Sensor fusion with machine learning for detection of partial discharge on electrical transmission infrastructure

Abstract

A monitoring device, system and method for detecting a condition of an infrastructure asset is provided. The monitoring device comprises: at least first and second sensors, a wireless communication interface and a memory, all of which are controlled by an application processor. When the monitoring device is mounted proximate to the infrastructure asset, the first sensor detects a first signal emitted by the infrastructure asset and the second sensor simultaneously detects a second signal emitted by the infrastructure asset, the first and second detected signals recorded to memory. The application processor processes each of the first and second recorded signals via at least one digital signal processing (DSP) module configured to perform feature extraction on the first and second recorded signals. A classifier model analyzes the extracted features of the first and second signals to generate a classification indicating a condition of the infrastructure asset.

Description

[0001] SENSOR FUSION WITH MACHINE LEARNING FOR DETECTION OF PARTIAL DISCHARGE ON ELECTRICAL TRANSMISSION INFRASTRUCTURE

[0002] Field:

[0003] The present disclosure relates to monitoring electrical transmission infrastructure; in particular, the present disclosure relates to monitoring remotely located electrical transmission infrastructure utilizing sensors to detect changes in the infrastructure itself and in the environment surrounding the electrical transmission infrastructure.

[0004] Background:

[0005] Electrical partial discharge is an indication of the breakdown of insulation between an electrical conductor and the ground, or between two electrical conductors. When electrical partial discharges occur in remote areas near dry vegetation a fire may be ignited, causing significant damage to the environment and property.

[0006] Monitoring electrical transmission infrastructure for the breakdown of insulation is challenging, as electrical transmission infrastructure is often located in remote, forested areas, covering vast geographical areas. An existing method of detecting partial discharge on high voltage electrical transmission infrastructure is to monitor patterns in the low-frequency ultrasound range of acoustics (~20khz - ~50khz). For example, as described in the Applicant’s Patent Cooperation Treaty Publication No. WO2023235981 Al, the entirety of which is incorporated herein by reference, monitoring devices equipped with ultrasonic sensors are configured to periodically sample audio signals emitted by electrical transmission infrastructure within the ultrasonic range. Machine learning (ML) models are trained on large audio data sets, and configured to analyse the resulting audio data captured by the monitoring devices to detect patterns indicating that a breakdown in the insulators has occurred, or is developing. Other existing methods include detecting partial discharges on high voltage electrical transmission infrastructure through measuring radiofrequency (RF) signals that are emitted when a partial discharge event occurs.

[0007] An issue with the existing technology is that the sampled audio data or sampled RF data may be contaminated by non-electromagnetic events, which may inhibit the monitoring device from accurately detecting and interpreting the ultrasonic (or other audio) signals or the RF signals thatresult from problems developing on the electrical transmission infrastructure. Examples of non-electromagnetic events that emit audio signals include birds, wind, weather, mechanical vibration or other environmental events. The audio signals generated by such environmental events are typically outside of the ultrasonic range. Such non-ultrasonic audio signals may overwhelm the monitoring device and mask the relevant ultrasonic audio signals that the monitoring device is attempting to detect. As well, other events may occur in the environment that emit ultrasonic signals, but are not caused by electrical partial discharges. These ultrasonic signals need to be differentiated from the partial discharge ultrasonic signals that are relevant to the breakdown of insulators on the electrical transmission infrastructure. Sources of RF signals that are unrelated to partial discharges occurring on electrical infrastructure include, but are not limited to, telecommunications (including radio and television broadcasting, mobile phone networks, satellite communications, and other wireless communication systems) and radar systems, as well as the sun and the earth’s geomagnetic field.

[0008] Where ML models are used to detect patterns in the ultrasonic data obtained by the monitoring devices, ideally, the ultrasonic signals resulting from events unrelated to electrical partial discharges would be removed from the data set and not analysed. However, differentiating the relevant ultrasonic signals from the non-relevant ultrasonic signals in a data set may be a complex exercise, requiring additional computing and power resources on the monitoring device. Increased power and computing requirements for the monitoring device may be difficult and impractical to implement, given the nature of remotely- located electrical transmission infrastructure, which presents challenges in providing an adequate and reliable power source to the monitoring device that is attached to the remotely -located electrical transmission infrastructure.

[0009] In one aspect of the present disclosure, a system for detecting a condition of an infrastructure is provided. The infrastructure comprises a plurality of infrastructure assets. The system comprises: a plurality of monitoring devices, each monitoring device of the plurality of monitoring devices mounted proximate to an infrastructure asset of the plurality of infrastructure assets; and a central server located remotely from the electrical transmission infrastructure, the central server in wireless communication with each monitoring device of the plurality of monitoring devices. Eachmonitoring device of the plurality of monitoring devices comprises: first and second partial discharge sensors, a wireless communication interface and a memory, each of the first and second partial discharge sensors, the wireless communication interface and the memory controlled by an application processor. The first partial discharge sensor is configured to detect a first signal emitted by the infrastructure asset to generate a first data stream and the second partial discharge sensor is configured to simultaneously detect a second signal emitted by the infrastructure asset to generate a second data stream, and the first and second data streams are recorded to the memory. The application processor performs pre-processing on each of the first and second data streams via at least one digital signal processing (DSP) module, wherein the at least one DSP module is configured to perform feature extraction on the first and second data streams to generate first and second feature vectors. At least one classifier model analyzes the first and second feature vectors to generate a classification, the classification indicating a condition of the infrastructure asset.

[0010] In some embodiments of the system, the first partial discharge sensor is an ultrasonic microphone and the first signal is an audio signal in the ultrasonic range. In some embodiments, the second sensor is a radiofrequency (RF) sensor and the second signal is an RF signal. The RF sensor may be selected from one or more of the group comprising: a superheterodyne RF sensor, an SDR, a logarithmic envelope detector. In some embodiments, the monitoring device further comprises an environmental sensor, the environmental sensor configured to measure at least one characteristic of an environment surrounding the infrastructure asset to generate an environmental data stream. The environmental data stream is analyzed by the central server or the application processor to validate the generated classification. The at least one characteristic of the surrounding environment may be selected from a group comprising: temperature, air pressure, humidity.

[0011] In some embodiments of the system, the pre-processing performed on each of the at least first and second data streams by the application processor includes a first DSP module performing feature extraction on the first data stream to generate the first feature vector and a second DSP module performing feature extraction on the second data stream to generate the second feature vector. In some embodiments, the pre-processing performed on each of the at least first and second data streams by the application processor further includes a feature-level fusion module configured tocombine the first and second feature vectors generated by the at least first and second DSP modules to generate a combined feature vector.

[0012] In some embodiments of the system, the at least one classifier model resides on the application processor and the generated classification is transmitted via the wireless communication interface of the monitoring device to the central server over a wireless network. In some embodiments, the at least one classifier model resides on the central server and wherein the first and second feature vectors generated by the at least one DSP module are transmitted, via the wireless communication interface of the monitoring device, to the central server over a wireless network. In such embodiments, the analysis of the extracted features is performed by the classifier model, on the central server, to generate the classification.

[0013] In some embodiments of the system, the at least one classifier model of each monitoring device is configured as one multi-channel ML model for ingesting the at least first and second feature vectors extracted from the first and second data streams to generate the classification. In some embodiments, the at least one classifier model of each monitoring device is configured as one single-channel ML model for ingesting the combined feature vector generated by the feature-level fusion module to generate the classification. In some embodiments, the at least one classifier model is configured as a first ML model for ingesting the first feature vector to generate a first classification and a second ML model for ingesting the second feature vector to generate a second classification, and wherein the first classification indicating the condition of the infrastructure asset is validated by the second classification.

[0014] In some embodiments of the system, each monitoring device further comprises a low-power coprocessor in communication with the application processor and at least one low-power continuous sensor. The at least one low-power continuous sensor continuously monitors for a transient signal exceeding a predetermined threshold value. When the transient signal detected by the at least one low-power continuous sensor exceeds the predetermined threshold value, the low-power coprocessor signals the application processor to initiate a data processing operation and a logging operation. The logging operation includes at least initiating a simultaneous recording of the first and second data streams obtained by the first and second partial discharge sensors, and the dataprocessing operation including the at least one DSP module generating first and second feature vectors from the first and second data streams.

[0015] In some embodiments of the system, each monitoring device further comprises at least one low-power continuous sensor in communication with the application processor. The at least one low-power continuous sensor is configured to continuously monitor for a transient signal exceeding a predetermined threshold value, and the application processor is a multi-core processor. When the transient signal received by the at least one low-power continuous sensor exceeds the predetermined threshold value, the multi-core application processor initiates a data processing operation and a logging operation. The logging operation includes at least initiating a simultaneous recording of the first and second data streams obtained by the first and second partial discharge sensors, and the data processing operation includes the at least one DSP module generating first and second feature vectors from the first and second data streams. The at least one low-power continuous sensor may be selected from a group comprising: arc detection sensor, high-frequency current transformer, audio sensor, RF sensor, optical sensor, thermal sensor, vibration sensor, environmental sensor.

[0016] In some embodiments of the system, each monitoring device further comprises a tilt sensor in communication with the application processor, the tilt sensor configured to periodically detect an orientation of the corresponding monitoring device relative to a ground to generate a tilt data stream, the tilt data stream transmitted wirelessly to the central server for monitoring movement of the respective infrastructure asset over time.

[0017] In some embodiments of the system, the infrastructure asset may be selected from a group of electrical infrastructure assets, the group comprising: switch-gear, insulator, conductor, subconductor, pole, tower. In such embodiments, the condition of the infrastructure asset may be selected from a group comprising: physical failure, corona discharge, partial discharge, arcing, flashover, tracking, pollution, operational baseline condition. In some embodiments, the infrastructure may be selected from a group comprising: electrical transmission infrastructure, electrical distribution infrastructure, electrical generation infrastructure, electrical substation infrastructure, transportation infrastructure, tunnels, roadways, overpasses, bridges, roadwaylighting systems, traffic control systems, pipeline infrastructure, oil and gas infrastructure, railway systems.

