A method and system for monitoring and early warning of power supply network state of a factory

Abstract

The application discloses a kind of factory power supply network state monitoring and early warning method and system, it is related to power grid intelligent detection field, including separating the electrical sampling sequence of target power supply network monitoring node from the industrial big data of factory, and synchronously extracting the production beat action sequence of nonlinear load equipment;In the dynamic harmonic benchmark template library that is pre-constructed based on historical industrial big data, corresponding benchmark template is retrieved and extracted as dynamic mask;Electrical sampling sequence and dynamic mask are input into sequence alignment model to perform alignment and subtraction operation on time axis, and output residual signal sequence;Residual signal sequence is input into abnormal identification model, and output whether there is the determination result of micro-arc event independent of production beat action sequence outside, and trigger early warning when determining that there is micro-arc event.The application solves the problem that micro-arc discharge signal is covered by high-frequency noise of normal production of nonlinear device, which leads to early warning lag or frequent false alarm.

Description

Technical Field

[0001] This invention relates to the field of smart power grid monitoring, and more specifically, to a method and system for monitoring and early warning of the status of a factory power supply network. Background Technology

[0002] In modern industrial automation, the operational status of a factory's power supply network directly impacts the continuity of the manufacturing process and electrical safety. Under long-term heavy-load operation, aging cable insulation, loose connections within distribution cabinets, or poor terminal contact can trigger early partial discharge or micro-arc phenomena. While these phenomena initially involve low energy and may not immediately trigger conventional overcurrent protection, their continued development can affect the insulation safety and operational reliability of power supply equipment. Therefore, monitoring the condition and providing early warnings of anomalies in factory power supply networks is of practical significance.

[0003] Existing technologies already include status monitoring solutions for power distribution networks or equipment. For example, invention patent CN116094163A discloses a feeder terminal with circuit breaker body fault early warning. This feeder terminal includes a smart switch and a smart terminal. It uses a DC sampling module, an AC sampling module, a control circuit detection module, and a wireless signal receiving module, combined with temperature, voltage, current, and control circuit indicator light data, to provide early warnings for circuit breaker body temperature rise, mechanical faults, and circuit breaks. Another example is invention patent CN116111727A, which discloses an abnormal monitoring method for integrated distribution boxes based on dynamic temperature thresholds. This method collects internal temperature data, ambient temperature data, load data, current data, and voltage data from the integrated distribution box, and sets a dynamic temperature threshold based on environmental hysteresis, load hysteresis, and aging degree to complete the abnormal monitoring of the integrated distribution box.

[0004] The aforementioned existing technologies can monitor power distribution equipment status, temperature changes, load changes, or electrical quantity changes, and are effective for early warning of common anomalies. However, in highly automated factory scenarios, there are often a large number of nonlinear load devices within the power supply network coverage area. Automated equipment, such as frequency converters, CNC machining equipment, and automatic welding equipment, injects harmonic components and high-frequency disturbances into the power supply network when performing different production actions. These disturbances are not faults themselves, but rather accompanying electrical characteristics of normal equipment operation.

[0005] In this context, the high-frequency discharge characteristics of early micro-arcs are often superimposed on electrical disturbances caused by normal production operations. If only a fixed threshold is used for monitoring, fluctuations caused by production operations are easily misjudged as anomalies. If the mixed electrical waveforms are directly input into a neural network model for end-to-end classification, the model may simultaneously learn production cycle disturbances and fault discharge characteristics, leading to anomaly identification results being significantly affected by changes in production operations. Therefore, how to distinguish between deterministic electrical disturbances caused by production operations and micro-arc events independent of production cycles in complex industrial power environments is a problem that needs further resolution in factory power supply network condition monitoring. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method and system for monitoring and early warning of the status of a factory power supply network, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A method for monitoring and early warning of the status of a factory power supply network includes the following steps: Acquire industrial big data from the factory, separate the electrical sampling sequence of the target power supply network monitoring node from the industrial big data, and simultaneously extract the production cycle action sequence of nonlinear load equipment within the coverage area of ​​the target power supply network; In a dynamic harmonic reference template library pre-built based on historical industrial big data, the reference template corresponding to the production cycle action sequence is retrieved and extracted as a dynamic mask. The electrical sampling sequence and the dynamic mask are input into the sequence alignment model to perform alignment and subtraction operations on the time axis, and the residual signal sequence is output. The residual signal sequence is input into the anomaly identification model, which outputs a determination result of whether there is a micro-arc event independent of the production cycle action sequence, and triggers an early warning when a micro-arc event is determined to exist.

[0008] Specifically, the construction process of the dynamic harmonic reference template library includes: Under the normal operating condition of the target power supply network without insulation faults, historical electrical sampling sequences corresponding to different production cycle action sequences are extracted from the historical industrial big data. The historical electrical sampling sequence is input into a time series clustering model to perform time series clustering processing; Extract the center sequence of each cluster output by the time series clustering model, and establish a mapping relationship between the center sequence and the corresponding production beat action sequence to form the dynamic harmonic reference template library.

[0009] Specifically, the training process of the anomaly detection model includes: Historical residual signal sequences labeled with records of micro-arc events were collected as positive sample data, and historical residual signal sequences without micro-arc events were collected as negative sample data. The positive sample data and the negative sample data are merged to form a training dataset, and the initial neural network model is trained under supervision using the training dataset. During supervised training, the cross-entropy loss function is used to calculate the loss value between the predicted classification result of the initial neural network model and the actual labeled result of the sample. The backpropagation algorithm is used to update the model parameters until the model converges, thus obtaining the trained anomaly recognition model.

