to edge feature extraction and preliminary fault diagnosis, and is pushed to the effective ferry carrier in a short-range radio frequency mode, so that the real-time operation state data can be transferred to a network coverage area by the movement of the effective ferry carrier and then returned to the cloud for deep analysis. The application realizes efficient monitoring and predictive maintenance of electromechanical equipment in a no-electricity and no-network area by using vehicle vibration self-power generation and physical ferry communication.

Description

Technical Field

[0001] This invention relates to the field of electromechanical equipment monitoring technology for highways, and particularly to a method and system for real-time monitoring of the operating status of electromechanical equipment on highways. Background Technology

[0002] The electromechanical system of highways is a crucial infrastructure for ensuring road safety and efficient operation. It mainly includes various electromechanical equipment such as tunnel ventilation systems, lighting systems, traffic monitoring systems, and information dissemination systems. These electromechanical devices are widely distributed and stretch over long distances along the route. A large number of these devices are located inside tunnels, in mountainous sections, on bridges, and other remote locations. Real-time monitoring and maintenance management of their operational status has always been a technical challenge in the industry.

[0003] Currently, the monitoring of the operational status of electromechanical equipment on highways mainly relies on the following technical solutions: Firstly, there is the monitoring system based on wired networks. This solution involves installing sensors on electromechanical equipment and transmitting the collected data to the nearest monitoring sub-center or road section management center via fiber optic cables or twisted-pair cables. While this technology offers stable data transmission and high real-time performance, its limitations include: high cost of wired network deployment, especially in complex terrain conditions such as tunnels and mountainous areas, leading to significant construction difficulties and maintenance costs; furthermore, the coverage of wired networks is limited, making it difficult to achieve comprehensive coverage for dispersed electromechanical equipment.

[0004] Secondly, there is the monitoring system based on public wireless networks. This solution uses wireless communication modules such as 4G, 5G, or NB-IoT to transmit the status data of electromechanical equipment back to the monitoring center via the operator's network. This solution solves the problem of wired deployment to some extent, but its application is limited by the coverage of public network signals. Along highways, especially inside tunnels and in remote mountainous areas, there are often signal blind spots or weak signal areas, resulting in unstable data transmission and the formation of "data islands." In addition, the continuous online operation of wireless communication modules consumes a lot of power, which places high demands on the power supply system.

[0005] Thirdly, there is the battery-powered monitoring system with timed reporting. To reduce power consumption, some monitoring equipment uses battery power and is set to a timed wake-up mechanism to periodically collect and upload data. The drawbacks of this approach are: timed reporting creates data acquisition blind spots, making it impossible to detect sudden equipment malfunctions in a timely manner; batteries need to be replaced regularly, resulting in a huge maintenance workload for widely distributed and numerous equipment, and the special installation locations of some equipment make battery replacement difficult.

[0006] The aforementioned existing technical solutions generally suffer from the following technical defects in practical applications: First, they are highly dependent on power supply. Most monitoring terminals rely on mains power or batteries. Mains power is expensive to deploy, and battery power requires frequent replacements, making it difficult to achieve long-term, continuous monitoring in areas without electricity.

[0007] Second, there are blind spots in communication coverage. Public network signal coverage is insufficient along highways, especially in tunnels and mountainous sections, resulting in a large number of electromechanical devices being in a "data island" state, unable to achieve real-time transmission of status data.

[0008] Third, data acquisition and processing capabilities are limited. Existing monitoring systems mostly use simple threshold judgment methods, which have a weak ability to identify the status of electromechanical equipment and make it difficult to detect complex faults and early signs of faults; moreover, all raw data needs to be uploaded to the cloud for processing, resulting in large data transmission volume and high energy consumption.

[0009] Fourth, the operation and maintenance mode is passive. Existing systems generally adopt the "repair after failure" or "periodic inspection" mode, which lacks predictive maintenance capabilities and makes it difficult to achieve intelligent management of the entire life cycle of electromechanical equipment.

[0010] Therefore, how to achieve low-power, high-reliability, and intelligent monitoring of the operating status of electromechanical equipment along highways without electricity or network coverage has become a pressing technical problem to be solved in this field. Summary of the Invention

[0011] In view of the above, the main objective of this invention is to provide a method and system for real-time monitoring of the operating status of electromechanical equipment on highways, so as to solve the above-mentioned technical problems.

[0012] This invention proposes a method for real-time monitoring of the operating status of electromechanical equipment on highways. The method is executed by a passive monitoring terminal deployed on the electromechanical equipment of the highway, and the specific steps are as follows: Step 1: Capture the mechanical vibration energy generated by the vehicle moving on the road surface and convert the mechanical vibration energy into a wake-up electrical signal; Step 2: Based on the wake-up electrical signal, trigger the feature extraction operation to extract the waveform features of the current vibration event, and perform pattern recognition on the waveform features to determine the type of vibration source that generates the vibration. Step 3: Determine whether it is a preset effective transfer carrier based on the vibration source type; If so, the data acquisition function will be activated to collect real-time operating status data of the electromechanical equipment; Step 4: Extract edge features and perform preliminary fault diagnosis on the collected real-time operating status data to generate semantic data packets; Step 5: Push the semantic data packets to an effective transfer carrier via short-range radio frequency, so that the semantic data packets can be transferred to the network coverage area by the movement of the effective transfer carrier and then transmitted back to the cloud for in-depth analysis in the cloud.

[0013] This invention also proposes a real-time monitoring system for the operating status of electromechanical equipment on highways. The system is applied to the aforementioned real-time monitoring method for the operating status of electromechanical equipment on highways. The system includes passive monitoring terminals deployed on the electromechanical equipment of the highway; the passive monitoring terminal includes: Piezoelectric energy harvesting module, used for: It captures the mechanical vibration energy generated by the vehicle's movement on the road surface and converts the mechanical vibration energy into a wake-up electrical signal; TinyML processing chip, used for: Based on the wake-up electrical signal, a feature extraction operation is triggered to extract the waveform features of the current vibration event, and pattern recognition is performed on the waveform features to determine the type of vibration source that generates the vibration. Based on the type of vibration source, determine whether it is a preset effective ferry carrier; Edge feature extraction and preliminary fault diagnosis are performed on the collected real-time operating status data to generate semantic data packets; If so, then activate the data acquisition function; Sensor modules are used for: In response to the wake-up data acquisition function signal triggered by the TinyML processing chip, the real-time operating status data of the electromechanical equipment is collected and sent to the TinyML processing chip. Short-range communication module, used for: Semantic data packets are pushed to an effective ferry carrier via short-range radio frequency, so that the semantic data packets can be transferred to the network coverage area by the movement of the effective ferry carrier and then transmitted back to the cloud for in-depth analysis in the cloud.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Achieve passive monitoring and completely solve the power supply problem; By utilizing a piezoelectric energy harvesting module to capture ground vibrations generated by vehicle movement, energy harvesting and information sensing are simultaneously achieved through dual signal processing channels, eliminating the need for external power supply or battery replacement. Dynamic matching of energy budget with task mode enables the terminal to operate stably for extended periods under passive conditions, significantly reducing maintenance costs.

[0015] 2. Overcome communication blind spots and achieve full road segment data coverage; By employing a "vehicle physical shuttle" mechanism, semantic data packets are sent to passing vehicles via short-range communication, using the vehicles as data transport carriers, and then transmitted back to the cloud via networked areas. Combined with negative state monitoring technology, it is possible to infer power loss or communication failures of silent devices, thus solving the "data island" problem in areas without public network signals, such as tunnels and mountainous areas.