[0018] In some embodiments of the system, the central server is configured to receive at least a plurality of classifications generated by the plurality of monitoring devices to generate an aggregated data set. In such embodiments, the central server analyses the aggregated data set to generate a recommendation for maintenance of each infrastructure asset of the plurality of infrastructure assets so as to generate a plurality of recommendations for maintenance of the plurality of assets. Each recommendation of the plurality of recommendations is prioritized based on an assessed criticality of the generated condition of each asset of the plurality of assets.

[0019] In some embodiments of the present disclosure, a monitoring device for detecting a condition of an infrastructure asset is provided. The monitoring device comprises: at least first and second partial discharge sensors, a wireless communication interface and a memory, each of the first and second partial discharge sensors, the wireless communication interface and the memory controlled by an application processor. The first partial discharge sensor is configured to detect a first signal emitted by the infrastructure asset to generate a first data stream and the second partial discharge sensor is configured to simultaneously detect a second signal emitted by the infrastructure asset to generate a second data stream, the first and second data streams recorded to the memory. The application processor performs pre-processing on each of the first and second data streams via at least one digital signal processing (DSP) module, wherein the at least one DSP module is configured to perform feature extraction on the first and second data streams to generate first and second feature vectors. At least one classifier model residing on the application processor analyzes the first and second feature vectors to generate a classification, the classification indicating a condition of the infrastructure asset.

[0020] In some embodiments of the present disclosure, a method for monitoring a condition of an infrastructure network is provided. The infrastructure network comprises a plurality of infrastructure assets. The method comprises: receiving, on a central server in wireless communication with the plurality of monitoring devices of the present disclosure deployed on the plurality of infrastructure assets, at least a plurality of classifications generated by the plurality ofmonitoring devices to generate an aggregated data set; performing a comparative data analysis on the aggregated dataset to identify at least one trend indicating the performance or condition of one or more assets of the plurality of infrastructure assets.

[0021] In some embodiments of the method, each device of the plurality of monitoring devices includes an environmental sensor that generates an environmental data stream, the environmental data stream received by the central server and included in the aggregated data set. In some embodiments of the method, each device of the plurality of monitoring devices includes a tilt sensor that generates a tilt data stream, the tilt data stream received by the central server and included in the aggregated data set.

[0022] In some embodiments of the method, each monitoring device further comprises at least one low-power continuous sensor in communication with the application processor, wherein the at least one low-power continuous sensor continuously monitors for a transient signal exceeding a predetermined threshold value. In such embodiments, when the transient signal detected by the at least one low-power continuous sensor exceeds the predetermined threshold value, the application processor initiates a data processing operation and a logging operation. The logging operation includes at least initiating a simultaneous recording of the first and second data streams obtained by the first and second partial discharge sensors. The data processing operation includes the at least one DSP module generating first and second feature vectors from the first and second data streams. In some embodiments of the method, the at least one low-power continuous sensor may be selected from a group comprising: arc detection sensor, high-frequency current transformer, audio sensor, RF sensor, optical sensor, thermal sensor, vibration sensor, environmental sensor.

[0023] In some embodiments of the method, the infrastructure asset may be selected from a group of electrical infrastructure assets, the group comprising: switch-gear, insulator, conductor, subconductor, pole, tower. In some embodiments of the method, the condition of the infrastructure asset may be selected from a group comprising: physical failure, corona discharge, partial discharge, arcing, flashover, tracking, pollution, operational baseline condition. In some embodiments of the method, the infrastructure may be selected from a group comprising: electrical transmission infrastructure, electrical distribution infrastructure, electrical generationinfrastructure, electrical substation infrastructure, transportation infrastructure, tunnels, roadways, overpasses, bridges, roadway lighting systems, traffic control systems, pipeline infrastructure, oil and gas infrastructure, railway systems.

[0024] In some embodiments of the method, the central server analyses the aggregated data set to generate a plurality of recommendations for maintenance of the plurality of assets, wherein each recommendation of the plurality of recommendations is prioritized based on an assessed criticality of the generated condition of each asset of the plurality of assets. In some embodiments of the method, the at least one long-term trend provides an analysis of the behavior of partial discharge events in relation to seasonal cycles and geographical factors, and wherein the method further includes a step of refining the at least one classifier model of the plurality of monitoring devices. In some embodiments of the method, the at least one trend is utilized to establish an operational baseline for at least one group of infrastructure assets of the plurality of infrastructure assets, wherein each infrastructure asset in the at least one group of infrastructure assets shares a common characteristic, the common characteristic selected from a group comprising: asset type, asset location, asset environmental conditions.

[0025]

[0026] of the Drawings:

[0027] FIG. 1 is an example of an ultrasound signal waveform and the corresponding radiofrequency signal waveform, representing the ultrasound and radio frequency signals emitted by an electrical discharge event.

[0028] FIG. 2 is a block diagram of the system and sensor components of an embodiment of the monitoring device.

[0029] FIG. 3 is a block diagram of the system and sensor components of another embodiment of the monitoring device.

[0030] FIG. 4 is a block diagram of the system and sensor components of another embodiment of the monitoring device.FIG. 5 is a block diagram of DSP modules and a multi-channel ML model, where both RF data and audio data are fed into the multi-channel ML model.

[0031] FIG. 5 A is a block diagram of DSP modules, a feature-level fusion module and a ML model, where both RF data and audio data are fed into the ML model.

[0032] FIG. 6 is a block diagram of a plurality of ML models where separate data streams are fed into corresponding separate ML models.

[0033] FIG. 7 is a block diagram of a plurality of ML models where separate data streams are fed into corresponding separate ML models and the ML models are applied on the edge computing device.

[0034] FIG. 8 is a block diagram of a plurality of ML models where separate data streams are fed into corresponding separate ML models and the ML models are applied in the cloud.

[0035] FIG. 9 is a schematic diagram of a system comprising a plurality of monitoring devices.

[0036] Detailed Description:

[0037] In solving the challenge of monitoring remotely located electrical transmission infrastructure in order to detect partial electrical discharges or other events that indicate a breakdown in the insulators or other components of the infrastructure is occurring, a plurality of monitoring devices, which either incorporate or interact with an adjacent edge computing device, may be installed directly on, or in close proximity to, the infrastructure. In some embodiments of the present disclosure, the monitoring devices are provided with at least two different sensors and at least one application processor. The application processor controls the inputs and outputs of the device and performs digital signal processing on the data recorded by the at least two sensors.

[0038] In one aspect, a pre-processing step is performed on the raw sensor signal data via one or more Digital Signal Processing (DSP) modules. In such embodiments, the DSP module performs feature extraction to produce data representations, otherwise referred to herein as feature vectors, that are optimized for ML inference. Analysis of the pre-processed signal data is performed by a machine learning (ML) model, otherwise referred to herein as a classifier model. The classifier model is used to analyse the at least two recorded data streams sensed by the at least two sensors, identify patterns in the recorded sensor signal data which indicate the condition of the electricalinfrastructure, and thereby generate a classification of the recorded sensor signal data. The classification may indicate, for example, that a probable partial electrical discharge or a corona discharge has been detected on the electrical infrastructure proximate to the monitoring device, thereby indicating that an electrical insulator is breaking down. In other examples, the classification may indicate that the condition of the proximate electrical infrastructure is in a normal state, or in other cases, that the condition of the infrastructure is beginning to deteriorate and may lead to a partial discharge event in the future. Once the recorded data has been classified, the classification may be transmitted to a central monitoring location via a wireless communication interface.

[0039] In some embodiments, the pre-processing step may further include providing the outputs of the one or more DSP modules to a feature-level fusion module, which combines multiple feature vectors extracted from each of the recorded data streams into a single, enriched and more comprehensive feature vector. In such embodiments, the resulting, combined feature vector is then provided to the classifier model for characterization of the detected event.

[0040] The central monitoring location may be a central server or plurality of central servers at a location that is remote from the infrastructure being monitored by the monitoring devices. In other embodiments, the central monitoring location may be located in a cloud network and accessible from any physical location. At the central monitoring location, access to the classification data received from a plurality of monitoring devices, mounted across a network of electrical infrastructure, may be monitored in order to detect developing problems on the infrastructure, which in some embodiments is being monitored in real time or in near-real time. Advantageously, the operator of the electrical infrastructure may be alerted to a high risk condition detected on the electrical infrastructure, and perform maintenance or other preventative measures before a catastrophic failure of the electrical transmission line occurs. Additionally, providing such real time, near-real time, or frequent monitoring capability (for example, checking every hour, every 12 hours or every 24 hours), may facilitate the efficient scheduling of regular repair and maintenance in order to maintain the electrical infrastructure in an optimum operating condition.

[0041] This ability to monitor the condition of the electrical infrastructure network, from a central location, advantageously provides the benefit of more effectively and efficiently maintaining thenetwork, while potentially conserving significant resources that would otherwise be spent in an attempt to physically attend to and monitor sections of the electrical infrastructure on the ground, which is costly and inefficient given the vast geographic area that such electrical infrastructure networks will typically cover. In some embodiments, analysis of the aggregated data set of data collected from the plurality of monitoring devices 200 across a monitoring system may allow for generating a recommended maintenance schedule, whereby the maintenance or repair of assets is ranked in order of priority, according to the assessed criticality of the condition of a given asset or grouping of assets. In another aspect of the present disclosure, the detection of a transient signal or event, which may indicate critical conditions or events that have occurred within the network of electrical infrastructure being monitored, may result in the central server generating an alert so that the utility operator may take immediate action to address the issue. For example, if an arcing event or a catastrophic failure is detected, such as a downed powerline or a fire, an alert may be sent to the utility operator to take immediate action.