[0010] Specifically, the process of inputting the electrical sampling sequence and the dynamic mask into the sequence alignment model to perform alignment and subtraction operations on the time axis includes: The sequence alignment model uses a dynamic time warping algorithm to calculate the matching path between the electrical sampling sequence and the dynamic mask; According to the optimal matching path corresponding to minimizing the objective function, each data point in the dynamic mask is shifted to the time node corresponding to the electrical sampling sequence; The electrical sampling sequence and the shifted dynamic mask are subjected to point-by-point interpolation at corresponding time nodes to obtain the residual signal sequence.

[0011] Specifically, the electrical sampling sequence includes a voltage sampling sequence and a current sampling sequence with a sampling frequency greater than a preset frequency threshold, and the nonlinear load device includes an automated operation device for which production action instructions are provided by a manufacturing execution system or a programmable logic controller.

[0012] Specifically, before inputting the electrical sampling sequence and the dynamic mask into the sequence alignment model to perform alignment and subtraction operations on the time axis, the method further includes: The envelope sequence of the power frequency fundamental current is extracted from the electrical sampling sequence; Within the preset time window corresponding to the production cycle action sequence, detect the abrupt start point of the rising edge of the envelope sequence and mark the abrupt start point as the physical action anchor point; The starting time point of the dynamic mask is shifted and calibrated based on the physical action anchor point, and the calibrated dynamic mask is input into the sequence alignment model for alignment and subtraction operations.

[0013] Specifically, the process of detecting the abrupt change start point of the rising edge of the envelope sequence includes: Calculate the sequence of first-order derivative values ​​of the envelope sequence with respect to time; The first time sampling point in the sequence of first-order derivative values ​​where the first-order derivative value changes abruptly from zero to a value greater than a preset derivative change threshold is determined as the change initiation point.

[0014] Specifically, before triggering the alert, it also includes: Extract the abnormal feature segments from the residual signal sequence that are determined by the anomaly identification model to contain micro-arc events, and store them in the reproduction buffer pool; The monitoring and judgment situation when the production cycle action sequence appears continuously within a preset number of monitoring cycles; The recurrence rate of the abnormal feature fragment appearing synchronously with the production cycle action sequence within the preset number of monitoring times is statistically analyzed. When the reproducibility rate is lower than a preset reproducibility rate threshold, the step of triggering an early warning is executed; When the reproducibility rate is higher than or equal to the preset reproducibility rate threshold, it is determined that the abnormal feature segment originates from the deterministic mechanical aging or component aging of the device, and the warning trigger operation for the current residual signal sequence is intercepted.

[0015] Specifically, after determining that the abnormal feature fragment originates from deterministic mechanical aging of the equipment or component aging, the process further includes: Weighting coefficients are assigned to the abnormal feature segments and the corresponding reference templates in the dynamic harmonic reference template library, and a weighted summation operation is performed on the two to generate an updated reference template. The updated reference template is used to replace the original reference template, thereby updating the template data in the dynamic harmonic reference template library. The abnormal feature fragments are extracted as equipment aging feature quantities and pushed to the equipment maintenance module of the manufacturing execution system.

[0016] This invention also discloses a factory power supply network status monitoring and early warning system for implementing the above method, comprising: The data synchronization acquisition module is used to acquire the factory's industrial big data, separate the electrical sampling sequence of the target power supply network monitoring node from the industrial big data, and synchronously extract the production cycle action sequence of nonlinear load equipment within the coverage area of ​​the target power supply network. The dynamic mask extraction module is used to retrieve and extract the reference template corresponding to the production cycle action sequence as a dynamic mask from the dynamic harmonic reference template library pre-built based on historical industrial big data. The residual decoupling module is used to input the electrical sampling sequence and the dynamic mask into the sequence alignment model to perform alignment and subtraction operations on the time axis, and output the residual signal sequence. An anomaly identification and early warning module is used to input the residual signal sequence into the anomaly identification model, and the anomaly identification model outputs a determination result of whether there is a micro-arc event independent of the production cycle action sequence, and triggers an early warning when a micro-arc event is determined to exist.

[0017] The advantage of this invention over existing technologies lies in its effective solution to the technical challenge of traditional monitoring methods facing frequent false alarms or serious underreporting due to the complete masking of the weak high-frequency characteristics of early micro-arcs in modern factory power supply networks by the massive and dynamically changing high-frequency noise of normal production. This invention breaks through the limitations of merely searching for patterns within pure electrical data, innovatively utilizing the cross-domain physical law that strictly binds the massive background electrical noise generated by the factory's nonlinear load equipment to its own mechanical production control cycle in time. By introducing industrial big data streams, the system uses the previous production cycle action sequence as an index to retrieve the dynamic mask specific to that action from the dynamic harmonic reference template library. Subsequently, a sequence alignment model is used to precisely align and subtract the real-time captured complex electrical waveform from this dynamic mask point by point. This decoupling method, which transcends electromechanical business logic, can accurately eliminate the massive background interference caused by the normal operation of mechanical equipment, clearly exposing the real micro-arc residual signal hidden under the mask, which is completely random and independent of production instructions. Thus, it achieves highly sensitive early warning of potential electrical insulation breakdown risks with zero false alarms and without interfering with normal production scheduling.