[0016] 3. End-edge-cloud collaboration significantly improves diagnostic efficiency and accuracy; The TinyML processing chip at the terminal layer completes vibration source identification, anomaly detection, and preliminary fault diagnosis, generating ultra-low data volume semantic packets (compression ratio exceeding 99%). The edge layer vehicle-mounted intelligent gateway performs spatial correlation analysis and time series analysis to achieve cluster fault identification and degradation trend early warning. The cloud platform aggregates all data for remaining life prediction and maintenance decision-making. The clear division of labor among the three layers reduces communication and storage pressure while improving the real-time performance and accuracy of fault diagnosis.

[0017] 4. Predictive maintenance and model self-evolution reduce operating costs; Based on equipment status time series and swarm intelligence analysis, the remaining service life and failure probability are output, and optimized maintenance work orders are automatically generated. The cloud uses actual maintenance feedback to incrementally or fully train the terminal, edge, and cloud models, and updates the model parameters through OTA (over-the-air) (using vehicle shuttle to the terminal), so as to continuously improve the system's diagnostic capabilities.

[0018] 5. Highly adaptable and can be expanded to various transportation infrastructures; This invention does not rely on specific network infrastructure and can be widely applied to the monitoring of electromechanical equipment such as highway tunnel fans, information boards, and lighting fixtures. It can also be extended to bridges, railways, and other scenarios, and has extremely high promotional value.

[0019] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by means of embodiments of the invention. Attached Figure Description

[0020] Figure 1 This is a flowchart of a method for real-time monitoring of the operating status of electromechanical equipment on highways, as proposed in this invention. Figure 2 This is a diagram illustrating the overall architecture of a real-time monitoring system for the operating status of electromechanical equipment on highways, as proposed in this invention. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0022] These and other aspects of the embodiments of the present invention will become clear from the following description and accompanying drawings. In these descriptions and drawings, some specific embodiments of the present invention are specifically disclosed to illustrate some ways of implementing the principles of the embodiments of the present invention; however, it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0023] Please see Figure 1 This embodiment proposes a method for real-time monitoring of the operating status of electromechanical equipment on highways, the method comprising the following steps: Step 1: Capture the mechanical vibration energy generated by the vehicle moving on the road surface and convert the mechanical vibration energy into a wake-up electrical signal; In a preferred embodiment of the present invention, capturing the mechanical vibration energy generated by the vehicle driving on the road surface and converting the mechanical vibration energy into a wake-up electrical signal specifically includes the following steps: S101. When the vehicle travels to the preset range of the passive monitoring terminal, the road vibration causes the piezoelectric material of the piezoelectric energy acquisition module to deform and output an alternating voltage signal. S102. In the information channel, the alternating voltage signal is amplified and filtered for noise, and then converted into a digital signal sequence. After being buffered and amplified, the alternating voltage signal is filtered by an anti-aliasing filter to remove high-frequency noise above 1kHz, and then converted into a digital signal sequence by an analog-to-digital converter. The digital signal sequence is used as a wake-up electrical signal.

[0024] Step 2: Based on the wake-up electrical signal, trigger the feature extraction operation to extract the waveform features of the current vibration event, and perform pattern recognition on the waveform features to determine the type of vibration source that generates the vibration. In a preferred embodiment of the present invention, a feature extraction operation is triggered based on a wake-up electrical signal to extract the waveform features of the current vibration event, and pattern recognition is performed on the waveform features to determine the type of vibration source. Specifically, the steps include the following: S201, the passive monitoring terminal is equipped with a supercapacitor and has a dual signal processing channel on the piezoelectric energy acquisition module, which includes an energy channel and an information channel; In this step, the energy channel is connected in sequence to a full-bridge rectifier, a filter capacitor, a DC-DC converter (with maximum power point tracking MPPT), and a supercapacitor at the output. Information channel: connected in sequence to buffer amplifier, anti-aliasing filter, and analog-to-digital converter, with the output terminal connected to the analog-to-digital converter input pin of TinyML processing chip.

[0025] S202. When the vehicle travels to the preset range of the passive monitoring terminal, the road vibration causes the piezoelectric material of the piezoelectric energy acquisition module to deform and output an alternating voltage signal. S203. In the energy channel, the alternating voltage signal is rectified and converted to charge the supercapacitor, which then serves as the power supply for the passive monitoring terminal after it is woken up. The alternating voltage signal is rectified by a full-bridge rectifier into a pulsating DC voltage, and the maximum power is extracted by a DC-DC converter in MPPT mode to charge the supercapacitor. The electrical energy stored in the supercapacitor is used for data acquisition, processing, and transmission after the terminal is woken up.

[0026] S204. In the information channel, the alternating voltage signal is amplified and filtered for noise, and then converted into a digital signal sequence. After being buffered and amplified, the alternating voltage signal is filtered by an anti-aliasing filter to remove high-frequency noise above 1kHz, and then converted into a digital signal sequence by an analog-to-digital converter.

[0027] S205. Perform bandpass filtering on the digital signal sequence to obtain the vibration waveform; S206. Extract the time-domain and frequency-domain features of the vibration waveform, and then perform splicing and normalization operations in sequence to obtain the normalized feature vector. S207. Input the normalized feature vector into the pre-trained lightweight neural network model (pre-trained 1D-CNN classifier) ​​for classification and identify the vibration source type, which is one of heavy trucks, small passenger cars, maintenance vehicles or non-vehicle interference. S208. When the vibration source type is identified as a maintenance vehicle or a heavy truck, it is determined to be a valid shuttle carrier; otherwise, it is determined to be an invalid vibration source, and the passive monitoring terminal returns to sleep mode.

[0028] This step achieves efficient acquisition of vibration energy and real-time analysis of vibration waveforms using the same piezoelectric signal source, enabling vehicle type identification without additional sensors. Subsequent acquisition and communication processes are only activated when a valid carrier passes through, reducing the invalid wake-up rate by over 90% compared to traditional timed or simple threshold wake-up methods, significantly improving the energy utilization efficiency of the passive monitoring terminal.

[0029] To further reduce communication latency, this embodiment also provides a predictive activation function. By analyzing the phase difference of the vibration waveform, the vehicle's direction of travel and speed are calculated, and the time window for the vehicle to arrive at the monitoring point is predicted. Before the predicted time window arrives, the sensor and communication module are woken up in advance to achieve seamless communication with the vehicle.

[0030] Specifically, once the TinyML processing chip identifies a valid carrier, it does not immediately wake up the sensors. Instead, it estimates the vehicle's speed and direction by analyzing the phase difference of the vibration waveform. The method involves integrating two piezoelectric units spaced 0.5m apart. The vehicle triggers the two units sequentially, and the vehicle speed is calculated based on the time difference and the distance between them. Based on the vehicle speed and its current distance (estimated through vibration amplitude), the time it takes for the vehicle to arrive at the communication area is predicted. The TinyML processing chip sets a timer to wake up the sensors and communication module 10ms before the arrival time in the communication area, ensuring data acquisition and transmission are synchronized with the vehicle's arrival and reducing standby power consumption.

[0031] The predictive activation steps are as follows: Two independent piezoelectric energy acquisition units, denoted as piezoelectric unit P1 and piezoelectric unit P2, are installed at the bottom of the passive monitoring terminal casing along the vehicle's travel direction (i.e., the highway driving direction). The center-to-center distance between the two piezoelectric units is a fixed value D, which is 0.5 meters (this distance is based on the fact that at common highway speeds of 80-120 km / h, the time difference between the vibration signal reaching the two units is within the range of 15-22.5 milliseconds, which is convenient for the analog-to-digital converter to sample and distinguish).

[0032] When a vehicle travels to the preset range of the passive monitoring terminal, the impact of the vehicle's tires on the road surface generates vibration waves, which propagate along the road surface in the form of longitudinal waves and surface waves. Since piezoelectric units P1 and P2 are arranged along the direction of travel, the vibration waves first reach the piezoelectric unit P1, which is closer to the oncoming vehicle, and then reach the other unit P2. After receiving the wake-up electrical signal, the TinyML processing chip does not immediately enter the data acquisition mode, but first starts the predictive activation process. The chip simultaneously acquires the digital signal sequences output by the information channels of piezoelectric units P1 and P2 at a sampling rate of 2kHz, denoted as S1[n] and S2[n], respectively, with a sampling window length of 200ms (covering the time it takes for the vehicle to completely pass through the two units).