[0042] An example embodiment of the monitoring device 200 is schematically illustrated in FIG. 2. The monitoring device 200 may include at least two sensors; in the example shown, the at least two sensors includes a radiofrequency (RF) sensor 240 and an ultrasonic microphone 250. The RF sensor 240 and the ultrasonic microphone 250 are in electronic communication with the application processor 220, which in some embodiments may be a microcontroller unit (MCU), although it will be appreciated that any suitable processor may be utilized. The application processor 220 receives and processes the signals recorded by the RF sensor 240 and the ultrasonic microphone 250. The application processor 220 stores onto, and reads data from, a memory unit 230. In addition, the application processor 220 sends signals to a wireless communication interface 260, for wirelessly transmitting data to a server of a central monitoring location.

[0043] Optionally, in some embodiments the monitoring device 200 may also include an environmental sensor 270. The environmental sensor 270 may be configured to measure the conditions of the environment surrounding the monitoring device 200, which may include, for example, the surrounding ambient temperature, air pressure and humidity. For embodiments including an environmental sensor 270, the data stream obtained from the environmental sensor 270 is also provided to the application processor 220. Each of the components of the monitoring device 200, described herein, may be contained within, or supported on, a housing 210 of the device.Preferably, the housing 210 is constructed to protect the electronics of the device from the surrounding environment and built to withstand harsh weather conditions, including extreme heat, cold, wind, precipitation, exposure to the sun, etc.

[0044] In some embodiments of the present disclosure, the microphone of the ultrasound sensor 250 includes a microphone array, including two or more ultrasound microphones, whereby each ultrasound microphone in the array is positioned in a different direction to capture audio signals emitted by specific sub-components of an infrastructure asset being monitored. For example, this may occur where a monitoring device is mounted to a transmission line pole supporting a bundle of three sub-conductors suspended from insulators. In this example, the monitoring device may include a microphone array of three ultrasound microphones, each ultrasound microphone positioned to sample and record audio signals emitted by one insulator supporting one subconductor of the bundle of three sub-conductors. In such configurations of monitoring devices, the monitoring device may capture three different signals and three different data sets during each sampling period, and the subsequent classifier steps are performed by the edge computing device on each of the three audio signals recorded by the monitoring device.

[0045] In use, the partial discharge sensors 240, 250 sample measurements of the surrounding environment. In some embodiments, the sensors 240, 250 may be configured to continually sample measurements; in other embodiments, the sensors 240, 250 may be configured to sample signals at regular intervals, such as anywhere from once every minute to once every hour to once every few hours or once every 24 hours, or any other suitable sampling frequency. In the illustrated example, the sensors 240, 250 would receive RF signals and ultrasound signals emitted by the adjacent infrastructure onto which the device is mounted. These signals are communicated to the application processor 220, which, as will be further described below, performs any pre-processing steps on the raw recorded data stream, such as applying DSP to extract features from the recorded data, and then applies the classifier model to the extracted features to analyze the signals and generate a classification of the two data streams obtained from the partial discharge sensors 240, 250.

[0046] The generated classification data may indicate, for example, that a partial discharge event, or other characteristic of the infrastructure, has been detected; in other examples, the generatedclassification may indicate that the infrastructure is operating under normal conditions or within expected parameters. The generated classification data may be stored on the memory 230, and may be transmitted via the wireless communication interface 260 at regular intervals to a central server or central monitoring location. In this example embodiment, because the classifier model is operating on the edge computing device (or in other words, on the monitoring device 200), the generated classification data may typically be a condition code or other small amount of data. Such small data packets may be readily transmitted on the remotely located wireless networks accessible by the remotely located monitoring devices 200, even where there is limited bandwidth available on those wireless networks. Thus, the ability to transmit a classification code, rather than the data streams themselves, to a central server location, may address the issue that access to wireless networks with large bandwidth may be scarce, expensive or non-existent.

[0047] In another embodiment of the monitoring device 200, shown in FIG. 3, in addition to the application processor 220 there is a low power co-processor 280. The low power co-processor 280 may be configured for low-power, always-on monitoring of unexpected events that may require continuous monitoring in order to be captured by the monitoring device. Examples of such unexpected events include, but are not limited to, arcing or flashover events, poles or towers that have been brought down by wind, lightning, wildfires or other causes, active fires on poles or towers, etc. Because such events are relatively rare and unexpected, it is unlikely that the monitoring device will capture sensor signals related to these events when the monitoring device 200 is only sampling on a pre-determined frequency, such as every few minutes or every few hours. Thus, in such embodiments, the low-power co-processor 280 is in communication with the application processor 220 and is optimized for power efficiency, thus allowing continuous operation of the low-power co-processor 280 without significantly impacting the overall power consumption of the monitoring device 200. Upon detecting a signal from an arc detection sensor 290 that exceeds a predetermined threshold value, the low-power co-processor 280 signals the application processor 220 to initiate higher-level data processing or logging operations. Higher-level data processing may include, for example, additional data collection from the ultrasonic microphone (250) and RF sensor (240) to validate the detected arcing or flashover event. Advantageously, the use of a low-power co-processor, in communication with an arc detection sensor 290, may assist with maintaining low power consumption by only engaging the applicationprocessor when a likely arcing event, or other event of interest, has been detected by one of the low-power, always-on sensors (such as arc detection sensor 290).

[0048] Similarly, and optionally, any one of the following sensors may be continuously monitored by the low-power co-processor 280, such that when any of these sensors sense a signal that exceeds a predetermined threshold value, the low-power co-processor 280 signals the application processor 220 to initiate higher-level data processing or logging operations; such sensors include, but are not limited to: the environmental sensor 270; the vibration sensor 291 and the arc detection sensor 290.

[0049] Referring now to FIG. 4, another embodiment of the monitoring device 200 omits the low power co-processor 280, and all sensors are configured to be in direct electronic communication with the application processor 220. This includes the primary sensors, such as the RF sensor 240 and the ultrasonic microphone 250, as well as the suite of continuous monitoring sensors, which may include one or more of an environmental sensor 270, vibration sensor 291, arc detection sensor 290, and tilt sensor 292, and / or any other sensor or sensors which may be configured as a continuous monitoring sensor, represented as the Nth sensor 293. In such embodiments, the application processor 220 may be a multi-core unit, comprising two or more CPU cores, which may be configured for employing sophisticated power management techniques, allowing for the maintenance of low-power operation without requiring a dedicated and separate co-processor. For example, a single, highly efficient CPU core may be designated for processing the suite of continuous monitoring sensors, while one or more of the other CPU cores of the application processor 220 may be configured to remain in a deep sleep state, for processing the signal data obtained from the primary sensors, such as the RF sensor 240 and the ultrasonic microphone 250, which are scheduled to perform sampling at a pre-determined sampling frequency. This allows the device to achieve a low-power, always-on monitoring capability, similar to the co-processor embodiment, but within a single integrated circuit.

[0050] Advantageously, a single multi-core application processor architecture may still perform concurrent tasks effectively, while potentially reducing manufacturing costs for the monitoring devices. A multi-core application processor 220 may dedicate specific processing resources to handle simultaneous inputs from multiple sensors without compromising performance. Forinstance, one core may be polling the vibration sensor 291 while another core processes a data buffer from the RF sensor 240, providing a configuration where no data is lost and high-level data processing and logging operations may be initiated without delay when an unexpected and infrequent event is detected.

[0051] The arc detection sensor 290, in some embodiments, may be a versatile sensor capable of monitoring for transient arcing or flashover events in near-real time. The arc detection sensor 290 may employ one or more of the following technologies or their equivalents, as would be known to a person skilled in the art, depending on the deployment specifications: a) high-frequency current transformer (HFCT), which utilizes high-frequency electromagnetic signals to detect arcing events through current pulses detected in conductors or in grounding systems, such as ground cables. The HFCT may be effective for detecting electrical anomalies that generate high-frequency emissions, such as arcing or flashover events; b) audio sensors, including for example microphones, which may be configured to detect loud bangs, impulsive noises, or other specific sound waves or sound wave patterns that indicate an arcing or flashover event. Such audio sensors may be configured to operate in both the audible range (<20 kHz) and ultrasonic range (>20 kHz) to capture a wide spectrum of transient sound signals; c) other sensor types, such as wideband RF antennas, optical sensors, thermal sensors, and the like, which are capable of detecting signals or conditions associated with arcing or flashover events.

[0052] In an exemplary embodiment, the arc detection sensor 290 may be a specialized, low-power integrated circuit designed for transient RF event detection. This implementation comprises a narrow-band antenna tuned to a specific frequency range known to be associated with electrical arcing (e.g., in the 300 kHz - 3 MHz range) and integrated signal processing circuitry. In this embodiment of an arc detection sensor 290, the functionality is characterized by a low-power listening state, where the sensor is able to continuously monitor for activity in its dedicated RF band while consuming minimal power, typically in the microampere range. The sensor's circuitry may include a hardwired or programmable algorithm that analyzes the temporal shape and energy of an incoming signal to distinguish the unique signature of an electrical arc from other RF signatures. Upon positive identification of an arcing or flashover event, the sensor 290 generates a signal to either the low power co-processor 280, which then wakes the application processor 220 (as shown in the embodiment of FIG. 3), or directly to the application processor 220 (as shown inthe embodiment of FIG. 4). By offloading the initial event detection and signature validation to a dedicated, ultra-low-power component, the monitoring device 200 may be configured for continuous monitoring and detection of critical arcing events while maximizing the operational longevity of the device 200.

[0053] In some embodiments, the arc detection sensor 290 may be configured to generate a signal that is proportional to the detected event, which signal is processed by the low-power co-processor 280 to determine if further actions are required, such as signaling the application processor 220 to initiate higher-level data processing or logging operations (as in the embodiment shown in FIG.

[0054] 3), or the arc detection sensor signal may be processed by the multi-core application processor 220 (as in the embodiment of FIG. 4). Advantageously, the flexibility in configuring an arc detection sensor 290 to employ one or more technologies, such as the integrated low-power circuit, HFCT, audio sensors, or other various sensors to detect transient arcing or flashover events, or other transient events, allows the system to adapt to a range of environmental and operational conditions, which may increase reliable detection and response. By providing for efficient detection of critical, yet transient, events, while minimizing energy consumption, such devices employing a low-power co-processor 280 or a multi-core application processor 220, in combination with an arc detection sensor 290, may render the monitoring device 200 suitable for long-term deployment in remote and / or resource-constrained environments.