[0018] Furthermore, to address the hard alignment misalignment of timestamps caused by transmission delays in industrial Ethernet networks and lag in the transmission of pneumatic or mechanical components, this invention designs an alignment calibration mechanism based on low-frequency physical anchor points. This invention utilizes the objective law that the sudden burst of the macroscopic fundamental current and the injection of microscopic high-frequency harmonic noise are absolutely synchronized in physical time at the moment when electrical equipment is truly powered on. By capturing the abrupt start point of the rising edge of the fundamental current envelope as the absolute time anchor point, the system performs high-precision pre-shift calibration on the extracted dynamic mask. This scheme leverages the natural synchronicity between physical phenomena, completely eliminating the need to deploy an extremely expensive hardware-level high-precision clock synchronization network within the factory. Software cross-frequency band calibration alone overcomes the random delay errors of electromechanical execution, eliminating artificial glitches and false alarms caused by misalignment subtraction.

[0019] Furthermore, addressing the problem of irreversible, slow distortion of the reference noise characteristics during normal operation caused by gear wear or natural aging of internal components after long-term, high-intensity operation of factory equipment, this invention effectively resolves the dilemma of long-term slow failure and false positives due to outdated templates. This invention utilizes the fact that early dangerous arcs are occasional, random charge release events, while the aging of mechanical parts is a deterministic physical degradation law that recurs stably with every action once it occurs. By frequently statistically analyzing and tracking the recurrence rate of abnormal features in a buffer pool, the system can accurately distinguish whether the current waveform distortion is a fatal random arc or mild machine aging. For aging features with high recurrence rates, the system not only blocks invalid electrical fire alarms but also automatically integrates them into new templates to achieve system self-adaptation and evolution, and pushes this feature across domains to the equipment maintenance system. This design successfully transforms redundant noise that originally interfered with electrical monitoring into valuable resources, providing highly valuable health diagnostic indicators for predictive maintenance of factory equipment. Attached Figure Description

[0020] Figure 1 A schematic diagram of the factory physical environment and data acquisition in this invention; Figure 2 This is a schematic diagram of data separation and sequence extraction in this invention; Figure 3 This is a schematic diagram illustrating the construction of the dynamic harmonic reference template library of this invention; Figure 4 This is a schematic diagram of the sequence alignment and residual decoupling operation of the present invention; Figure 5 This is a schematic diagram of the micro-arc event determination and early warning triggering of the present invention. Detailed Implementation

[0021] The specific embodiments of the present invention will now be described with reference to the accompanying drawings.

[0022] This invention provides a method for monitoring and early warning of factory power supply network status, which can be deployed in a factory energy management system, power distribution monitoring system, edge computing gateway, or industrial server. Multiple monitoring nodes can be set up in the factory power supply network. These nodes can be located at the main incoming line cabinet, branch distribution cabinets, busbar branches, key production line distribution branches, or the power supply inlet of key nonlinear load equipment. The monitoring nodes are equipped with voltage transformers, current transformers, high-frequency current sensors, power quality acquisition units, or high-speed sampling modules to collect voltage and current signals. Production-side data can come from manufacturing execution systems, programmable logic controllers, CNC systems, robot controllers, frequency converter controllers, and field industrial Ethernet data. This forms industrial big data that simultaneously includes electrical quantities, equipment action commands, process cycles, equipment numbers, timestamps, and operating condition identifiers, such as... Figure 1 As shown.

[0023] In one embodiment, the system first acquires industrial big data from the factory and extracts the electrical sampling sequence of the target power supply network monitoring node. The target power supply network monitoring node refers to the sampling location corresponding to the distribution branch where insulation risk, poor contact risk, or partial discharge risk needs to be monitored. The electrical sampling sequence includes a voltage sampling sequence and a current sampling sequence, and the sampling frequency of the sampling module is greater than a preset frequency threshold. This preset frequency threshold is used to ensure that the abrupt changes and high-frequency disturbances caused by micro-arc discharge are retained in the sampling data, and can be set from 10kHz to 2MHz according to the on-site electrical noise frequency band, target identification frequency band, and sensor capabilities. For scenarios with strong high-frequency disturbances, such as automatic welding, servo drives, variable frequency motors, and CNC machine tools, a higher preset frequency threshold can be used; for ordinary conveyor lines, pumps, or fan branches, a lower preset frequency threshold can be used.

[0024] While separating the electrical sampling sequence, the system simultaneously extracts the production cycle action sequence of nonlinear load equipment within the target power supply network coverage area. Nonlinear load equipment can include automated operation equipment whose production action instructions are provided by a manufacturing execution system or programmable logic controller, such as frequency converter driven equipment, CNC machining equipment, servo press-fitting equipment, automatic welding equipment, robotic handling equipment, laser processing equipment, and switching power supply-intensive equipment. The production cycle action sequence is not simply the switching state of the equipment, but rather a sequence of data reflecting the actual operation stages of the equipment over a period of time, such as servo acceleration, spindle start, automatic welding arc initiation, fixture clamping, cylinder extension, tool feed, load holding, and reset return. Among these, automatic welding arc initiation is a normal production action, and its electrical disturbances can be used as dynamic mask learning objects, but are not equivalent to the micro-arc events caused by power supply network insulation defects or poor contact that this invention requires warning of. The system can generate the production cycle action sequence according to the equipment number, action instruction code, action start time, action end time, and action duration, such as... Figure 2 As shown.