[0033] Bandpass filtering is performed on S1[n] and S2[n] respectively to remove low-frequency road surface fluctuations and high-frequency environmental noise, resulting in filtered sequences, denoted as S1. filt [n] and S2 filt [n]; For S1 filt [n] and S2 filt [n] Perform Hilbert transform to extract the envelope, obtaining envelope S1 and envelope S2; set the envelope threshold (30% of the maximum envelope value in this embodiment), and detect the sampling point index of envelope S1 and envelope S2 that first exceed the threshold, converting them into arrival time t1 and arrival time t2.

[0034] Based on arrival time t1 and arrival time t2, calculate the time difference ΔT = t2 - t1. Vehicle speed V = D / |ΔT| (unit: m / s). Direction determination: If t1 < t2, the driving direction is from piezoelectric unit P1 to piezoelectric unit P2; otherwise, it is the opposite. Verify that the vehicle speed V is within the range of 20-150 km / h. If it exceeds this range, abandon predictive activation and revert to the normal wake-up mode.

[0035] The optimal communication position relative to the geometric centers of the two piezoelectric units (along the direction of travel) was calibrated during terminal installation. Using the last piezoelectric unit triggered as the reference point, the remaining distance L was calculated. rem (Constant, stored in the terminal). Predicted arrival time T arrival = L rem / V.

[0036] The TinyML processing chip sets a timer to trigger an interrupt at the predicted arrival time minus the warm-up time (5ms). During the timer countdown, the chip enters a light sleep state. After the interrupt is triggered, the TinyML processing chip wakes up the sensors (accelerometer, current transformer, temperature sensor) and communication module, and synchronously collects data and sends semantic data packets when the vehicle reaches the optimal communication position.

[0037] Step 3: Determine whether it is a preset effective transfer carrier based on the vibration source type; If so, the data acquisition function will be activated to collect real-time operating status data of the electromechanical equipment; Step 4: Extract edge features and perform preliminary fault diagnosis on the collected real-time operating status data to generate semantic data packets; In a preferred embodiment of the present invention, edge feature extraction and preliminary fault diagnosis are performed on the collected real-time operating status data to generate semantic data packets, specifically including the following steps: S401. The raw data includes vibration signals, current signals and temperature signals. The vibration signals, current signals and temperature signals are preprocessed respectively to obtain the preprocessed signals. In this step, after the TinyML processing chip identifies the vibration source and determines it to be a valid transfer carrier, it wakes up the sensor array and collects the following raw data: vibration signals are collected using a triaxial accelerometer, current signals are collected using a current transformer, and temperature signals are collected using a temperature sensor; the DC component of the vibration signal is removed and a Hanning window is applied; the DC component of the current signal is removed; and the temperature signal is used directly.

[0038] S402. Using the feature extraction model built into the TinyML processing chip, fault feature vectors are extracted from the preprocessed signal. The fault feature vectors include vibration features, current features, and temperature features. In this step, the TinyML processing chip's built-in feature extraction model contains three parallel sub-modules, which process vibration, current, and temperature data respectively. Vibration feature extraction: Perform a fast Fourier transform on the main vibration direction of the preprocessed vibration signal to obtain the vibration spectrum; find the preset fault feature frequency (such as bearing outer ring fault frequency, gear meshing frequency) in the vibration spectrum, extract the preset fault feature frequency amplitude, and calculate the root mean square and kurtosis of the entire frequency band. These data finally form the vibration feature vector. Current feature extraction: Perform a fast Fourier transform on the preprocessed current signal to obtain the current spectrum; based on the current spectrum, calculate the total harmonic distortion rate and the fundamental current amplitude to form a current feature vector; Temperature feature extraction: Calculate the temperature gradient to obtain the temperature feature vector; S403. Input the fault feature vector into the lightweight fault diagnosis model to obtain the initial diagnosis conclusion; In this step, the vibration feature vector, current feature vector, and temperature feature vector are concatenated and fed into a pre-trained isolated forest anomaly detection model to output anomaly scores for the electromechanical equipment. The higher the score, the more significantly the electromechanical equipment deviates from the normal mode. If the anomaly score of the electromechanical equipment exceeds a preset threshold, the concatenated vector is further input into a pre-trained support vector machine to output the fault type and corresponding fault confidence. The fault type can include bearing wear, power phase loss, overheating, or normal operation. Then, the anomaly score, fault type, and corresponding fault confidence of the electromechanical equipment are used as the initial diagnostic conclusion.

[0039] S404 Finally, the extracted fault feature vector, along with the initial diagnosis conclusion, electromechanical equipment ID, and timestamp, are structurally encapsulated to generate a semantic data packet.

[0040] Step 5: Push the semantic data packets to the effective transfer carrier via short-range radio frequency so that the semantic data packets can be transferred to the network coverage area by the movement of the effective transfer carrier and then transmitted back to the cloud for in-depth analysis in the cloud. In a preferred embodiment, the effective shuttle carrier is equipped with an on-board intelligent gateway. Semantic data packets are received and cached by the on-board intelligent gateway. After pushing the semantic data packets to the effective shuttle carrier via short-range radio frequency, the on-board intelligent gateway also performs fusion analysis and secondary diagnostic steps, as detailed below. S501A performs spatial correlation analysis on semantic data packets sent by multiple passive monitoring terminals on the same road segment, calculates the correlation coefficient of fault feature vectors of each electromechanical device, and determines that there is a power supply abnormality or environmental disturbance when the correlation coefficient exceeds the preset threshold, thus obtaining the spatial analysis results. In this step, the in-vehicle intelligent gateway continuously receives semantic data packets via the UWB module. Each semantic data packet contains: electromechanical equipment ID, timestamp, fault feature vector, and initial diagnosis conclusion. The edge AI computing unit categorizes the data packets by electromechanical equipment ID and stores them in a local circular buffer, while simultaneously recording the vehicle's location and reception time. For the same electromechanical equipment ID, the buffer retains the data from the most recent 10 visits for trend analysis.

[0041] Specifically, based on the equipment installation location mapping table, all data packets in the current buffer are grouped according to their respective road segments. Suppose there are M pieces of equipment in a certain road segment, and the data packet corresponding to each piece of equipment contains the initial diagnosis conclusion and feature vector.

[0042] The number of electromechanical devices with an "abnormal score > 0.6" in the initial diagnosis of electromechanical equipment in this section is counted. If more than 70% of the electromechanical devices are judged to be abnormal, the cluster abnormality flag is triggered.

[0043] For the electromechanical equipment that triggers the cluster anomaly flag, the vibration RMS value of the vibration feature vector and the current THD value of the current feature vector are extracted. The Pearson correlation coefficients of the vibration RMS values ​​and the current THD values ​​between different electromechanical equipment are calculated and averaged. If the average value is greater than 0.8 (preset), it is determined that the fault types of these electromechanical equipment are different (i.e., not the same fault type), and it is determined to be "systematic disturbance" rather than multiple independent faults.

[0044] If the main manifestation is a synchronous increase in current THD and synchronous fluctuation in vibration RMS, it is inferred to be "supply voltage fluctuation"; If the main manifestation is a synchronous increase in vibration RMS while the current THD is normal, it is inferred to be "environmental wind pressure resonance"; If the initial diagnosis shows that most of the mechanical and electrical equipment fault types are "normal" but the abnormal score is high, it is inferred to be "common sensor interference".