[0055] Regarding the vibration sensor 291, such a sensor may be configured to monitor for wind- induced galloping or oscillations in transmission lines and / or support structures, which events may indicate abnormal stress or environmental conditions impacting infrastructure stability. Additionally, a vibration sensor 291 may detect loose hardware or mechanical fatigue by identifying characteristic vibration patterns or frequency signatures associated with structural degradation, loose hardware and / or failure. The vibration sensor 291 may also be used to capture seismic activity or external impacts, such as earthquakes or physical collisions, allowing the system to identify and respond to potentially catastrophic events in near real-time.

[0056] A tilt sensor 292, which may be an inclinometer, an accelerometer, or other types of sensors capable of detecting orientation of an object relative to the ground, may be configured to periodically monitor for gradual changes in the inclination or alignment of transmission line poles,towers, or other infrastructure, which could indicate structural instability, ground shifts, or foundation failure. A tilt sensor 292 detects long-term settling or creeping movements, such as gradual tilting caused by environmental factors like soil erosion, frost heave, or heavy loading conditions. The tilt sensor 292 may also detect sudden angular shifts or rotations caused by external events, such as collisions, extreme weather (e.g., high winds or ice loads), or seismic activity. In some embodiments, the tilt sensor 292 is in communication with the application processor 220 and configured to record measurements at periodic intervals, which may be at the same sampling frequency or a different sampling frequency as the measurements that are taken by the at least two partial discharge sensors 240, 250. Such periodic measurements by the tilt sensor 292 may therefore track gradual changes in the orientation of the electrical transmission support infrastructure, such as a pole or a tower, that may take place over time.

[0057] An environmental sensor 270, as illustrated in FIG. 2, may be in communication with the application processor 220 and configured to take readings of surrounding environmental conditions in accordance with the same predetermined frequency or schedule as the RF sensor 240 and the ultrasonic microphone 250. Alternatively, in embodiments comprising a low-power coprocessor 280 (as in FIG. 3) or a multi-core application processor 220 (as in FIG. 4), the environmental sensor 270 may be configured to communicate with the low-power co-processor or the multi-core application processor such that the environmental conditions are continuously monitored by the environmental sensor 270. When a signal is detected by the environmental sensor 270 that exceeds a predetermined threshold, the signal is processed by the low-power co-processor 280 to signal the application processor 220 to initiate higher-level data processing or logging operations. In embodiments employing a multi-core application processor 220 having two or more cores, as shown in FIG. 4, when a signal is detected by the environmental sensor 270 that exceeds a predetermined threshold, the signal is detected by a secondary core of the application processor 220, and the secondary core signals the primary core of the application processor 220 to initiate higher fidelity sensor data capture (logging) and / or processing operations.

[0058] In a configuration of the environmental sensor 270 as shown in FIG. 3 and FIG. 4, the environmental sensor may additionally be configured to take readings of the surrounding environmental conditions in accordance with the same predetermined frequency or schedule as the RF sensor 240 and the ultrasonic microphone 250. Advantageously, such an implementation ofan environmental sensor 270 allows for the continuous monitoring and detection of critical environmental events such as wildfires, extreme temperature changes, or abnormal humidity levels, allowing for the identification of events that could directly threaten the transmission line or surrounding infrastructure, while at the same time providing for the tracking of gradual environmental changes (e.g., sustained high temperatures, increasing humidity, or air quality variations), with the capability to transmit telemetry data to a cloud system for long-term trend analysis, providing for proactive maintenance and risk assessment. Examples of environmental sensor 270 functions include, but are not limited to, monitoring local atmospheric conditions (e.g., pressure drops, wind speed, or smoke particulates) that could provide early warnings for storms, high wind stress, or air quality issues, ensuring the system proactively responds to environmental threats. In some embodiments, this approach provides the dual functionality of real-time continuous monitoring and long-term data analysis for broader operational benefits, providing for both immediate response and strategic decision-making.

[0059] As compared to a one-channel classifier model, which is trained on a single set of data from a single sensor type, the Applicant has found that a multi-channel classifier model, trained on at least two sets of data obtained from at least two different types of s sensors, provides a more accurate classifier model. For example, not intended to be limiting, the Applicant has discovered that a multi-channel classifier model, utilizing both ultrasonic data and radiofrequency (RF) data recorded by a microphone and an RF sensor, respectively, yields a more accurate classifier model as compared to a one-channel classifier model that utilizes ultrasonic data alone. The Applicant has discovered that RF waveforms emitted by electrical infrastructure displays detectable patterns that correlate with partial electrical discharge events, and other events or conditions of interest occurring on electrical transmission infrastructure. Although the relationship between the RF signals and the ultrasonic signals emitted by electrical infrastructure is complex, the Applicant has utilized multi-channel ML models, as described in more detail below, to increase the accuracy of identifying relevant ultrasonic signals by detecting correlating patterns in the RF signals emitted simultaneously by the electrical infrastructure. Thus, in some embodiments of the present disclosure, a more accurate classifier model is obtained for performing classification on the edge computing device.In one aspect, a multi-channel ML model is trained on large sensor signal data sets, where the data has been obtained from monitoring equipment placed on infrastructure in the environment and / or obtained from a lab or testing environment, wherein events or conditions of interest are labelled in the training data sets. The trained multi-channel ML model, or in other words the classifier model, is subsequently deployed on the edge computing devices for continual or frequent sampling and monitoring of the remotely located infrastructure. In some embodiments, the large sensor signal data sets may comprise raw analog or digital signals recorded by the sensors. In some embodiments, the large sensor signal data sets may comprise pre-processed sensor signals, wherein a feature extraction process is applied to the raw sensor signal data in order to extract the relevant features from the recorded signals to generate feature vectors. An example of a feature extraction process includes, but is not limited to, DSP. In still other embodiments, pulse-code modulation (PCM) may be applied to the raw analog signals recorded by the sensors, to produce a digital representation of the analog signals. Thus, for example, the large sensor signal data sets may comprise 240 kHz PCM data captured from ultrasound microphones, either in the field or in a laboratory. It will be appreciated that the above examples of the data that may comprise the large sensor signal data sets, utilized for training the classifier models, are provided as illustrative examples only and are not intended to be limiting.

[0060] In one aspect of the present disclosure it is proposed to use at least two partial discharge sensors on the monitoring device, wherein both partial discharge sensors are configured to simultaneously monitor and record signals that may be generated by an electrical transmission infrastructure. The recorded signals may indicate when a partial discharge event, a corona event, an arcing event, or some other event indicating a breakdown in the infrastructure has occurred or may be developing. In some embodiments, the monitoring device may be equipped with both a radio frequency sensor (RF sensor) and an ultrasound sensor. Electrical discharges generated by the breakdown of insulators on electrical transmission infrastructure produce both ultrasonic signals and radiofrequency signals that are characteristic of such electrical discharges. An example of a suitable ultrasound sensor includes a single microphone, or a plurality of microphones, which have a frequency response in the ultrasonic range of the audio frequency spectrum.

[0061] An RF sensor 240 comprises at least an antenna operatively coupled to signal processing circuitry. The signal processing circuitry is designed to detect and quantify RF signals emitted by electricalinfrastructure, particularly those associated with corona and partial discharge (PD) events. The RF sensor 240 may be implemented in different ways. Three exemplary embodiments of an RF sensor 240 are described below.

[0062] In one embodiment, the RF sensor 240 utilizes a logarithmic envelope detector. This approach extracts the amplitude envelope of a broadband RF signal, providing a direct representation of the energy present across a wide spectrum of frequencies. The signal processing circuitry includes a logarithmic detector, which converts the incoming RF signal's power into an equivalent DC voltage. This logarithmic scaling provides a very wide dynamic range, providing for the measurement of both faint signals from incipient partial discharge events, and strong signals from severe corona discharge events. The single-value output of the logarithmic envelope detector embodiment of an RF sensor 240 is highly energy-efficient and may be particularly well-suited for deployment across a network of monitoring devices 200

[0063] In another embodiment, the RF Sensor 240 may be a software-defined radio (SDR). An SDR digitizes a wide band of the RF spectrum, allowing the application processor 220 to perform sophisticated digital signal processing. This embodiment of an RF sensor 240 would allow for the analysis of a detected signal’s spectral "fingerprint" in order to classify different types of discharges. For example, the spectral fingerprint of a detected RF signal may provide for distinguishing between weather-induced corona and fault-based partial discharges.

[0064] In a further embodiment, the RF sensor 240 may be configured to use the superheterodyne approach. Using an antenna, a transistor, and an amplifier, RF signals are mixed with a local oscillator to produce an intermediate frequency that is subsequently filtered, amplified, and processed. This superheterodyne approach may be best deployed for isolating specific frequency ranges associated with partial discharge phenomena, reducing the interference from noise and unrelated signals, and provides for focused monitoring on narrow frequency slices that may correlate with certain partial discharge events. The superheterodyne approach targets specific frequency bands where partial discharge activity is hypothesized to occur, such as within medium frequency ranges (for example, in the range of 300 kHz - 3 MHz). However, this range is not intended to be limiting, and the RF detection methodology used in the sensor device may be configured to receive signals emitted or caused by a partial discharge event in any other frequencyrange, such as, for example: in the high frequency (HF) range (3 to 30 MHz); in the very high frequency (VHF) range (30 to 300 MHz); and the ultra-high frequency (UHF) range (300 MHz to 3 GHz). The superheterodyne approach may be effective for focusing on pre-defined frequencies that are known to carry modulated signals of partial discharge events.

[0065] For any of these embodiments of an RF sensor 240, the sensor may be designed to interface with external antennas through standard antenna hardware connectors, like an SMA connector. For detecting partial discharge events on transmission infrastructure, some embodiments may include an antenna tuned for high sensitivity in the Very High Frequency (VHF) range (30 MHz to 300 MHz), such as a monopole or helical monopole antenna.