[0025] The core idea of ​​this invention is to isolate deterministic electrical disturbances caused by normal production operations from real-time electrical waveforms, and then determine whether micro-arc events independent of the production cycle sequence exist in the remaining waveform. Since nonlinear load devices typically generate similar harmonic components, high-frequency ripple, and inrush current characteristics under the same production operations, these electrical disturbances caused by normal operations can be pre-learned and formed into a reference template. This invention uses this reference template as a dynamic mask. The dynamic mask refers to the typical electrical disturbance waveform corresponding to a certain type of production cycle sequence under normal operating conditions where the target power supply network has no insulation faults, no loose terminals, and no partial discharges. It is not a fixed threshold, nor a single-frequency filter template, but rather a time-series background waveform that switches according to the type of production operation.

[0026] During real-time monitoring, the system retrieves and extracts the reference template corresponding to the current production cycle action sequence from a pre-built dynamic harmonic reference template library based on historical industrial big data, using it as a dynamic mask. The retrieval process can use action command codes, equipment numbers, process step numbers, action durations, and load levels as indexes. When the production cycle action sequence is a discrete action combination, the template can be directly found through the action combination code; when the production cycle action sequence has duration fluctuations or load level changes, the action sequence can be encoded first, and then the closest reference template can be selected from the template library based on similarity. The purpose of this setting is that only dynamic masks that match the current action have canceling properties; using mismatched templates for subtraction may generate artificial residual spikes and increase false alarms.

[0027] The dynamic harmonic reference template library can be built before the system is put into operation or during the initial learning phase. Under normal operating conditions with no insulation faults in the target power supply network, the system extracts historical electrical sampling sequences corresponding to different production cycle action sequences from historical industrial big data. Normal operating conditions can be determined jointly by insulation detection records, inspection records, maintenance records, arc-free alarm records, and manual confirmation results. For each production cycle action sequence, the system collects historical electrical sampling sequences from multiple repetitions of that action. Because on-site mechanical delays, load fluctuations, and sampling noise can cause waveforms to be inconsistent under the same action, a time series clustering model is needed to obtain more stable and representative waveforms.

[0028] Time series clustering models can employ dynamic time warping distance (DTDM), k-shape clustering, autoencoder-based clustering, or Gaussian mixture clustering. For a more easily implemented approach in a factory setting, DTDM can be used to measure the morphological similarity between two historical electrical sampling sequences. Then, a center sequence clustering method based on DTDM is used to divide multiple samples within the same production cycle sequence into several clusters. The clustering model outputs the center sequence for each cluster. This center sequence can be the average sequence of samples within the cluster after time alignment, or it can be the true historical sequence with the smallest cumulative distance from other samples within the cluster. The system establishes a mapping relationship between the center sequences and the corresponding production cycle sequences, thereby forming a dynamic harmonic reference template library, such as... Figure 3 As shown, using cluster centers as templates can filter out sporadic noise, making the template closer to the stable electrical background of repeated normal production operations.

[0029] After extracting the dynamic mask, the system inputs the electrical sampling sequence and the dynamic mask into the sequence alignment model, performs alignment and subtraction operations on the time axis, and outputs the residual signal sequence. The sequence alignment model is used to solve the problem of slight differences in the speed of the same action in different production cycles. For example, for the same cylinder extension action, due to differences in air pressure, friction, load, and control response, its electrical disturbance waveform may be advanced, delayed, or locally stretched. If the dynamic mask is directly subtracted point by point according to the timestamp, the dynamic mask and the real-time electrical waveform may be misaligned, and the misaligned part may be mistakenly identified as abnormal. Therefore, this invention uses the sequence alignment model to first calibrate the dynamic mask on the time axis.

[0030] In one embodiment, the sequence alignment model employs a dynamic time warping algorithm to calculate the matching path between the electrical sampling sequence and the dynamic mask. The dynamic time warping algorithm finds the path with the minimum cumulative matching distance by comparing the distances between sampling points in two time sequences. This optimal matching path indicates that the data points in the dynamic mask should correspond to time nodes in the real-time electrical sampling sequence. Following this optimal matching path, the system adjusts, resamples, or maps each data point in the dynamic mask to the time node corresponding to the electrical sampling sequence; when multiple template points are mapped to the same real-time sampling point, the average value can be taken; when some real-time sampling points do not have directly corresponding template points, linear interpolation can be used to fill in the gaps. Subsequently, the system performs point-by-point interpolation on the electrical sampling sequence and the aligned dynamic mask at the corresponding time nodes to obtain the residual signal sequence, such as... Figure 4 As shown. For multi-channel sampled data containing voltage and current, alignment and subtraction can be performed on the voltage channel and the current channel respectively, and then the residuals of multiple channels can be concatenated as the input of the anomaly detection model.

[0031] The residual signal sequence represents the remaining components in the real-time electrical waveform that are not explained by the current production cycle. Under normal circumstances, if electrical disturbances mainly originate from equipment executing actions according to the cycle, the residual signal amplitude is small and exhibits random noise characteristics after subtracting the dynamic mask from the real-time electrical sampling sequence. If there is early breakdown of the cable insulation, poor terminal contact, loose connections inside the distribution cabinet, or a weak arc forming between conductors at different potentials, the high-frequency impact released by the micro-arc is not strictly bound to the production action command. In the residual signal sequence, it will manifest as spikes, short-term oscillations, aperiodic pulses, or localized energy surges. This invention does not directly search for anomalies in the mixed waveform, but rather identifies micro-arcs after subtracting the background of production actions, thus reducing misjudgments caused by fluctuations in normal production actions.