[0045] S502A performs time series analysis on the semantic data packets sent by the same passive monitoring terminal when it passes through multiple times, calculates the rate of change of the fault feature vector, and determines that there is a slow degradation fault when the rate of change continues to rise, thus obtaining the time analysis results. The main objective of this step is to utilize the accumulated data from multiple passes of the inspection vehicle over the same electromechanical equipment, combined with historical changes in the initial diagnostic conclusions, to identify slow performance degradation and achieve early warning.

[0046] Specifically, for the electromechanical equipment ID, the most recent K (K≥3) historical data are read from the buffer, including the fault feature vector and initial diagnosis conclusion for each time, and sorted in ascending order by timestamp.

[0047] Observe the changing trend of abnormal scores. If the slope of the linear fit of the abnormal scores is greater than zero and increases for three consecutive times, then an "abnormal development trend" is determined to exist.

[0048] Calculate the rate of change for key feature dimensions (such as bearing amplitude and temperature). If the rate of change of a feature exceeds 30% and the slope of the linear fit of the outlier score is greater than zero, then further confirmation of degradation is made.

[0049] Based on the current initial diagnosis of the fault type and historical changes, for example: If the current fault type is "normal", but the abnormal score continues to rise and the bearing amplitude increases significantly, it is determined to be "early bearing wear" (the fault type is updated to "bearing wear warning"). If the current fault type is "normal" but the abnormal score continues to rise, it is judged as "overheating trend"; If the current fault type is already a certain fault and the fault confidence level is gradually increasing, it is determined as "fault aggravation".

[0050] Finally, based on these analysis results, time series analysis results are generated, including: degradation trend slope, updated fault type (if necessary), and warning level.

[0051] S503A: Based on the spatial analysis results and the temporal analysis results, the initial diagnosis conclusion is revised to generate a more accurate secondary diagnosis conclusion; The main purpose of this step is to revise the original initial diagnosis based on the results of spatial correlation analysis and time series analysis, and generate a more accurate secondary diagnosis.

[0052] Specifically, if the spatial correlation analysis determines that it is a systematic disturbance, the fault type in the initial diagnosis of the electromechanical equipment will be overridden as "systemic disturbance", the anomaly score will be reset to 0.5 (moderate anomaly), and the fault confidence will be reduced by 0.2. If the time series analysis determines a degradation trend and the original initial diagnosis was "normal", then the fault type will be updated to "early warning of a certain condition" (such as "bearing wear early warning"), the abnormal score will be increased to 0.6, and the fault confidence will be the fault confidence of the trend fit. If neither is abnormal, the original initial diagnosis is retained, and a secondary diagnosis is obtained.

[0053] S504A. Prioritize abnormal data based on more accurate secondary diagnostic conclusions to obtain the classification results. This includes the following steps: Mark data that has already failed or is about to fail as urgent priority; Data showing performance degradation or signs of failure will be marked as warning priority; Data with minor anomalies requiring observation should be marked as high priority data. Mark data without anomalies as normal priority; The emergency priority, early warning priority, attention priority, and normal priority are used as the policy basis for the vehicle-mounted intelligent gateway to upload enhanced semantic data packets in that order.

[0054] In this step, the abnormal data is prioritized based on more accurate secondary diagnostic conclusions. An example of this is as follows:

[0055] S505A Finally, the hierarchical results, spatial analysis results, temporal analysis results, and more accurate secondary diagnostic conclusions are added to the semantic data packet to generate an enhanced semantic data packet.

[0056] After completing secondary diagnostics and priority classification, the vehicle intelligent gateway traverses all enhanced semantic data packets in the local cache; it assigns enhanced semantic data packets of normal priority to the queue to be compressed, and assigns enhanced semantic data packets of other priorities to the direct transmission queue without compression.

[0057] Statistical summaries are generated from enhanced semantic data packets representing the normal status of multiple passive monitoring terminals within the same road segment and time window. The statistical summaries include road segment identification, total number of electromechanical devices, number of normal electromechanical devices, average temperature, and average vibration amplitude. The statistical summary is combined with the enhanced semantic data packets of the abnormal electromechanical equipment to form the final upload queue. Data is uploaded in priority order: first, emergency / early warning / concern data packets are uploaded, and finally, the statistical summary (replacing the original detailed normal data packets) is uploaded, thus reducing the amount of data uploaded. When a network connection is established, the data in the final upload queue is uploaded to the cloud sequentially.

[0058] As a preferred embodiment of the present invention, the specific steps of cloud-based deep analysis are as follows: S501B: The enhanced semantic data packets from different vehicles and time points are spatiotemporally aligned according to the electromechanical equipment ID and time axis to construct the state time series of each electromechanical equipment. In this step, the enhanced semantic data packets are first cleaned and validated. Then, for each electromechanical device, the data reported by different vehicles at different times is sorted by timestamp to form a continuous state time series.

[0059] Because the same electromechanical equipment may be passed and collected by multiple vehicles equipped with onboard intelligent gateways within the same time period, the cloud platform may receive multiple reported data from different vehicles for the same electromechanical equipment at the same time. To avoid data redundancy and conflicting conclusions, the cloud platform executes the following time window aggregation and conflict resolution rules: Time window segmentation: Divide the time axis into continuous statistical windows according to a preset time granularity (e.g., 15 minutes). For each statistical window, group all reported data for the same electromechanical equipment that fall within that window.

[0060] Vehicle deduplication: If the same vehicle has multiple reporting records for the same electromechanical equipment within the same statistics window, only the most recent data packet reported by the vehicle will be retained (the latest one will be selected based on the timestamp).

[0061] Multi-source data conflict resolution: When multiple valid reported data from different vehicles exist within the same statistical window, the status data representing that window is determined according to the following priority rules: data with higher diagnostic conclusion priority is adopted first (emergency > warning > attention > normal); if the priorities are the same, data with higher anomaly score is adopted; if the anomaly scores are the same, data with a later timestamp is adopted.

[0062] State sequence construction: After the above deduplication and fusion processing, each statistical window retains at most one valid data point representing the state of the electromechanical equipment during that time period. The representative data from each statistical window are arranged in chronological order to form a continuous and non-redundant state time series.

[0063] For statistical windows with no reported data (such as periods when no vehicles pass by), no interpolation is performed to maintain the original sparse sampling characteristics. At the same time, vehicle location information is used to label each data point with "perceived vehicle type" (maintenance vehicle / private vehicle) and "perceived angle" (vehicle driving direction) for subsequent swarm intelligence analysis.

[0064] S502B: Calculate the theoretical frequency at which each electromechanical device is sensed by the vehicle, and obtain the number of records reported by deduplicated vehicles as the actual sensing frequency. S503B: When the actual sensing frequency is lower than the theoretical frequency and the traffic flow on the road section is normal, negative state monitoring is triggered to infer that the electromechanical equipment has a power failure or communication failure fault, and a power failure / failure maintenance work order is directly generated. In this step, the cloud platform retrieves traffic flow data for each road segment from roadside traffic flow detectors (geomagnetic coils, radar) every 5 minutes. For road segments without detectors, the historical average value for the same period (in weekly mode) is used to fill in the gaps to obtain the basic traffic flow.

[0065] Each electromechanical device has a known effective communication distance (the effective communication distance varies depending on the location; for example, a nominal distance of 100 meters, but 50 meters after considering signal attenuation in a tunnel). This determines the length of the communication window for a single vehicle passing through the electromechanical device, which is the length of the road segment covered by one effective communication distance before and after the electromechanical device.

[0066] Based on the time required for a single vehicle to pass through the communication window, and combined with traffic flow, it is possible to estimate how many vehicles are expected to successfully communicate within a fixed time period, thus obtaining the theoretical frequency.