[0066] The data gathered from any such RF sensor configuration may be used for a comparative analysis, wherein a plurality of monitoring devices 200 are deployed across a network of infrastructure assets and report to a central server. The central server aggregates data from the plurality of monitoring devices 200 to facilitate long-term trend analysis and enhance system-wide insights. By monitoring of each asset over an extended period of time, a baseline for normal RF and ultrasonic activity may be established across assets with similar characteristics and environmental conditions. This baseline allows the system to more accurately identify statistically significant deviations or gradual changes in the signal data samples of a given electrical asset, thereby improving confidence in the early detection of incipient faults.

[0067] In one aspect, the Applicant has found that the superheterodyne RF method may be most suitable when the specific frequency bands where partial discharge (PD) activity is likely to occur have been previously identified. If the specific frequency bands are well-defined and predictable, the superheterodyne RF method may effectively isolate and amplify relevant RF signals. However, if infrastructure variations (e.g., different conductor types, insulators, or substation components) or environmental condition variations (e.g., humidity, EMI, or nearby electrical noise sources) cause partial discharge signals to shift in frequency or behave unpredictably, the superheterodyne method may not consistently detect and classify partial discharge events. The envelope detection approach may be comparatively less computationally intensive, as compared to the superheterodyne approach, thus potentially allowing for continuous monitoring while conserving energy. Furthermore, envelope detection methodology may adapt to a broad range of operational scenarioswithout requiring precise tuning or adjustments for specific frequencies. On the other hand, envelope detection may lack the frequency specificity of the superheterodyne detection method, which may limit the ability of the envelope detection method to isolate certain partial discharge events if precise frequency information is required.

[0068] In some embodiments, the RF sensor 240 may be configured to utilize two or more of the approaches to RF signal detection and processing that are described above. For example, without intending to be limiting, the RF sensor 240 may be configured to utilize both the envelope detection method and the SDR method. Such a combined configuration of an RF sensor 240 may thereby provide both the energy-efficient advantages of the envelope detection method, and the ability to distinguish between different types of discharge events through analysing the spectral fingerprint of the RF signal, as provided by the SDR method.

[0069] By sampling, recording and analyzing both ultrasonic and RF signals simultaneously, both types of signals may be provided as inputs into a multi-channel ML algorithm, which may produce a more robust method of differentiating relevant ultrasound data from non-relevant ultrasound data. Generally speaking, the relevant ultrasound data is generated from electromagnetic events (ie: partial electrical discharges) that occur in the air and in close proximity to the electrical transmission infrastructure. The same electromagnetic events also emit RF signals which are characteristic of the electromagnetic events, and different from the RF signals that may be emitted by other, non-relevant events, such as the different environmental variables that are described above (including but not limited to birds and other wildlife, weather events, lightning strikes, mechanical vibration, etc.)

[0070] The classifier model may be configured in different ways. In some embodiments, with reference to FIG. 5, a single, multi-channel ML model 480 may ingest the data from the two or more input channels, each channel characterized by one of the two or more sensor signal data streams 410, 415, 416, and analyse the data from the two or more sensor streams simultaneously in the neural network of the ML model 480.

[0071] The sensor signal data streams 410, 415 (and optional signal data stream 416) may firstly go through a pre-processing step, whereby respective digital signal processor (DSP) modules 420, 425 perform feature extraction on each sensor signal data stream 410, 415, 416 in which case it isthe extracted features that are provided by the application processor to the single, multi-channel ML model 480 for classification. A single ML model 480 may be a preferred embodiment, as both sensor signal data streams 410, 415 are analyzed simultaneously by a single model. A convolutional neural network (CNN) may make intuitive predictions based on what it considers to be meaningful patterns detected within both the sensor signal data streams 410, 415, 416. As an example of the two sensor signal data streams, not intended to be limiting, a first sensor signal data stream 410 may be signals recorded by an RF sensor, and a second sensor signal data stream 415 may be signals recorded by an ultrasound microphone recording sound signals in the ultrasonic range. Optionally, as shown in FIG. 5, one or more additional sensor signal data streams 416, such as recorded by one or more additional sensors, may also be processed by a DSP module, such as DSP module 426. The one or more additional sensors may be any of the sensors described herein, configured for detecting signals emitted by electrical infrastructure when a partial discharge event or other event occurs, indicating the breakdown of an insulator or the impending breakdown of an insulator.

[0072] As shown in FIG. 5A, in some embodiments an additional pre-processing step may be employed, such as a feature-level fusion module 427. As previously described, each DSP module 420, 425, 426 performs feature extraction on the raw recorded data of each respective data stream, such as the ultrasonic recording data stream 410, the RF recording data stream 415 and the Nth sensor data stream 416. The feature extraction step may include processes such as filtering, signal conditioning and spectral analysis to produce a feature vector for its respective data stream. The extracted feature vectors obtained by each of the DSP modules 420, 425 and 426 are then fed into a feature-level fusion module 427. The function of module 427 is to combine the multiple feature vectors into a single, enriched, and more comprehensive feature vector. To effectively combine the feature vectors extracted from each of the raw data streams, advanced spatio-temporal methods may be employed to model dependencies between the sensor streams. Examples of such methods include, but are not limited to, Bayesian inference or Spatio-Temporal Kalman Filtering. The resulting combined feature vector obtained from module 427, which contains correlated information from each of the sensor raw data streams 410, 415 and optionally 416, is then provided as the input to a single-channel ML model 480 for classification.Advantageously, embodiments comprising a feature-level fusion module 427 may allow the ML model to make more accurate and robust classifications by leveraging the combined insights from a scalable number of diverse sensor inputs. It will be appreciated that the embodiments incorporating a feature-level fusion module 427 are scalable, in that this architecture provides for analyzing and correlating the raw data streams obtained from two or more sensors, with the additional sensors represented by the Nth sensor data stream 416, which represents the data stream obtained by one or more additional sensors, beyond the two primary partial discharge sensors 240 and 250 described herein. Likewise, each additional Nth sensor producing an additional Nth data stream has a correlated Nth DSP module 426, for performing the pre-processing step of extracting the relevant features from the raw data stream 416. This same convention of referring to additional sensors, sensor data streams, DSP modules and ML models as the “Nth” sensor, sensor data stream, DSP module and / or ML model is used throughout the present disclosure.

[0073] In some embodiments, with reference to FIG. 6, the classifier model may be configured as two or more independent, single-channel models 430, 435, wherein each independent ML model ingests the data from one input channel. Each input channel is characterized by one of the two (or more) sensor signal data streams 410, 415. Each of the sensor signal data streams 410, 415 may firstly go through a pre-processing step, whereby respective DSP modules 420, 425 perform feature extraction on its corresponding sensor signal data stream 410, 415, in which case it is the extracted features that are provided by the application processor to the corresponding, independent ML model 430 or 435 for classification.

[0074] For example, not intended to be limiting, an ultrasonic microphone may generate an ultrasonic sensor signal data stream 410, which is pre-processed through the DSP module 420 to extract the relevant features, and then the generated feature vector is input into the independent audio ML model 430 for classification. Likewise, an RF sensor may generate an RF sensor signal data stream 415, which is pre-processed through the DSP module 425 to extract the relevant features, and then the generated feature vector is input into the independent RF ML model 435 for classification. Optionally, if additional sensors are included in the monitoring device 200, each additional sensor will generate a data stream, represented in FIG. 6 as the Nth sensor data stream 416; in that case, the Nth sensor data stream 416 undergoes pre-processing by a corresponding Nth DSP module 426, and is analysed by a separate Nth sensor ML model 436. This approach to ML modelling forseparate data streams is otherwise referred to as Decision-Level Fusion, wherein each independent ML model produces an output and wherein a final, single output is then determined by combining these individual outputs from each ML model, for example through averaging or via a weighted-sum rule.

[0075] As a further example, not intended to be limiting, a first sensor signal data stream 410 (in either FIGS. 5 or 6) may be signals recorded by a High-Frequency Current Transformer (HFCT) and the second sensor signal data stream 415 may be signals recorded by an ultrasound microphone. The two sensor signal data streams 410, 415 are recorded simultaneously, thus providing additional data generated by a single event of interest that has occurred on the electrical transmission infrastructure.

[0076] Although not shown in the Figures, a potential third sensor, or an alternative sensor to the RF sensor, may be an HFCT for partial discharge detection, offering direct monitoring of high-frequency current pulses through a ground conductor. For embodiments implementing an HFCT sensor, if the partial discharge signal is below the noise floor of the HFCT's detection capability, advanced signal processing techniques, such as wavelet analysis, may be used to isolate and identify transient partial discharge events. Conversely, if the HFCT proves overly sensitive, such as picking up partial discharge signals from nearby poles or other electrical infrastructure components that are not being monitored by the specific monitoring device 200, then it may be necessary to apply attenuation or to correlate the HFCT data with other sensors on the monitoring device, such as RF or ultrasonic sensors, to accurately localize partial discharge events. Additionally, the HFCT’s output may be sampled at a high enough frequency to process it through similar DSP methods, as described elsewhere in this application with respect to pre-processing performed on the raw data obtained by the ultrasound and RF sensors, and feed the HFCT data (specifically, the features extracted from that data by the DSPs) to the ML model(s), which may provide for deeper analysis and classification of partial discharge events. While the HFCT requires a ground conductor for operation, installing these ground conductors on poles that do not have an existing ground conductor may also enhance the monitoring device’s resistance to electrostatic discharge and surges, improving overall reliability and standardizing the infrastructure for longterm monitoring.Referring again to FIG. 3, one possible implementation of the HFCT sensor (not shown in FIG. 3) is for the HFCT to be in direct communication with the application processor 220, as is the case for the RF sensor 240 and the ultrasonic microphone sensor 250 shown in FIG. 3. Alternatively, if the HFCT sensor is implemented as an arc detection sensor 290, the application processor 220 may be configured to prompt the low-power co-processor 280 to increase its sampling rate and transfer the data directly to the application processor’s memory 230 via a serial communication interface. Although two partial discharge sensors 240, 250 are described in this present disclosure, it will be appreciated that a possible alternative implementation may include utilizing an HFCT sensor to be a third partial discharge sensor, in addition to the RF sensor 240 and the ultrasonic microphone 250. In such embodiments, there would be three sensor signal data streams being recorded, simultaneously, by each of the three partial discharge sensors, and the three sensor signal data streams would undergo pre-processing by a DSP to extract the features, prior to being fed into a single, three-channel ML model (similar to the Nth-channel ML model embodiment illustrated in FIG. 6). Alternatively, in some embodiments, after each sensor signal data stream from the three partial discharge sensors undergoes pre-processing by a DSP to extract the features, the extracted features would each be fed into a dedicated ML model (similar to the Nth-channel ML model illustrated in FIGS. 5 and 5A).