[0032] An anomaly detection model receives a sequence of residual signals and outputs a determination of whether a micro-arc event exists independently of the production beat sequence. This model can employ a one-dimensional convolutional neural network, a temporal convolutional network, a long short-term memory network, a gated recurrent unit network, or a neural network combining one-dimensional convolutional networks and recurrent networks. More specifically, it first extracts local spikes, local high-frequency oscillations, and short-term energy surges from the residual signal using one-dimensional convolutional layers. Then, it learns the temporal persistence of anomalous segments using recurrent or temporal convolutional layers. Finally, it outputs the probability of the micro-arc event through a fully connected layer. Model parameters include convolutional kernel weights, recurrent layer state weights, fully connected layer weights, and bias terms, which are automatically updated during training via backpropagation.

[0033] Training the anomaly detection model can be completed before system deployment or updated periodically on a factory edge server. During training, the system collects historical residual signal sequences labeled with records of micro-arc events as positive sample data. These records can come from arc detection devices, maintenance records, insulation anomaly location records, infrared inspection records, manual verification records, or post-fault tracing records. Simultaneously, the system collects historical residual signal sequences without micro-arc events as negative sample data. The negative samples should cover different production cycles, load levels, and shift environments to reduce the probability of the model misclassifying a normal operating condition as an anomaly. After merging the positive and negative sample data to form the training dataset, the initial neural network model undergoes supervised training.

[0034] During supervised training, the initial neural network model outputs a classification result for each historical residual signal sequence. This classification result can be either "existence of micro-arc event" or "absence of micro-arc event," or it can be the probability value of the corresponding micro-arc event. The system uses the cross-entropy loss function to calculate the loss value between the model's predicted classification result and the actual labeled result of the sample, and then uses the backpropagation algorithm to update the model parameters. The training process continues until the loss value tends to stabilize, the validation set recognition result no longer significantly improves, or the preset convergence condition is reached, thus obtaining the trained anomaly recognition model. In real-time operation, when the anomaly recognition model determines that a micro-arc event exists in the residual signal sequence, the system enters the early warning triggering process and can send the target monitoring node, occurrence time, residual signal segment, corresponding production cycle action sequence, and equipment information to the power distribution monitoring platform, such as... Figure 5 As shown.

[0035] In a further embodiment, before the electrical sampling sequence and the dynamic mask input sequence alignment model perform time axis alignment and subtraction, the system also performs translation calibration based on low-frequency physical action anchor points. This step is used to solve the overall misalignment problems caused by industrial Ethernet transmission delay, PLC scan cycle delay, pneumatic component action hysteresis, mechanical transmission hysteresis, and equipment controller timestamp errors. The timestamp of the production action command is not necessarily equivalent to the actual physical time point when the equipment is powered on and exerts force, and if the dynamic mask is aligned from the command time, it may have a deviation of tens of milliseconds or even greater from the real-time electrical waveform. Therefore, this invention finds the physical anchor point where the action actually occurs from the electrical sampling sequence itself.

[0036] In one embodiment, the system extracts the envelope sequence of the fundamental frequency current from the electrical sampling sequence. For 50Hz or 60Hz power frequency systems, digital bandpass filtering can be used to retain the fundamental frequency current component, and then the fundamental current envelope can be obtained through Hilbert transform, sliding RMS calculation, or peak tracking. The fundamental current envelope reflects the power-on process of the equipment at the macroscopic level. Compared with high-frequency harmonic noise, it has stronger anti-interference capabilities and is easier to pinpoint the actual start of the action.

[0037] Within a preset time window corresponding to the production cycle action sequence, the system detects the abrupt start point of the rising edge of the envelope sequence and marks this abrupt start point as the physical action anchor point. The preset time window is used to limit the search range and prevent the system from searching for anchor points in irrelevant fluctuations. This preset time window can be set from 20ms to 2000ms according to the maximum delay from PLC instruction to electrical response. For devices with fast response, such as servo drives and frequency converter starts, the preset time window can be from 20ms to 200ms; for pneumatic actuators, hydraulic mechanisms, or heavy-duty mechanical mechanisms, the preset time window can be from 200ms to 2000ms.

[0038] When detecting the abrupt start point of the rising edge of the envelope sequence, the system calculates the first-order derivative of the envelope sequence with respect to time. The first-order derivative reflects how quickly the envelope changes over time. When the device has not yet started operating, the envelope is in a stable state, corresponding to a first-order derivative value close to zero. When the device starts to operate, the fundamental current envelope rises rapidly, corresponding to a sudden change in the first-order derivative value from a stable state close to zero to a larger positive value. The system identifies the first time sampling point in the first-order derivative value sequence where the stable state close to zero abruptly changes to a value greater than a preset derivative abrupt change threshold as the abrupt start point. In engineering implementation, a stable state close to zero can be understood as the derivative value being in a stable and flat range after smoothing. The preset derivative abrupt change threshold can be set from 0.02 to 0.5 based on the normalized change amplitude of the envelope. A smaller threshold makes it easier to capture slight starts, while a larger threshold suppresses false triggers caused by background fluctuations.