[0067] The cloud platform counts the number of vehicle reports received from the electromechanical equipment within the same time period, after deduplication, and uses this as the actual sensing frequency. Then, based on the theoretical frequency and the actual sensing frequency, it calculates the missing rate M = (theoretical frequency - actual sensing frequency) / theoretical frequency, and uses the magnitude of the missing rate as the trigger condition for negative state monitoring, for example: Missing rate M > 0.7 (missing rate exceeds 70%); Actual traffic volume ≥ 0.5 × base traffic volume (excluding natural losses due to low traffic volume); The above conditions must be met for three consecutive time periods (45 minutes); Furthermore, during the aforementioned period, at least one maintenance vehicle or heavy truck that was confirmed to be equipped with an on-board intelligent gateway passed through the section of road where the electromechanical equipment was located but did not receive data. If no vehicle equipped with a gateway passes through during this period, the observation window will be extended and negative status monitoring alarms will not be triggered temporarily to eliminate statistical bias caused by ordinary vehicles without receiving equipment.

[0068] After triggering negative status monitoring, query all vehicle IDs that passed through the road segment where the current electromechanical equipment is located during that time period, and check whether these vehicles uploaded data from other electromechanical equipment. If at least 3 vehicles reported data from other electromechanical equipment but not the current electromechanical equipment, it is determined to be an electromechanical equipment fault; if all vehicles did not upload any data, it may be a problem with the vehicle gateway or network, and no electromechanical equipment fault alarm will be triggered.

[0069] Query the last successfully reported data packet of the current electromechanical equipment before triggering negative state monitoring: If the last report includes a "low supercapacitor voltage" warning (energy storage capacity <10% in the feature vector), then the fault type is inferred to be "power outage". If the last report contains the "communication module self-test abnormal" flag, it is inferred that the "communication module has failed"; If the electromechanical equipment has never had any reporting records (new installation), it is inferred to be "initial deployment anomaly"; If the last report was normal and more than 7 days ago, it is inferred to be an "unknown fault (on-site investigation is recommended)".

[0070] The cross-validation consistency rate is defined as the ratio of the number of vehicles that passed through the section where the electromechanical equipment is located and reported data from other electromechanical equipment but lacked data from their own equipment within the time window that triggers negative state monitoring, to the total number of candidate vehicles that passed through that section. The higher the ratio, the greater the possibility that the problem is not a general issue with the vehicle gateway or the network upload link, but rather a specific fault in the local electromechanical equipment.

[0071] Therefore, the cross-validation consistency rate, the sensing missing rate, and the number of consecutive missing periods can quantify the reliability of negative state monitoring inferences, that is, the degree of confidence that electromechanical equipment has power loss or communication failure faults. The higher the value, the more reliable the inference. If the confidence level is greater than a preset threshold, an emergency maintenance work order is generated. This work order has the highest priority and is immediately pushed to the maintenance personnel. This branch (all results of this step) does not enter the subsequent predictive maintenance process.

[0072] S504B: For electromechanical equipment that has not triggered negative condition monitoring, input the corresponding state time series data into the predictive maintenance model, calculate the remaining service life and failure probability of the equipment, and generate a predictive maintenance work order.

[0073] In this step, the predictive maintenance model includes a remaining useful life prediction model and a failure probability prediction model. The remaining useful life prediction model uses a pre-trained temporal convolutional network, while the failure probability prediction model uses an XGBoost classifier.

[0074] Before inputting the state time series data into the predictive maintenance model, the state time series of the most recent 30 days of electromechanical equipment that has not triggered negative state monitoring is first extracted, and the series is aligned and the difference is filled to obtain the aligned state time series.

[0075] The fault feature vector is extracted from the state time series. The daily median, the daily mean of the anomaly score, and the number of vehicle perceptions after deduplication are used to form a daily feature vector. The daily feature vectors are stacked in chronological order to form a time series tensor. The time series tensors are normalized and fed into a pre-trained temporal convolutional network to predict the remaining lifespan of each electromechanical device. By fitting the historical linear regression slope of the daily feature vector (using 7 days as the standard in this example), the trend characteristics are obtained; The fluctuation characteristics are obtained by calculating the coefficient of variation (standard deviation / mean) of the daily feature vectors over 7 days. The median and standard deviation of the corresponding characteristics of all electromechanical equipment in the same road segment on the same day (to calculate the relative deviation) are used to obtain the group characteristics; Calculate the average traffic volume, average temperature, and number of rainy days over the past 7 days to obtain environmental characteristics; The historical health index is obtained by calculating the percentage of days with an abnormal score exceeding 0.6 in the past 30 days. The columns of trend features, fluctuation features, group features, environmental features, historical health indicators, and the feature vector of the day are independently normalized and then horizontally concatenated to form a complete feature matrix, thus obtaining the feature matrix of the day. Input the feature matrix of the day into the XGBoost classifier to obtain the probability of failure of each electromechanical device in the future.

[0076] Finally, the remaining service life of the electromechanical equipment and the probability of its failure are combined to define a health score for the electromechanical equipment.

[0077] Based on the above prediction results, the following are examples of predictive maintenance work order triggering conditions: Remaining service life < 72 hours; The probability of failure in the next 24 hours is > 0.7; The probability of failure in the next 7 days is > 0.85; The health score has declined for 7 consecutive days and the current score is < 0.4.

[0078] Work order content generation: subject to the following constraints: Location of mechanical and electrical equipment (latitude and longitude, mileage marker); Predict the type of failure (inferred from characteristic trends); Recommended repair time window (prioritize periods with low traffic volume); Required spare parts list (mapped according to fault type); Operating instructions (standardized maintenance procedures); Personnel Assignment: Shortest path optimization based on maintenance personnel's current location, skill tags, and shift schedule.

[0079] In a preferred embodiment of the present invention, the method further includes a model self-evolution step: S601. The cloud platform optimizes at least one machine learning model used in the implementation of the method based on actual fault repair records to obtain optimized model data. S602. The optimized model data is sent to the effective shuttle carrier via OTA over-the-air upgrade. Then, the mobile attributes of the effective shuttle carrier are used to transmit model parameters to each passive monitoring terminal via short-range radio frequency. After receiving the optimized model data, the S603 passive monitoring terminal updates or replaces the corresponding local model based on the optimized model data, thereby continuously improving the diagnostic accuracy.

[0080] In this step, the cloud platform utilizes the massive amounts of aggregated data and actual fault repair records to centrally train and optimize the isomorphic version of the model deployed at the terminal layer (or a teacher model that can be converted through knowledge distillation). The optimized model parameters are then sent to the edge layer via OTA (Over-The-Air) updates, and transmitted from the edge to the terminal. The terminal only loads and updates the model parameters (without performing local training), thereby continuously improving diagnostic accuracy. Since the edge layer does not involve model training, the cloud platform uses the aggregated massive amounts of data to determine the parameters and sends them to the edge layer.

[0081] Terminal layer model training data: 1D-CNN classifier (vehicle model recognition): Manually labeled feature vectors (four categories: heavy trucks, small passenger cars, maintenance vehicles, and non-vehicle interference) are extracted from historical vibration waveforms stored in the cloud. The labeling sources include roadside camera alignment labels and patrol vehicle self-reports.

[0082] Isolated Forest Anomaly Detection Model (Anomaly Detection): Use unlabeled feature vectors from the normal operation of electromechanical equipment as normal samples.

[0083] Support Vector Machine (Fault Classification): Feature vectors of fault samples confirmed using historical maintenance records.

[0084] Edge layer parameter determination (not model training): The cloud-based system statistically analyzes the distribution of correlation coefficients, rates of change, etc., determines spatial correlation thresholds (e.g., 0.8), degradation warning thresholds (e.g., 30%), etc., and sends the data to the vehicle-mounted intelligent gateway.

[0085] Cloud-based model training data: Temporal Convolutional Network (Remaining Lifetime Prediction): Extracts the state time series of faulty electromechanical equipment 30 days before the failure, and uses healthy electromechanical equipment as negative samples.