[0077] In some embodiments, referring to FIGS. 2 and 3, the monitoring device 200 may include an environmental sensor 270. An environmental sensor 270 may record one or more characteristics of the surrounding ambient environmental conditions, including but not limited to the surrounding ambient temperature, air pressure and humidity. Such environmental conditions or characteristics may additionally be used to validate the classification result of the multi-channel classifier model. The real-time environmental condition data measured by the environmental sensor 270 may be transmitted alongside the ML classification data over a wireless transmission to the centralized server system and may be used to qualify the classification of partial discharge events, based on environmental / ambient conditions obtained during the audio and RF sampling recordings. For example, not intended to be limiting, such ambient environmental condition data may be used to distinguish a lightning strike event from a partial discharge event, on the basis that the ambient environmental conditions strongly indicate that a lightning strike event has occurred. Additionally, environmental sensors may be utilized to detect localized conditions surrounding the support pole or other structure onto which the monitoring device is mounted, providing valuable monitoringinformation about the condition of the physical structure. For example, environmental sensors may detect a sudden increase in temperature beyond normal values, indicating the presence of a fire on or near the support structure.

[0078] Environmental conditions, such as temperature, humidity, and other atmospheric factors, may provide valuable contextual information to aid in the classification of partial discharge events detected by the at least two partial discharge sensors 240, 250 of the monitoring device 200. For instance, high humidity levels may promote surface discharge by creating conductive paths on insulator surfaces, whereas lower humidity conditions may correlate more closely with internal partial discharge within insulating materials. Temperature extremes, whether high or low, can also influence partial discharge activity by degrading insulation strength, inducing mechanical stress, or causing surface condensation in cold, humid conditions. While the exact relationships between these environmental parameters and partial discharge phenomena may require further validation, advantageously, monitoring these environmental parameters allows the system to enhance its ability to differentiate between discharge types and potential external influences. By combining environmental data with RF and ultrasonic sensor outputs, the system may achieve more robust and accurate event classification while adapting to varying operational conditions.

[0079] In the embodiment of an edge computing device 460 illustrated in FIG. 7, the classification step is performed by the independent classifier models 430, 435 (and optionally, 436) on the edge computing device 460 (otherwise referred to herein as the monitoring device). In such embodiments, because each model 430, 435 (and optionally, 436) is independently trained to generate classification results for the same partial discharge events contained in the training sensor signal data sets, it may be difficult to determine which model is more accurate in detecting any particular instance of a partial discharge event, which complicates weighting and decision making to be performed by the classifier models 430, 435, 436 as performed on the edge computing device. In one aspect, the system consolidates ML classifications generated by the classifier models 430, 435, 436 which may be packaged and transmitted wirelessly to a remote cloud database for further analysis, storage, or long-term trend evaluation.

[0080] Referring now to FIG. 8, rather than deploying the classifier model (or models) on the edge computing device 460, the classifier model (or models) may be deployed on a remote central server(or plurality of central servers) 470, including but not limited to deployment on a cloud network 470. In such embodiments, the sensors are located on the monitoring device (also referred to herein as the edge computing device 460). Each sensor, including at least the two partial discharge sensors such as, for example, an ultrasonic microphone and an RF sensor, acquires respective ultrasound and RF sensor signal data streams 410, 415 (and optionally, additional partial discharge sensor data streams represented by the Nth sensor data stream 416). In the illustrated embodiment, the acquired sensor signal data streams 410, 415, 416 undergo local pre-processing, on the edge computing device 460, by dedicated DSP modules 420, 425, 426. The DSP modules 420, 425, 426 perform feature extraction to reduce data complexity and optimize it for transmission. Feature extraction may include, for example, spectral analysis, filtering, and / or signal conditioning to identify key characteristics of the input signals. The extracted feature vectors generated by the DSP modules 420, 425, 426 are then transmitted wirelessly to the cloud network or other central server 470 for further analysis, which may facilitate the efficient use of bandwidth and device power.

[0081] In the remote central server or cloud network 470, as shown in FIG. 8, the extracted and transmitted features are processed by a plurality of corresponding ML models 430, 435, 436 hosted by the cloud network 470. In the example embodiment illustrated in FIG. 8, the plurality of corresponding ML models may be configured as, for example, an individual audio ML model 430, which processes extracted features from the ultrasonic signals, and the RF ML model 435, which processes extracted features derived from RF signals, in order to classify potential electrical phenomena or system conditions, and / or to detect events such as partial discharge, arcing, or environmental anomalies. If there are additional sensors and data streams, then there may be additional individual ML models for each additional sensor and data stream, represented by the Nth sensor ML model 436.

[0082] Although not shown in FIGS. 6 to 8, it will be appreciated that in each of these illustrated embodiments, the individual ML models 430, 435 and 436 may, instead, be configured as one multi-channel ML model 480, such as shown in FIGS. 5 and 5 A. Additionally, although not shown in FIGS. 6 to 8, it will be appreciated that the additional pre-processing step of generating a combined feature vector, performed by the feature-level fusion module 427 (as shown in FIG. 5 A), may be incorporated into any of the embodiments illustrated in FIGS. 6 to 8.In some embodiments, such as illustrated in FIG. 9, it will be appreciated that a plurality 610 of monitoring devices 200 may be deployed on a plurality of infrastructure assets within an electrical transmission network. For example, the plurality of monitoring devices 200 may be mounted to a plurality of transmission line poles 620 or other electrical transmission support structure. Each monitoring device 200 may be equipped to capture data related to partial discharge events. The partial discharge classifications are wirelessly transmitted, via the wireless communication interface 260 of the device 200, to a cloud server 470 for centralized analysis, storage, and longterm trend evaluation.

[0083] In another aspect of the present disclosure, such as illustrated for example in FIGS. 8 and 9, the extracted features of the RF and ultrasound data (for example) may be wirelessly transmitted, via the wireless communication interface 260 of the device 200, to a cloud server 470 for centralized analysis, storage, and long-term trend evaluation. Deploying a plurality of monitoring devices 200 across an electrical transmission network, provides for the distributed monitoring system that may detect and analyze partial discharge events in real-time or near-real time, with the cloud server 470 aggregating data from multiple devices 200 to enhance system-wide insights and reliability. The central server, in performing an analysis on the aggregated data obtained for a plurality of devices 200 across an electrical infrastructure network, plays a role in transforming the data collected by the plurality of monitoring devices into actionable intelligence for managing the infrastructure network. By aggregating data from multiple devices, the server facilitates a comprehensive, system-wide comparative analysis and long-term trend analysis that cannot be obtained at the individual device level. This centralized processing provides valuable functions, including the analysis of historical data to establish operational baselines for different asset types under various environmental conditions; the use of trend projections for predictive maintenance scheduling; and the correlation of sensor data across the network to facilitate root cause analysis of detected faults.

[0084] Furthermore, this aggregated dataset, which may include correlated environmental data streams and / or tilt data streams, provides for a holistic analysis of the behavior of partial discharge in relation to seasonal cycles and geographical factors. This holistic analysis allows for the continuous retraining and refinement of the ML models, which may then be dynamically allocated and deployed back to the devices 200 located, for example, in specific climates, via over-the-air updates. Ultimately, this creates a powerful feedback loop where the aggregated historical datafrom the network continually enhances the predictive accuracy and contextual awareness of each device, allowing the system to move beyond simple event detection to provide a holistic view of network health, risk assessment, and maintenance prioritization

[0085] In another aspect, the monitoring devices 200 may be configured to receive remote, over-the-air (OTA) firmware and software updates. This capability allows for the enhancement of device functionality, security and performance throughout its operational lifecycle without requiring costly and impractical physical service visits. Implementing OTA firmware and software update functionality allows for the network of monitoring devices to remain effective, secure, and adaptable to new insights and operational requirements over long-term deployments. Examples of firmware and software updates that may be performed include, but are not limited to, the following:

[0086] a) Modifying or replacing the digital signal processing (DSP) algorithms to enhance or change the methods of feature extraction;

[0087] b) Deploying new ML model weights or entirely new model architectures as they are refined on larger datasets;

[0088] c) Adjusting data sampling schedules, power management schemes, and trigger thresholds for various sensors;

[0089] d) Deploying critical security patches to address new vulnerabilities, along with general bug fixes to improve device reliability;

[0090] e) Remotely updating sensor calibration data to account for hardware drift and changing environmental conditions.

[0091] In another aspect of the present disclosure, features in ultrasound data sampled by the monitoring device that indicated discharge events are distributed widely across a range of ultrasound and high audible frequencies. Different types of electrical discharges may emit different audio signals that are characteristic of those electrical discharge types. As such, there is no single “tone” or set of discrete tones that are characteristic of all discharge events. Factors that may influence the audio signals emitted by a particular electrical discharge event include, but are not limited to: variation in the insulator materials, insulator design, and varying conditions in the environment surrounding the electrical infrastructure being monitored. However, advantageously, the Applicant has discovered that partial discharge events are typically accompanied by an RF discharge that isdetectable above the ambient “noise floor” of the RF domain. Thus, by sampling ultrasound data and RF data simultaneously at a given monitoring site, and utilizing ML models to identify correlations between the sampled ultrasound data and the sampled RF data, the Applicant has developed methods for increasing the accuracy of identifying and detecting patterns in the sampled data that is produced by an electrical discharge event, as compared to only sampling and analyzing ultrasound data alone.