[0039] After determining the physical action anchor point, the system performs translation calibration on the starting time point of the dynamic mask based on this anchor point. That is, instead of using only the production action command time as the starting point of the dynamic mask, the system moves the starting position of the dynamic mask's action to the time position corresponding to the abrupt change in the rising edge of the envelope in the real-time electrical sampling sequence. The calibrated dynamic mask is then input into the sequence alignment model for fine-grained alignment and subtraction. The purpose of this design is to first complete the overall time translation through the fundamental current envelope, and then complete the local shape alignment through dynamic time warping, which can simultaneously eliminate overall delay and local beat differences, and reduce spurious residuals caused by misaligned subtraction.

[0040] In a further embodiment, before triggering the warning, the system also sets up a recurrence rate judgment mechanism for abnormal feature segments. Due to long-term operation of factory equipment, wear of gears, bearings, aging of contactor contacts, aging of inverter power devices, or changes in mechanical damping may cause slow, stable, and repeatable distortions of normal electrical disturbances under the same production operation. These distortions may also appear in the residual signal sequence before the template is updated and be identified as abnormal by the anomaly identification model. To distinguish between occasional random micro-arcs and stable, recurring aging features, this invention performs recurrence rate tracking before issuing a warning.

[0041] When the anomaly detection model determines that a micro-arc event exists in the residual signal sequence, the system first extracts the anomalous feature segments from the residual signal sequence and stores them in the reproducibility buffer. The anomalous feature segments can be the time segments with the highest model attention, the residual segments with the highest classification contribution, or residual segments exceeding the anomalous probability region. The reproducibility buffer is used to temporarily store the waveform, occurrence time, relative production action phase, target monitoring node number, corresponding equipment number, and corresponding production cycle action sequence of the anomalous segments. Subsequently, the system tracks the monitoring and judgment when the same production cycle action sequence occurs consecutively within a preset number of monitoring cycles. The preset number of monitoring cycles can be set to 3 to 50 times; for fast-paced automated production lines, it can be set to 10 to 50 times; for equipment with long single-action intervals, it can be set to 3 to 10 times.

[0042] Within the aforementioned preset monitoring count, the system statistically analyzes the recurrence rate of abnormal feature segments appearing synchronously with the production cycle sequence. The recurrence rate can be understood as the proportion of subsequent monitoring of the same production cycle sequence where the abnormal feature segment reappears under similar action phases, waveform morphologies, or residual energy distributions compared to other abnormal feature segments in the recurrence buffer. If the abnormal feature segment consistently appears alongside the same production action, it is more likely to originate from equipment aging or component aging; if the abnormal feature segment only appears sporadically and does not consistently reappear with production actions, it is more likely to be a random micro-arc event independent of the production cycle.

[0043] The system compares the statistically obtained recurrence rate with a preset recurrence rate threshold. The preset recurrence rate threshold can be set between 0.6 and 0.95. For power distribution scenarios with high safety requirements and a desire to detect sporadic discharge hazards as early as possible, a lower threshold closer to 0.6 can be used; for scenarios with more normal aging disturbances and where early maintenance judgment is permissible, a higher threshold closer to 0.95 can be used. When the recurrence rate is lower than the preset recurrence rate threshold, the system executes an early warning trigger step, indicating that the target power supply network monitoring node may have early micro-arcs, weakened insulation, loose terminals, or poor contact risks. When the recurrence rate is higher than or equal to the preset recurrence rate threshold, the system determines that the abnormal feature segment originates from deterministic mechanical aging of the equipment or component aging, and intercepts the micro-arc early warning trigger operation for the current residual signal sequence. This interception does not discard the abnormal information, but rather transfers the abnormal information from the electrical safety early warning channel to the equipment maintenance analysis channel to avoid repeatedly misreporting stable aging features as random micro-arc events.

[0044] After determining that the abnormal feature segment originates from the deterministic mechanical aging of the equipment or the aging of its components, the system can also update the dynamic harmonic reference template library. Since the aging characteristics have become the stable electrical background of the equipment at the current stage, if the template is not updated, similar residuals will be generated every time the same production action occurs, thus continuously interfering with electrical monitoring. This invention assigns weight coefficients to the abnormal feature segment and the corresponding reference template in the dynamic harmonic reference template library, and performs a weighted summation operation on the two to generate an updated reference template. To avoid inconsistencies between the length of the abnormal feature segment and the complete reference template, before weighting, the abnormal feature segment can be aligned to the same action phase of the corresponding reference template along the time axis, and an aging compensation sequence can be constructed within the length range of the reference template. The aging compensation sequence retains the abnormal feature segment within the corresponding time period, and can be set to be consistent with the original reference template or set to not change the compensation value of the original reference template in other time periods. Subsequently, the system performs a weighted summation of the aging compensation sequence and the original reference template to obtain the updated reference template. The weighting coefficient of the baseline template can be set to 0.7 to 0.98, and the weighting coefficient of the aging compensation sequence can be set to 0.02 to 0.3, with the two added together to 1. When the weight of the baseline template is larger, the template update is more stable; when the weight of the aging compensation sequence is larger, the template adapts to changes in equipment status faster.