[0086] XGBoost classifier (failure probability prediction): uses the feature vector at the current time step and labels it with whether a failure will occur in the future.

[0087] The training strategy is as follows: Supervised models (1D-CNN classifier, support vector machine, temporal convolutional network, XGBoost classifier): When there is little new data, incremental fine-tuning is used (small learning rate, only updating some layers). When a large amount of new data has accumulated, full retraining is used (merging historical data and cross-validating to select the best).

[0088] Unsupervised model (Isolated Forest Anomaly Detection Model): Retrained weekly using all normal data from the most recent 30 days.

[0089] Edge layer parameters: will be distributed after monthly statistical updates.

[0090] It should be noted that the term "model data" in this embodiment should be interpreted broadly. For parametric models trained using iterative optimization algorithms such as gradient descent (e.g., 1D-CNN classifiers, support vector machines, temporal convolutional networks, XGBoost classifiers), model data includes differentiable parameters such as the network's weight matrix and bias vectors. For non-parametric models built using non-iterative methods (e.g., isolated forest anomaly detection models), model data includes all configuration data constituting a complete description of the model, such as the number of isolated trees, the structural definition of each tree, split feature indexes, and split thresholds. Regardless of the type of model data, after completing the optimization training, the cloud encapsulates the model data into transmittable data blocks and distributes them via OTA (Over-The-Air) updates. Upon receiving the data blocks, the terminal performs loading and update operations to replace or update the local model, thereby continuously improving diagnostic accuracy.

[0091] Please see Figure 2 This embodiment also proposes a real-time monitoring system for the operating status of electromechanical equipment on highways. The system includes a passive monitoring terminal deployed on the electromechanical equipment of the highway. Specifically, the passive monitoring terminal includes a piezoelectric energy acquisition module, a TinyML processing chip, a sensor module, and a short-range communication module.

[0092] The piezoelectric energy acquisition module is used to capture the mechanical vibration energy generated by the vehicle driving on the road surface and convert it into a wake-up electrical signal. Specifically, when the vehicle travels to the preset range of the passive monitoring terminal, the piezoelectric energy acquisition module captures the ground vibration generated by the vehicle and generates a wake-up electrical signal.

[0093] The TinyML processing chip is used to: trigger feature extraction operation based on the wake-up electrical signal to extract the waveform features of the current vibration event, and perform pattern recognition on the waveform features to determine the type of vibration source; determine whether it is a preset valid transfer carrier based on the vibration source type; if so, wake up the data acquisition function; and perform edge feature extraction and preliminary fault diagnosis on the acquired real-time operating status data to generate semantic data packets.

[0094] The sensor module is used to respond to the wake-up data acquisition function signal triggered by the TinyML processing chip, collect real-time operating status data of the electromechanical equipment, and send the real-time operating status data to the TinyML processing chip. The sensor module includes a triaxial accelerometer (for collecting vibration signals), a current transformer (for collecting current signals), and a temperature sensor (for collecting temperature signals).

[0095] A short-range communication module is used to wirelessly push semantic data packets to an effective ferry carrier via short-range radio frequency. This allows the semantic data packets to be transferred to the network coverage area and then transmitted back to the cloud for in-depth cloud analysis, aided by the movement of the ferry carrier. The short-range communication module is preferably an ultra-wideband (UWB) communication module or a Bluetooth Low Energy (BLE) module, with a communication distance configured to be 50 to 150 meters to match the effective communication window when vehicles pass by.

[0096] It should be understood that the passive monitoring terminal described in this embodiment constitutes a real-time monitoring system for the operating status of highway electromechanical equipment. This system can be independently manufactured, sold, and used. When the terminal is deployed on highway electromechanical equipment, it works collaboratively with the effective shuttle carrier (vehicle) in the external environment and its onboard intelligent gateway and cloud platform to form a complete end-edge-cloud monitoring system. The specific composition and functions of the onboard intelligent gateway and cloud platform will be further explained in Embodiment 3 below.

[0097] Based on the aforementioned passive monitoring terminal, this embodiment further illustrates the complete system architecture and its collaborative working mechanism, which includes a terminal perception layer, a vehicle edge layer, and a cloud platform layer.

[0098] Please refer to it again. Figure 2 This embodiment provides a real-time monitoring system for the operating status of highway electromechanical equipment, comprising a three-layer architecture. The system includes: The terminal sensing layer refers to the aforementioned passive monitoring terminals deployed on electromechanical equipment. The vehicle edge layer includes an onboard intelligent gateway mounted on a driving vehicle (especially a maintenance vehicle and a heavy truck). The onboard intelligent gateway includes a short-range communication receiving module, an edge AI computing unit, and a cellular communication module. The cloud platform layer includes a data aggregation module and an operation and maintenance management module.

[0099] The following details the collaborative working mechanism of this three-tier architecture.

[0100] I. Terminal Perception Layer; The terminal sensing layer includes several passive monitoring terminals deployed on highway electromechanical equipment. Each passive monitoring terminal includes a piezoelectric energy acquisition module, a TinyML processing chip, a sensor module, and a short-range communication module.

[0101] The piezoelectric energy acquisition module is used to capture the mechanical vibration energy generated by the vehicle driving on the road surface and convert it into a wake-up electrical signal. Specifically, when the vehicle travels to the preset range of the passive monitoring terminal, the piezoelectric energy acquisition module captures the ground vibration generated by the vehicle and generates a wake-up electrical signal.

[0102] The TinyML processing chip is used to: trigger feature extraction operation based on the wake-up electrical signal to extract the waveform features of the current vibration event, and perform pattern recognition on the waveform features to determine the type of vibration source; determine whether it is a preset valid transfer carrier based on the vibration source type; if so, wake up the data acquisition function; and perform edge feature extraction and preliminary fault diagnosis on the acquired real-time operating status data to generate semantic data packets.

[0103] The sensor module is used to respond to the wake-up data acquisition function signal triggered by the TinyML processing chip, collect real-time operating status data of the electromechanical equipment, and send the real-time operating status data to the TinyML processing chip. The sensor module includes a triaxial accelerometer (for collecting vibration signals), a current transformer (for collecting current signals), and a temperature sensor (for collecting temperature signals).

[0104] A short-range communication module is used to wirelessly push semantic data packets to an effective ferry carrier via short-range radio frequency. This allows the semantic data packets to be transferred to the network coverage area and then transmitted back to the cloud for in-depth cloud analysis, aided by the movement of the ferry carrier. The short-range communication module is preferably an ultra-wideband (UWB) communication module or a Bluetooth Low Energy (BLE) module, with a communication distance configured to be 50 to 150 meters to match the effective communication window when vehicles pass by.

[0105] II. Vehicle edge layer; The vehicle edge layer includes onboard intelligent gateways mounted on vehicles (especially maintenance vehicles and heavy-duty trucks). These onboard intelligent gateways include short-range communication receiving modules, edge AI computing units, and cellular communication modules.

[0106] The short-range communication receiving module is used to receive and buffer semantic data packets sent by multiple passive monitoring terminals along the route.

[0107] The edge AI computing unit is used to perform fusion analysis and secondary diagnosis on cached semantic data packets to generate enhanced semantic data packets. Specifically, the edge AI computing unit performs the following operations: spatial correlation analysis on semantic data packets sent by multiple passive monitoring terminals on the same road segment, calculates the correlation coefficient of fault feature vectors of each electromechanical device, and determines that there is a power supply abnormality or environmental disturbance when the correlation coefficient exceeds a preset threshold; time series analysis on semantic data packets sent by the same passive monitoring terminal when it passes by multiple times, calculates the rate of change of fault feature vectors, and determines that there is a slow degradation fault when the rate of change continues to rise; the initial diagnosis conclusion is corrected based on the spatial and temporal analysis results to generate a secondary diagnosis conclusion; the data is prioritized (urgent, warning, attention, normal) based on the secondary diagnosis conclusion, and a statistical summary is generated for the normal priority data to replace the detailed data packet.