[0092] The relationship between ultrasonic signals and radio frequency signals generated by partial electrical discharge events is non-linear, because a change in amplitude in a recorded ultrasound signal does not correspond to an equivalent change in amplitude of the simultaneously recorded RF signal. Additionally, different frequencies in the ultrasound signal correspond differently to their analogous RF signals. Although there is a correlation between the ultrasound and RF signals emitted by electrical discharge events occurring on electrical infrastructure, these signals do not mirror each other and the correlation between these signals is complex. Given that the audio and RF signals emitted during an electrical discharge event are imperfect analogs, it is impractical to use a simple algorithm that requires direct concordance between the two signals to validate a sample. Thus, in one aspect, the Applicants propose the use of a multi-channel ML model that compares the two or more signals interpretively, looking for coincidence rather than direct correlation between the multiple data streams. Although the true relationship between the features of the audio and RF signals emitted when an electrical discharge event occurs is unknown, the comparison may be stochastically reliable.

[0093] In some embodiments, the audio signals are filtered and processed, utilizing the ML models, to differentiate between different types of partial electrical discharges. The RF waveform signals, sampled simultaneously with the audio signals, are utilized to validate the classification of the partial electrical discharge events, which classification is initially assigned based on the processing and analysis of the audio signals. However, in other embodiments, the RF waveform signals may be initially utilized to perform the classification of the different types of partial electrical discharge events, and the audio signals may be utilized to validate the classifications that are initially assigned based on processing and analysis of the RF waveform signals. In still other embodiments which utilize two independent ML models, the classification of the different types of partial electrical discharge events may be independently performed based on the audio and RF waveformdata streams respectively. Then, the independent RF and audio ML models are utilized to crossvalidate the classification as between the two independent ML models. The approach of cross-validation of the classification performed by each of the independent RF and audio ML models may advantageously increase the overall accuracy of the classification, taking into account that each of the audio and RF data streams are subject to different types and sources of interference.

[0094] In some embodiments, the classifier models may be implemented using a Convolutional Neural Network (CNN). In other embodiments, the classifier models will be implemented using a decision tree model. The main advantage of a decision tree model is lower computational costs and faster model training. However, one drawback of a decision tree model is that it tends to “overfit” on training data sets, meaning that the resulting model may not generalize well when a data set to be analyzed falls outside of the data sets the model was trained on. In addition, a decision tree approach requires the engineering of “features” extracted from the raw data that are determined to be important by the data science team, which may be difficult to identify in scenarios where the data science team is unable to accurately identify “features” that are relevant

[0095] In this respect, the relative advantage of a CNN approach to the multi-channel ML model is that it requires less preparation of the data that the model is being trained on. Advantageously, the CNN “discovers” features in the data that are relevant to the classification. A CNN model provides confidence intervals and therefore there is a reduced need for overfitting. Whereas, a decision tree approach is more deterministic, with an increased likelihood of overfitting. In one aspect of the present disclosure, a CNN architecture for a multi-channel ML model is preferred, although, a decision tree architecture may also work and is intended to be included in the scope of the present disclosure.

[0096] In one aspect of the present disclosure, monitoring of each asset over an extended period of time, on the order of several months or years, may allow for prioritization of the generated maintenance recommendations, taking factors into account such as the severity or criticality of a detected condition of an asset, the location of the asset, the proximity of a number of assets to one another requiring maintenance, and the scope of the potential impact on customers should the asset fail.

[0097] Advantageously, valuable information may be derived from the periodic sampling of each electrical asset, as gradual or sudden changes in the audio and RF signal data samples of a givenelectrical asset may indicate changes in the condition of that asset. As well, in the case of a catastrophic failure of a given asset, such events may be detected by the absence of classification results of the model transmitted to the centrally located server, as under normal operating conditions, the monitoring device would have transmitted the classification results at the predetermined time intervals, which for example may occur every minute, to every few hours.

[0098] In some embodiments of the present disclosure, the audio and RF signals, or other types of signals, emitted by an asset may be recorded by the monitoring device at pre-determined intervals, such as, taking a one second recording of the sound signals emitted by the asset every five minutes; it will be appreciated that this is only one example of a sampling scheme, and that other sampling schemes may be utilized in accordance with the present disclosure. For example, some sampling schemes may be configured to record audio and RF signals every minute, every few minutes, every few hours, or every 12 hours or 24 hours. Sample lengths may be a one second recording, or the sample length may be shorter than one second, may be several seconds, or may be several minutes, depending on the memory, power and computing resources and constraints applicable to a particular monitoring device deployment. Frequently recording the audio and RF signals, for a short duration, of many assets over a period of months or years may generate a sufficiently large data set that may be used to continually or frequently train and refine the classifier models, resulting in improved classification accuracy over time. This sampling approach may be appropriate for systems in which the edge computing devices face power constraints; for example, where each edge computing device is powered by a battery and a small solar panel. However, in systems without such power constraints, such as edge computing devices that are powered by power harvesting devices capable of harvesting electricity from the electrical transmission lines, sampling the audio and RF signals with increased frequency and recording duration may be desirable for obtaining larger data sets, as well as increasing the likelihood that an anomaly in the audio and RF signals, emitted by the infrastructure assets, will be captured by the monitoring device. In systems where the edge computing device has no power constraints, continuous sampling and monitoring of the infrastructure assets may be provided.

[0099] In some embodiments, large sensor signal data sets may be collected by physically attending to the monitoring devices and downloading higher-fidelity, raw data recordings stored on a memory of the monitoring device, such as via a Bluetooth or other wireless data connection between a datacollection device and the monitoring device. In other embodiments, large sensor signal data sets may be wirelessly transmitted to a central server location, which process may involve compressing such recorded data at the edge computing device prior to wireless transmission. From time to time, as classifier models are further refined, trained and / or updated, the classifier models residing on the individual monitoring devices may be updated via firmware over-the-air updates, as will be known to persons skilled in the art.

[0100] As more fully described in the Applicant’s International Patent Publication No. WO 2023235981 Al (the entirety of which is incorporated herein by reference), an autoencoder Al model (also referred to herein as the "autoencoder") comprises two components; the first component is the encoder that maps the raw input data to a latent variable space, and then compresses the data into a binary representation; and the second component is the decoder that decompresses the binary representation, and then maps the latent representation into a reconstruction of the input data. The encoder, which may be deployed on the edge computing device, may be used to compress the original sensor data obtained by the RF sensor 240 and the ultrasound sensor 250 to a size that meets the network constraints, and the compressed data may then be delivered to a central server (or other computing device or system) located remotely from the electrical assets, where computing and power resources are less constrained. At the central server, the binary representation of each sensor data stream may be decoded by the decoder component back into a representation of the original input data. The representation of the original input data may then be run through the classifier model on the central server to classify the latent variable space and identify any faults, defects or other conditions of the asset, as indicated in the original sensor data.

[0101] In some embodiments, utilizing an autoencoder may be useful where there is limited computing or power resources available at the edge computing device 200. Additionally, such an embodiment may be useful for collecting some or all of the originally recorded audio and RF signal data recorded by the sensors 240, 250, which data may be utilized to train, update and / or improve the classifier models. Alternatively, in some embodiments the strategy of utilizing an autoencoder to compress and send the originally recorded sensor data from the edge computing devices to a central server, for the purpose of updating the classifier models and / or for validating the classification of an event, may in some cases be performed in tandem with performing the classifier analysis at the edge computing device.Although the present disclosure describes monitoring remote electrical infrastructure using the methods and systems disclosed herein, the Applicant notes that the disclosed methods and systems may be applied to monitoring other types of infrastructure assets which emit signals detectable by the monitoring device 200, and these data streams may be analysed to detect defects or faults developing in the infrastructure asset being monitored. For example, not intended to be limiting, such methods and systems may be applied to monitoring the status of electrical distribution infrastructure, electrical generation infrastructure, electrical substation infrastructure, transportation infrastructure such as tunnels, roadways, overpasses, bridges and roadway lighting systems, traffic control systems, pipeline infrastructure, oil and gas infrastructure and railway systems.

[0102] Where the infrastructure network being monitored is electrical infrastructure, the infrastructure assets monitored by the plurality of monitoring devices may include, but is not limited to: switchgears, insulators, conductors, sub-conductors, and support structures including poles and towers. The types of conditions of the infrastructure assets that may be detected by the systems and methods disclosed herein include, but are not limited to: physical failures, corona discharges, partial discharges, arcing, flashover, electrical tracking (ie: dirt or carbon buildup on infrastructure assets that creates a path through which electrical current leakage may occur), pollution, and operational baseline conditions.

[0103] In some embodiments, the monitoring device and the edge computing device are contained within the housing of a single unit; however, it will be appreciated by a person skilled in the art that the monitoring device (comprising the sensors and an application processor) and the edge computing device (comprising a second application processor and a wireless communication interface) may be separate components located proximate to one another and that this is not intended to be limiting.

Claims

WHAT IS CLAIMED IS:

1. A system for detecting a condition of an infrastructure, the infrastructure comprising a plurality of infrastructure assets, the system comprising:a plurality of monitoring devices, each monitoring device of the plurality of monitoring devices mounted proximate to an infrastructure asset of the plurality of infrastructure assets,a central server located remotely from the infrastructure, the central server in wireless communication with each monitoring device of the plurality of monitoring devices,each monitoring device of the plurality of monitoring devices comprising: first and second partial discharge sensors, a wireless communication interface and a memory, each of the first and second partial discharge sensors, the wireless communication interface and the memory controlled by an application processor,wherein the first partial discharge sensor is configured to detect a first signal emitted by the infrastructure asset to generate a first data stream and the second partial discharge sensor is configured to simultaneously detect a second signal emitted by the infrastructure asset to generate a second data stream, the first and second data streams recorded to the memory, andwherein the application processor performs pre-processing on each of the first and second data streams via at least one digital signal processing (DSP) module, wherein the at least one DSP module is configured to perform feature extraction on the first and second data streams to generate first and second feature vectors, andwherein at least one classifier model analyzes the first and second feature vectors to generate a classification, the classification indicating a condition of the infrastructure asset.