[0045] After generating the updated baseline template, the system replaces the original baseline template in the dynamic harmonic baseline template library with the updated template, thus completing the template data update. In this way, when the same equipment subsequently executes the same production cycle sequence, the system uses the new template, which already contains stable aging characteristics, as a dynamic mask, allowing the residual signal to more effectively reflect new, sporadic, and production cycle-independent anomalies. Simultaneously, the system extracts the anomalous feature fragments as equipment aging characteristic quantities and pushes them to the equipment maintenance module of the manufacturing execution system. These equipment aging characteristic quantities can include the amplitude, duration, relative action phase, recurrence rate, and trend of the anomalous fragment, allowing the maintenance module to generate inspection work orders, predictive maintenance suggestions, or spare parts replacement reminders. Thus, stable distortions that might have caused false alarms during electrical monitoring are transformed into equipment health management data.

[0046] This invention also provides a factory power supply network status monitoring and early warning system. This system can be composed of an edge acquisition terminal, an industrial server, and a monitoring platform, or it can be integrated into a single industrial computer. The system includes a data synchronization acquisition module, a dynamic mask extraction module, a residual decoupling module, and an anomaly identification and early warning module.

[0047] The data synchronization acquisition module is used to acquire industrial big data from the factory, extract the electrical sampling sequence of the target power supply network monitoring nodes from the industrial big data, and simultaneously extract the production cycle action sequence of nonlinear load equipment within the coverage area of ​​the target power supply network. This module can connect to power quality acquisition equipment, high-speed sampling cards, PLCs, MES, SCADA, and industrial Ethernet switching equipment, and complete data association through timestamps, device numbers, and monitoring node numbers.

[0048] The dynamic mask extraction module is used to retrieve and extract reference templates corresponding to production cycle action sequences from a pre-built dynamic harmonic reference template library based on historical industrial big data, using them as dynamic masks. This module can include a template library construction unit, an action sequence encoding unit, and a template retrieval unit. The template library construction unit collects historical electrical sampling sequences under normal operating conditions and extracts cluster center sequences using a time series clustering model; the action sequence encoding unit converts equipment action commands, action sequences, and action durations into searchable sequence indexes; and the template retrieval unit extracts corresponding templates based on real-time production cycle action sequences.

[0049] The residual decoupling module is used to input the electrical sampling sequence and the dynamic mask into the sequence alignment model to perform alignment and subtraction operations on the time axis, outputting a residual signal sequence. This module may include a physical action anchor point calibration unit, a dynamic time warping alignment unit, and a point-by-point difference calculation unit. The physical action anchor point calibration unit separates the power frequency fundamental current envelope from the electrical sampling sequence and uses the abrupt start point of the rising edge of the envelope to perform translation calibration on the starting time point of the dynamic mask; the dynamic time warping alignment unit calculates the optimal matching path between the electrical sampling sequence and the dynamic mask; the point-by-point difference calculation unit performs difference operations at the corresponding time nodes and forms a residual signal sequence.

[0050] The anomaly identification and early warning module is used to input the residual signal sequence into the anomaly identification model. The anomaly identification model outputs a judgment result on whether there is a micro-arc event independent of the production cycle action sequence, and triggers an early warning when a micro-arc event is determined to exist. This module may also include a recurrence buffer pool, a recurrence rate statistics unit, an early warning interception unit, a template update unit, and a maintenance push unit. The recurrence buffer pool is used to store residual signal segments that are judged to be abnormal; the recurrence rate statistics unit is used to count whether abnormal segments are synchronously reproduced in subsequent production cycle action sequences; the early warning interception unit is used to block micro-arc early warnings for the current residual signal sequence when the recurrence rate is higher than or equal to a preset recurrence rate threshold; the template update unit is used to integrate the stably reproduced aging characteristics into the dynamic harmonic reference template library; and the maintenance push unit is used to push equipment aging characteristic quantities to the equipment maintenance module of the manufacturing execution system.

[0051] Through the above implementation methods, this invention can decouple the deterministic harmonic disturbances caused by the normal production operations of nonlinear load equipment from the real-time electrical waveform, allowing the anomaly identification model to primarily address the random discharge characteristics in the residual signal. This processing method avoids the model simultaneously learning production cycle disturbances and micro-arc characteristics in the original mixed waveform, thereby improving the stability of early micro-arc event identification. Simultaneously, the pre-shift calibration based on physical action anchor points can reduce misalignment residuals caused by timestamp errors, and the early warning interception and template update mechanism based on recurrence rate can distinguish between random micro-arcs and stable equipment aging characteristics, enabling power supply network security early warning and predictive equipment maintenance to be collaboratively implemented within the same industrial data link.

[0052] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for monitoring and early warning of the status of a factory power supply network, characterized in that, Includes the following steps: Acquire industrial big data from the factory, separate the electrical sampling sequence of the target power supply network monitoring node from the industrial big data, and simultaneously extract the production cycle action sequence of nonlinear load equipment within the coverage area of ​​the target power supply network; In a dynamic harmonic reference template library pre-built based on historical industrial big data, the reference template corresponding to the production cycle action sequence is retrieved and extracted as a dynamic mask. The electrical sampling sequence and the dynamic mask are input into the sequence alignment model to perform alignment and subtraction operations on the time axis, and the residual signal sequence is output. The residual signal sequence is input into the anomaly identification model, which outputs a determination result of whether there is a micro-arc event independent of the production cycle action sequence, and triggers an early warning when a micro-arc event is determined to exist.