[0108] The cellular communication module is used to upload enhanced semantic data packets (and statistical summaries) to the cloud platform when a network connection is detected.

[0109] III. Cloud Platform Layer; The cloud platform includes a data aggregation module and an operation and maintenance management module.

[0110] The data aggregation module performs spatiotemporal alignment and swarm intelligence analysis on the received enhanced semantic data packets to generate health assessment results for electromechanical equipment and predictive maintenance decisions. Specifically, this includes: spatiotemporally aligning data from different vehicles and time points according to the electromechanical equipment ID and timeline to construct a state time series; calculating the theoretical frequency at which each electromechanical device is perceived by the vehicle; when the actual perception frequency is lower than the theoretical frequency, inferring that the electromechanical equipment has a power outage or communication failure, and directly generating an emergency maintenance work order; inputting the state time series data into the predictive maintenance model to calculate the remaining service life and failure probability of the electromechanical equipment, and generating predictive maintenance work orders.

[0111] The operation and maintenance management module is used to push maintenance work orders to the maintenance terminal.

[0112] Furthermore, the cloud platform is configured to optimize and train the machine learning models involved in the terminal and cloud layers based on actual fault repair records. The optimized model parameters are then sent to the vehicle's intelligent gateway via OTA (Over-The-Air). The intelligent gateway forwards these parameters to the passive monitoring terminal via a short-range communication module when the terminal passes by. After receiving the optimized model parameters, the passive monitoring terminal updates its local model using the TinyML processing chip, thereby continuously improving the diagnostic accuracy of the entire system.

[0113] IV. Detailed Explanation of the End-Edge-Cloud Collaborative Working Mechanism

[0114] To enable those skilled in the art to more clearly understand the complete workflow of the present invention, the collaborative working mechanism of the end-edge-cloud three-layer architecture is described in detail below.

[0115] The system's workflow begins at the terminal sensing layer. The passive monitoring terminal is normally in a deep sleep state, retaining only the information channel of the piezoelectric energy acquisition module to remain sensitive to external vibrations. When a vehicle passes through the section of road where the electromechanical equipment is located, the contact vibration between the tires and the road surface is converted into an electrical signal by the piezoelectric material. This signal charges the supercapacitor through the energy channel and outputs a digital signal sequence as a wake-up signal through the information channel. After being woken up, the TinyML processing chip first performs vibration source type identification, using a built-in lightweight neural network model (1D-CNN) to classify the vibration waveform and determine whether the approaching vehicle is a heavy truck, maintenance vehicle, small passenger car, or environmental interference. If it is determined to be a valid shuttle vehicle (maintenance vehicle or heavy truck), the sensor module is immediately woken up to collect three types of operating status data (vibration, current, temperature) of the electromechanical equipment, and the edge feature extraction and preliminary fault diagnosis process is initiated. The TinyML processing chip runs an isolated forest anomaly detection model and a support vector machine classifier to perform anomaly detection and fault classification on the collected data, generating a semantic data package containing fault feature vectors, preliminary diagnosis conclusions, equipment IDs, and timestamps. Subsequently, the short-range communication module (UWB / BLE) wirelessly pushes semantic data packets to passing vehicles within the predicted optimal communication window.

[0116] After the data enters the vehicle's edge layer, the short-range communication receiving module of the onboard intelligent gateway continuously receives and caches semantic data packets sent by various electromechanical equipment terminals along the route. The edge AI computing unit performs fusion analysis on the cached semantic data packets: on the one hand, it performs spatial correlation analysis on the data of multiple electromechanical equipment within the same road segment, identifying systemic disturbances such as power supply fluctuations and environmental wind pressure resonance by calculating the correlation coefficient of fault feature vectors; on the other hand, it performs time series analysis on the data of the same electromechanical equipment at different time points (multiple passes), calculating the rate of change of fault feature vectors to identify slowly evolving faults such as bearing wear and performance degradation. Based on the analysis results of both spatial and temporal dimensions, the edge AI computing unit corrects the initial diagnosis, generates a more accurate secondary diagnosis, and marks the data into four priorities: urgent, warning, attention, and normal, according to the urgency of the fault. For the large amount of data marked as normal priority, the gateway further performs compression processing, aggregating it into a statistical summary, thereby significantly reducing the amount of data uploaded. When the vehicle travels to an area with 4G / 5G public network signals, the cellular communication module uploads the enhanced semantic data packets (and statistical summaries) to the cloud platform according to priority.

[0117] The cloud platform layer is responsible for the aggregation and in-depth analysis of global data. After receiving uploaded data from each vehicle gateway, the data aggregation module first performs spatiotemporal alignment according to the electromechanical equipment ID and timeline, constructing a sparse but continuous state time series for each electromechanical equipment. The cloud platform utilizes swarm intelligence analysis methods. On one hand, by calculating the difference between the actual and theoretical sensing frequencies, it performs negative state monitoring. When an electromechanical equipment has not reported data for an extended period while traffic flow is normal on the road segment, it infers that the equipment may have experienced power failure or communication module malfunction, directly generating an emergency maintenance work order. On the other hand, it inputs the state time series into a predictive maintenance model (temporal convolutional network and XGBoost) to assess the remaining lifespan and future failure probability of each electromechanical equipment, generating predictive maintenance work orders. The operation and maintenance management module, based on the work order content and considering the location of the electromechanical equipment, spare parts inventory, personnel skills, and scheduling, intelligently dispatches maintenance tasks to the mobile terminals of maintenance personnel.

[0118] The system also possesses closed-loop self-evolution capabilities. The cloud platform periodically uses actual fault repair records as tags to optimize and train the 1D-CNN classifier, isolated forest model, and support vector machine at the terminal layer, as well as the temporal convolutional network and XGBoost classifier at the cloud layer. After training, the model parameters are sent to the vehicle-mounted intelligent gateway via OTA. When maintenance vehicles or heavy trucks pass by the passive monitoring terminals again, the vehicle-mounted intelligent gateway transmits the updated model parameters to the terminal via a short-range communication module. Upon receiving the parameters, the TinyML processing chip loads the new parameters to update or replace the local model, achieving continuous improvement in diagnostic accuracy without manual on-site intervention.

[0119] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for real-time monitoring of the operating status of electromechanical equipment on highways, characterized in that, The method is executed by a passive monitoring terminal deployed on the electromechanical equipment of the highway, and the specific steps are as follows: Step 1: Capture the mechanical vibration energy generated by the vehicle moving on the road surface and convert the mechanical vibration energy into a wake-up electrical signal; Step 2: Based on the wake-up electrical signal, trigger the feature extraction operation to extract the waveform features of the current vibration event, and perform pattern recognition on the waveform features to determine the type of vibration source that generates the vibration. Step 3: Determine whether it is a preset effective transfer carrier based on the vibration source type; If so, the data acquisition function will be activated to collect real-time operating status data of the electromechanical equipment; Step 4: Extract edge features and perform preliminary fault diagnosis on the collected real-time operating status data to generate semantic data packets; Step 5: Push the semantic data packets to an effective transfer carrier via short-range radio frequency, so that the semantic data packets can be transferred to the network coverage area by the movement of the effective transfer carrier and then transmitted back to the cloud for in-depth analysis in the cloud.

2. The method for real-time monitoring of the operating status of electromechanical equipment on highways according to claim 1, characterized in that, In step 1, the mechanical vibration energy generated by the vehicle's movement on the road surface is captured and converted into a wake-up electrical signal. This specifically includes the following steps: When the vehicle travels to the preset range of the passive monitoring terminal, the road vibration causes the piezoelectric material of the piezoelectric energy acquisition module to deform and output an alternating voltage signal. After amplification and noise filtering, the alternating voltage signal is converted into a digital signal sequence, which serves as the wake-up electrical signal.