2. The system of claim 1 wherein the first partial discharge sensor is an ultrasonic microphone and the first signal is an audio signal in the ultrasonic range.

3. The system of claim 2 wherein the second sensor is a radiofrequency (RF) sensor and the second signal is an RF signal.

4. The system of claim 3 wherein the RF sensor is selected from one or more of the group comprising: a superheterodyne RF sensor, an SDR, a logarithmic envelope detector.

5. The system of claim 1 wherein each monitoring device further comprises an environmental sensor, the environmental sensor configured to measure at least one characteristic of an environment surrounding the infrastructure asset to generate an environmental data stream, wherein the environmental data stream is analyzed by the central server or the application processor to validate the generated classification.

6. The system of claim 5 wherein the at least one characteristic of the environment is selected from a group comprising: temperature, air pressure, humidity.

7. The system of claim 1 wherein the at least one DSP module comprises first and second DSP modules, and wherein the pre-processing performed on each of the at least first and second data streams by the application processor includes the first DSP module performing feature extraction on the first data stream to generate the first feature vector and the second DSP module performing feature extraction on the second data stream to generate the second feature vector.

8. The system of claim 7 wherein the pre-processing performed on each of the at least first and second data streams by the application processor further includes a feature-level fusion module configured to combine the first and second feature vectors generated by the first and second DSP modules to generate a combined feature vector.

9. The system of claim 1 wherein the at least one classifier model resides on the application processor and wherein the generated classification is transmitted via the wireless communication interface of the monitoring device to the central server over a wireless network.

10. The system of claim 1 wherein the at least one classifier model resides on the central server and wherein the first and second feature vectors generated by the at least one DSP module are transmitted, via the wireless communication interface of the monitoring device, to the central server over a wireless network, and wherein the analysis of the extracted features is performed by the at least one classifier model, on the central server, to generate the classification.

11. The system of claim 1 wherein the at least one classifier model of each monitoring device is configured as one multi-channel ML model for ingesting the at least first and second feature vectors extracted from the first and second data streams to generate the classification.

12. The system of claim 8 wherein the at least one classifier model of each monitoring device is configured as one single-channel ML model for ingesting the combined feature vector generated by the feature-level fusion module to generate the classification.

13. The system of claim 1 wherein the at least one classifier model is configured as a first ML model for ingesting the first feature vector to generate a first classification and a second ML model for ingesting the second feature vector to generate a second classification, and wherein the first classification indicating the condition of the infrastructure asset is validated by the second classification.

14. The system of claim 1 wherein each monitoring device further comprises a low-power coprocessor in communication with the application processor and at least one low-powercontinuous sensor, wherein the at least one low-power continuous sensor continuously monitors for a transient signal exceeding a predetermined threshold value, andwherein, when the transient signal detected by the at least one low-power continuous sensor exceeds the predetermined threshold value, the low-power co-processor signals the application processor to initiate a data processing operation and a logging operation, the logging operation including at least initiating a simultaneous recording of the first and second data streams obtained by the first and second partial discharge sensors, and the data processing operation including the at least one DSP module generating first and second feature vectors from the first and second data streams.

15. The system of claim 1 wherein each monitoring device further comprises at least one low- power continuous sensor in communication with the application processor, the at least one low-power continuous sensor configured to continuously monitor for a transient signal exceeding a predetermined threshold value, and wherein the application processor is a multi-core processor, andwherein, when the transient signal received by the at least one low-power continuous sensor exceeds the predetermined threshold value, the multi-core application processor initiates a data processing operation and a logging operation, the logging operation including at least initiating a simultaneous recording of the first and second data streams obtained by the first and second partial discharge sensors, and the data processing operation including the at least one DSP module generating first and second feature vectors from the first and second data streams.

16. The system of any one of claims 14 and 15, wherein the at least one low-power continuous sensor is selected from a group comprising: arc detection sensor, high-frequency current transformer, audio sensor, RF sensor, optical sensor, thermal sensor, vibration sensor, environmental sensor.

17. The system of any one of claims 1 to 15, wherein each monitoring device further comprises a tilt sensor in communication with the application processor, the tilt sensor configured to periodically detect an orientation of the corresponding monitoring device relative to a ground to generate a tilt data stream, the tilt data stream transmitted wirelessly to the central server for monitoring movement of the respective infrastructure asset over time.

18. The system of any one of claims 1 to 17 wherein each infrastructure asset is selected from a group of electrical infrastructure assets, the group comprising: switch-gear, insulator, conductor, sub-conductor, pole, tower.

19. The system of claim 18 wherein the condition of the infrastructure asset is selected from a group comprising: physical failure, corona discharge, partial discharge, arcing, flashover, tracking, pollution, operational baseline condition.

20. The system of any one of claims 1 to 17 wherein the infrastructure is selected from a group comprising: electrical transmission infrastructure, electrical distribution infrastructure, electrical generation infrastructure, electrical substation infrastructure, transportation infrastructure, tunnels, roadways, overpasses, bridges, roadway lighting systems, traffic control systems, pipeline infrastructure, oil and gas infrastructure, railway systems.

21. The system of any one of claims 1 to 20 wherein the central server is configured to receive at least a plurality of classifications generated by the plurality of monitoring devices to generate an aggregated data set, and wherein the central server analyses the aggregated data set to generate a recommendation for maintenance of each infrastructure asset of the plurality of infrastructure assets so as to generate a plurality of recommendations for maintenance of the plurality of assets, wherein each recommendation of the plurality of recommendations is prioritized based on an assessed criticality of the generated condition of each asset of the plurality of assets.

22. A monitoring device for detecting a condition of an infrastructure asset, the monitoring device comprising:at least first and second partial discharge sensors, a wireless communication interface and a memory, each of the first and second partial discharge sensors, the wireless communication interface and the memory controlled by an application processor,wherein the first partial discharge sensor is configured to detect a first signal emitted by the infrastructure asset to generate a first data stream and the second partial discharge sensor is configured to simultaneously detect a second signal emitted by the infrastructure asset to generate a second data stream, the first and second data streams recorded to the memory, andwherein the application processor performs pre-processing on each of the first and second data streams via at least one digital signal processing (DSP) module, wherein the at least one DSP module is configured to perform feature extraction on the first and second data streams to generate first and second feature vectors, andwherein at least one classifier model residing on the application processor analyzes the first and second feature vectors to generate a classification, the classification indicating a condition of the infrastructure asset.

23. A method for monitoring a condition of an infrastructure network comprising a plurality of infrastructure assets, the method comprising:receiving, on a central server in wireless communication with the plurality of monitoring devices of claim 1 deployed on the plurality of infrastructure assets, at least a plurality of classifications generated by the plurality of monitoring devices to generate an aggregated data set,performing a comparative data analysis on the aggregated dataset to identify at least one trend indicating the performance or condition of one or more assets of the plurality of infrastructure assets.

24. The method of claim 23 wherein each device of the plurality of monitoring devices includes an environmental sensor that generates an environmental data stream, the environmental data stream received by the central server and included in the aggregated data set.

25. The method of any one of claims 23 and 24 wherein each device of the plurality of monitoring devices includes a tilt sensor that generates a tilt data stream, the tilt data stream received by the central server and included in the aggregated data set.

26. The method of any one of claims 23 to 25 wherein each monitoring device further comprises at least one low-power continuous sensor in communication with the application processor, wherein the at least one low-power continuous sensor continuously monitors for a transient signal exceeding a predetermined threshold value, andwherein, when the transient signal detected by the at least one low-power continuous sensor exceeds the predetermined threshold value, the application processor initiates a data processing operation and a logging operation, the logging operation including at least initiating a simultaneous recording of the first and second data streams obtained by the first and second partial discharge sensors, and the data processing operation including the at least one DSP module generating first and second feature vectors from the first and second data streams.

27. The method of claim 26 wherein the at least one low-power continuous sensor is selected from a group comprising: arc detection sensor, high-frequency current transformer, audio sensor, RF sensor, optical sensor, thermal sensor, vibration sensor, environmental sensor.

28. The method of any one of claims 23 to 27 wherein each infrastructure asset is selected from a group of electrical infrastructure assets, the group comprising: switch-gear, insulator, conductor, sub-conductor, pole, tower.

29. The method of any one of claims 23 to 28 wherein the condition of the infrastructure asset is selected from a group comprising: physical failure, corona discharge, partial discharge, arcing, flashover, tracking, pollution, operational baseline condition.

30. The method of any one of claims 23 to 27 wherein the infrastructure is selected from a group comprising: electrical transmission infrastructure, electrical distribution infrastructure, electrical generation infrastructure, electrical substation infrastructure, transportation infrastructure, tunnels, roadways, overpasses, bridges, roadway lighting systems, traffic control systems, pipeline infrastructure, oil and gas infrastructure, railway systems.

31. The method of any one of claims 23 to 30 wherein the central server analyses the aggregated data set to generate a plurality of recommendations for maintenance of the plurality of assets, wherein each recommendation of the plurality of recommendations is prioritized based on an assessed criticality of the generated condition of each asset of the plurality of assets.

32. The method of any one of claims 23 to 31 wherein the at least one trend provides an analysis of the behavior of partial discharge events in relation to seasonal cycles and geographical factors, and wherein the method further includes a step of refining the at least one classifier model of the plurality of monitoring devices.

33. The method of any one of claims 23 to 32 wherein the at least one trend is utilized to establish an operational baseline for at least one group of infrastructure assets of the plurality of infrastructure assets, wherein each infrastructure asset in the at least one groupof infrastructure assets shares a common characteristic, the common characteristic selected from a group comprising: asset type, asset location, asset environmental conditions.

Images