2. The method for monitoring and early warning of factory power supply network status according to claim 1, characterized in that, The construction process of the dynamic harmonic reference template library includes: Under the normal operating condition of the target power supply network without insulation faults, historical electrical sampling sequences corresponding to different production cycle action sequences are extracted from the historical industrial big data. The historical electrical sampling sequence is input into a time series clustering model to perform time series clustering processing; Extract the center sequence of each cluster output by the time series clustering model, and establish a mapping relationship between the center sequence and the corresponding production beat action sequence to form the dynamic harmonic reference template library.

3. The method for monitoring and early warning of factory power supply network status according to claim 1, characterized in that, The training process of the anomaly detection model includes: Historical residual signal sequences labeled with records of micro-arc events were collected as positive sample data, and historical residual signal sequences without micro-arc events were collected as negative sample data. The positive sample data and the negative sample data are merged to form a training dataset, and the initial neural network model is trained under supervision using the training dataset. During supervised training, the cross-entropy loss function is used to calculate the loss value between the predicted classification result of the initial neural network model and the actual labeled result of the sample. The backpropagation algorithm is used to update the model parameters until the model converges, thus obtaining the trained anomaly recognition model.

4. The method for monitoring and early warning of factory power supply network status according to claim 1, characterized in that, The process of inputting the electrical sampling sequence and the dynamic mask into the sequence alignment model to perform alignment and subtraction operations on the time axis includes: The matching path between the electrical sampling sequence and the dynamic mask is calculated using the dynamic time warping algorithm based on the sequence alignment model. According to the optimal matching path corresponding to minimizing the objective function, each data point in the dynamic mask is shifted to the time node corresponding to the electrical sampling sequence; The residual signal sequence is obtained by performing point-by-point difference operation between the electrical sampling sequence and the shifted dynamic mask at the corresponding time nodes.

5. The method for monitoring and early warning of factory power supply network status according to claim 1, characterized in that, The electrical sampling sequence includes a voltage sampling sequence and a current sampling sequence with a sampling frequency greater than a preset frequency threshold, and the nonlinear load device includes an automated operation device for which production action instructions are provided by a manufacturing execution system or a programmable logic controller.

6. The method for monitoring and early warning of factory power supply network status according to claim 1, characterized in that, Before inputting the electrical sampling sequence and the dynamic mask into the sequence alignment model to perform alignment and subtraction operations on the time axis, the method further includes: The envelope sequence of the power frequency fundamental current is extracted from the electrical sampling sequence; Within a preset time window corresponding to the production cycle action sequence, detect the abrupt start point of the rising edge of the envelope sequence and mark the abrupt start point as a physical action anchor point; The starting time point of the dynamic mask is shifted and calibrated based on the physical action anchor point, and the calibrated dynamic mask is input into the sequence alignment model for alignment and subtraction operations.

7. The method for monitoring and early warning of factory power supply network status according to claim 6, characterized in that, The process of detecting the abrupt start point of the rising edge of the envelope sequence includes: Calculate the sequence of first-order derivatives of the envelope sequence with respect to time; The first time sampling point in the sequence of first-order derivative values ​​where the first-order derivative value changes abruptly from zero to a value greater than a preset derivative change threshold is determined as the change initiation point.

8. The method for monitoring and early warning of factory power supply network status according to claim 1, characterized in that, Before triggering the alert, it also includes: Extract the abnormal feature segments from the residual signal sequence that are determined by the anomaly identification model to contain micro-arc events, and store them in the reproduction buffer pool; The monitoring and judgment situation when the production cycle action sequence appears continuously within a preset number of monitoring cycles; The recurrence rate of the abnormal feature fragment appearing synchronously with the production cycle action sequence within the preset number of monitoring times is statistically analyzed. When the reproducibility rate is lower than a preset reproducibility rate threshold, the step of triggering an early warning is executed; When the reproducibility rate is higher than or equal to the preset reproducibility rate threshold, it is determined that the abnormal feature segment originates from the deterministic mechanical aging or component aging of the device, and the warning trigger operation for the current residual signal sequence is intercepted.

9. The method for monitoring and early warning of factory power supply network status according to claim 8, characterized in that, After determining that the abnormal feature fragment originates from deterministic mechanical aging of the device or component aging, the method further includes: Weighting coefficients are assigned to the abnormal feature segments and the corresponding reference templates in the dynamic harmonic reference template library, and a weighted summation operation is performed on the two to generate an updated reference template. The updated reference template is used to replace the original reference template, thereby updating the template data in the dynamic harmonic reference template library. The abnormal feature fragments are extracted as equipment aging feature quantities and pushed to the equipment maintenance module of the manufacturing execution system.

10. A factory power supply network status monitoring and early warning system for implementing the method of claim 1, characterized in that, include: The data synchronization acquisition module is used to acquire the factory's industrial big data, separate the electrical sampling sequence of the target power supply network monitoring node from the industrial big data, and synchronously extract the production cycle action sequence of nonlinear load equipment within the coverage area of ​​the target power supply network. The dynamic mask extraction module is used to retrieve and extract the reference template corresponding to the production cycle action sequence as a dynamic mask from the dynamic harmonic reference template library pre-built based on historical industrial big data. The residual decoupling module is used to input the electrical sampling sequence and the dynamic mask into the sequence alignment model to perform alignment and subtraction operations on the time axis, and output the residual signal sequence. The anomaly identification and early warning module is used to input the residual signal sequence into the anomaly identification model, and the anomaly identification model outputs a determination result of whether there is a micro-arc event independent of the production cycle action sequence, and triggers an early warning when a micro-arc event is determined to exist.

Images