3. The method for real-time monitoring of the operating status of electromechanical equipment on highways according to claim 2, characterized in that, In step 2, a feature extraction operation is triggered based on the wake-up electrical signal to extract the waveform features of the current vibration event, and pattern recognition is performed on the waveform features to determine the type of vibration source. Specifically, this includes the following steps: The passive monitoring terminal is equipped with a supercapacitor and has a dual signal processing channel on the piezoelectric energy acquisition module, which includes an energy channel and an information channel. The information channel is used to generate a wake-up electrical signal and to perform bandpass filtering on the digital signal sequence to obtain the vibration waveform; The energy channel is used to rectify and convert alternating voltage signals to charge the supercapacitor, which serves as the power supply for the passive monitoring terminal after it is woken up. When the TinyML processing chip is awakened by the wake-up electrical signal, the TinyML processing chip extracts the time-domain and frequency-domain features of the vibration waveform, and performs splicing and normalization operations in sequence to obtain the normalized feature vector. The normalized feature vectors are input into a pre-trained lightweight neural network model for classification to identify the vibration source type, which is one of the following: heavy truck, small passenger car, maintenance vehicle, or non-vehicle interference. When the vibration source is identified as a maintenance vehicle or a heavy truck, it is determined to be a valid shuttle carrier; otherwise, it is determined to be an invalid vibration source, and the passive monitoring terminal returns to sleep mode.

4. The method for real-time monitoring of the operating status of electromechanical equipment on highways according to claim 3, characterized in that, After the TinyML processing chip is woken up by the wake-up electrical signal, a predictive activation step is also included. The specific steps of the predictive activation step are as follows: By analyzing the phase difference of the vibration waveform, the vehicle's direction of travel and speed are calculated, and the time window for the vehicle to reach the monitoring point is predicted. Before the predicted time window arrives, the sensors and short-range communication modules are activated in advance to achieve seamless communication with the vehicle.

5. The method for real-time monitoring of the operating status of electromechanical equipment on highways according to claim 4, characterized in that, In step 4, edge feature extraction and preliminary fault diagnosis are performed on the collected real-time operating status data to generate semantic data packets, specifically including the following steps: The raw data includes vibration signals, current signals, and temperature signals. The vibration signals, current signals, and temperature signals are preprocessed separately to obtain preprocessed signals. Using the feature extraction model built into the TinyML processing chip, fault feature vectors are extracted from the preprocessed signal. The fault feature vectors include vibration features, current features, and temperature features. The fault feature vector is input into a lightweight fault diagnosis model to obtain an initial diagnosis conclusion. The extracted fault feature vectors, along with the initial diagnosis, electromechanical equipment ID, and timestamp, are structurally encapsulated to generate semantic data packets.

6. The method for real-time monitoring of the operating status of electromechanical equipment on highways according to claim 5, characterized in that, The effective shuttle carrier is equipped with an in-vehicle intelligent gateway. Semantic data packets are received and cached by the in-vehicle intelligent gateway. After pushing the semantic data packets to the effective shuttle carrier via short-range radio frequency, the in-vehicle intelligent gateway also performs fusion analysis and secondary diagnostic steps, as follows: Spatial correlation analysis is performed on semantic data packets sent by multiple passive monitoring terminals on the same road segment. The correlation coefficient of fault feature vectors of each electromechanical device is calculated. When the correlation coefficient exceeds the preset threshold, it is determined that there is a power supply abnormality or environmental disturbance, and the spatial analysis results are obtained. Time series analysis is performed on the semantic data packets sent by the same passive monitoring terminal during multiple passes, the rate of change of the fault feature vector is calculated, and when the rate of change continues to rise, it is determined that there is a slow degradation fault, and the time analysis results are obtained. Based on the spatial and temporal analysis results, the initial diagnosis was revised to generate a more accurate secondary diagnosis. Based on more accurate secondary diagnostic conclusions, abnormal data are prioritized and classified to obtain classification results. The hierarchical results, spatial analysis results, temporal analysis results, and more accurate secondary diagnostic conclusions are added to the semantic data package to generate an enhanced semantic data package.

7. The method for real-time monitoring of the operating status of electromechanical equipment on highways according to claim 6, characterized in that, For multiple enhanced semantic data packets marked as normal priority in the classification results, the in-vehicle intelligent gateway also performs compression processing: The enhanced semantic data packets of the normal status of multiple passive monitoring terminals within the same road segment and the same time window are used to generate a statistical summary. The statistical summary includes the road segment identification, the total number of electromechanical devices, the number of normal electromechanical devices, the average temperature, and the average vibration amplitude. Upload only the statistical summary along with the enhanced semantic data package for abnormal electromechanical equipment to reduce the amount of data uploaded.

8. The method for real-time monitoring of the operating status of electromechanical equipment on highways according to claim 7, characterized in that, In step 5, the specific steps of the cloud-based deep analysis are as follows: The enhanced semantic data packets from different vehicles and time points are spatiotemporally aligned according to the electromechanical equipment ID and time axis to construct the state time series of each electromechanical equipment. Calculate the theoretical frequency at which each electromechanical device is sensed by the vehicle, and obtain the number of records reported by deduplicated vehicles as the actual sensing frequency. When the actual sensing frequency is lower than the theoretical frequency and the traffic flow on the road section is normal, negative state monitoring is triggered, inferring that the electromechanical equipment has a power failure or communication failure, and directly generating a power failure / failure type maintenance work order. For electromechanical equipment that has not triggered negative condition monitoring, the corresponding state time series data is input into the predictive maintenance model to calculate the remaining service life and failure probability of the equipment and generate predictive maintenance work orders.

9. The method for real-time monitoring of the operating status of electromechanical equipment on highways according to claim 8, characterized in that, It also includes the model self-evolution step: The cloud platform optimizes at least one machine learning model used in the implementation of the method based on actual fault repair records to obtain optimized model data. The optimized model data is sent to the effective shuttle carrier via OTA over-the-air upgrade. Then, the mobile attributes of the effective shuttle carrier are used to transmit model parameters to each passive monitoring terminal via short-range radio frequency. After receiving the optimized model data, the passive monitoring terminal updates or replaces the corresponding local model based on the optimized model data, thereby continuously improving the diagnostic accuracy.

10. A real-time monitoring system for the operating status of electromechanical equipment on highways, characterized in that, The system applies the real-time monitoring method for the operating status of electromechanical equipment on highways as described in any one of claims 1 to 9, and the system includes a passive monitoring terminal deployed on the electromechanical equipment of the highway; the passive monitoring terminal includes: Piezoelectric energy harvesting module, used for: It captures the mechanical vibration energy generated by the vehicle's movement on the road surface and converts the mechanical vibration energy into a wake-up electrical signal; TinyML processing chip, used for: Based on the wake-up electrical signal, a feature extraction operation is triggered to extract the waveform features of the current vibration event, and pattern recognition is performed on the waveform features to determine the type of vibration source that generates the vibration. Based on the type of vibration source, determine whether it is a preset effective ferry carrier; Edge feature extraction and preliminary fault diagnosis are performed on the collected real-time operating status data to generate semantic data packets; If so, then activate the data acquisition function; Sensor modules are used for: In response to the wake-up data acquisition function signal triggered by the TinyML processing chip, the real-time operating status data of the electromechanical equipment is collected and sent to the TinyML processing chip. Short-range communication module, used for: Semantic data packets are pushed to an effective ferry carrier via short-range radio frequency, so that the semantic data packets can be transferred to the network coverage area by the movement of the effective ferry carrier and then transmitted back to the cloud for in-depth analysis in the cloud.

Images