Electric power meter data real-time analysis and abnormity early warning method based on edge calculation

Abstract

The invention provides an electric power meter data real-time analysis and abnormity early warning method based on edge calculation, and belongs to the technical field of data analysis and early warning, and the method comprises the steps: collecting the multi-modal data of an electric power meter based on a multi-channel analog-to-digital converter, employing a hardware filter circuit to eliminate high-frequency noise, and obtaining the data of the electric power meter; signal gain self-adaptive adjustment is carried out based on a reconfigurable amplifier; a digital signal processing accelerator is used to carry out Fourier transform on a matrix constructed by the standardized data stream, and a multi-core processor is combined to carry out analysis to generate a feature vector; the method comprises the following steps: acquiring environmental parameters in real time based on an environmental sensor, dynamically adjusting an anomaly detection threshold in combination with a machine learning model trained by historical data, and comparing a feature vector with the dynamic threshold to generate an anomaly signal; the relay control circuit triggers sound-light alarm, the physically isolated communication module sends an early warning instruction to the cloud platform, and meanwhile, the redundant power supply module is started to control the electric power meter to operate continuously. And the efficiency and accuracy of subsequent abnormity early warning are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power systems, in particular to a power meter data real-time analysis and abnormal early warning method based on edge computing. BACKGROUND

[0002] In the process of intelligent upgrading of the power system, the real-time performance, accuracy and timeliness of abnormal early warning of the data of the power meter as a key monitoring node directly affect the safety of the power grid operation and the reliability of power supply. With the access of distributed energy and the popularity of power electronic equipment, the power system presents the characteristics of multi-source heterogeneity, dynamic fluctuation and complex interference. The traditional power meter data processing method gradually exposes technical bottlenecks that are difficult to overcome, which are as follows:

[0003] In the prior art, the hardware filter circuit is a passive filter with preset parameters (such as an RC filter with a fixed cutoff frequency), which cannot adapt to the dynamic changes of high-frequency interference (such as electromagnetic interference generated by motor start-stop and electric arc discharge) in the power system, resulting in incomplete noise suppression, low signal-to-noise ratio of raw data, and fixed gain design of traditional amplifiers, which is difficult to cope with wide range fluctuations of power parameters (such as dramatic changes in voltage and current when load suddenly changes), which is prone to signal saturation or insufficient gain, resulting in data distortion. The full data collected by the power meter is uniformly operated, and no feature enhancement is performed on local abnormal data (such as instantaneous distortion of a phase voltage), which causes abnormal features to be diluted by global data. The traditional abnormal early warning relies on the experience of setting static thresholds, but the environment of the power meter is not constant. The above factors will lead to low efficiency and low accuracy of subsequent abnormal early warning.

[0004] Therefore, the present application provides a power meter data real-time analysis and abnormal early warning method based on edge computing. SUMMARY

[0005] The present application provides a power meter data real-time analysis and abnormal early warning method based on edge computing to solve the above technical problems.

[0006] The present application provides a power meter data real-time analysis and abnormal early warning method based on edge computing, comprising:

[0007] Step 1: A multi-channel analog-to-digital converter integrated based on an edge computing node collects multi-modal data of a power meter, a hardware filter circuit is used to eliminate high-frequency noise, and a signal gain self-adaptive adjustment is performed based on a reconfigurable amplifier to generate a standardized data stream;

[0008] Step 2: Fast Fourier transform is performed on the global standardization matrix and the local abnormal standardization matrix constructed by the standardized data stream using a digital signal processing accelerator, and a corresponding local power parameter feature vector and a global power parameter feature vector are generated through operation analysis combined with a multi-core processor;

[0009] Step 3: Based on the real-time collection of environmental parameters by the environmental sensor and the machine learning model trained based on historical data, the abnormal detection threshold is dynamically adjusted, and the local power parameter feature vector and the global power parameter feature vector are compared with the dynamic threshold to generate an abnormal signal;

[0010] Step 4: According to the type and level of the abnormal signal, the relay control circuit is triggered to sound and light alarm, and the pre-warning instruction is sent to the cloud platform through the physically isolated communication module, and at the same time, the redundant power module is started to control the power meter to run continuously.

[0011] Preferably, before the high-frequency noise is eliminated by the hardware filtering circuit, it includes:

[0012] A plurality of sets of historical data are obtained and input into the initial filtering circuit to obtain historical filtering data, obtaining a double array of noise elimination and non-noise elimination of the same index, and inputting the double array into a linear analysis model to determine the linear coefficient of noise elimination, and at the same time, the filtering difference data of each historical filtering data is input into a difference analysis model, and the difference decomposition coefficient of each index is obtained in combination with the mutual dependence relationship between the indexes;

[0013] According to the linear coefficient and the difference decomposition coefficient, the failure problem of the initial filtering circuit is analyzed;

[0014] According to the failure problem, the improved structure of the initial filtering circuit is determined, and the improved structure and the initial filtering circuit are fused to obtain the hardware filtering circuit.

[0015] Preferably, the reconfigurable amplifier includes:

[0016] The gain intelligent control module composed of the signal feature extraction unit and the dynamic decision circuit, the reconfigurable cascade amplification array composed of the four-stage gain cooperative regulation architecture, the environment adaptive filtering module composed of the temperature drift compensation type active band-pass filter and the intelligent notch circuit, the intelligent analog-to-digital conversion module composed of the double-channel synchronous sampling ADC, and the data standardization processing module composed of the multi-level cache and the high-reliability transmission unit;

[0017] The gain intelligent control module, the reconfigurable cascade amplification array, the environment adaptive filtering module, the intelligent analog-to-digital conversion module, and the data standardization processing module are connected in sequence.

[0018] Preferably, the local anomaly standardization matrix is constructed based on the standardized data stream, comprising:

[0019] The power system monitored by the power meter is determined, and a standardized matrix is constructed according to the standardized data of the power system at each monitoring time point. The remaining sampling points except the first sampling point and the tail sampling point in each row of the standardized matrix are sequentially selected as second sampling points, wherein the standardized data stream is composed of the standardized data of multiple monitoring time points.

[0020] The absolute values of the first difference between the two adjacent column elements of the second sampling point are taken, and the nearest cycle boundary point from the second sampling point is determined according to the set state working cycle of the power system, and the absolute value of the second difference of the corresponding row of the locked element value is taken.

[0021] The sum of all first difference absolute values and second difference absolute values is obtained, and the step coefficient corresponding to the second sampling point is obtained in combination with the set state working cycle.

[0022] When the step coefficient of the second sampling point is greater than the preset coefficient, a to-be-verified label is set to the second sampling point.

[0023] Otherwise, a normal label is set to the second sampling point.

[0024] According to the label setting result, the to-be-verified column proportion of each column vector in the corresponding standardized matrix and the to-be-verified row proportion of each row vector are determined, and are compared with the preset proportion to lock the abnormal row and the abnormal column.

[0025] According to the first number of abnormal indicators corresponding to the abnormal rows involved in each abnormal column, the second number of to-be-verified labels corresponding to the abnormal columns involved, and the abnormal factor corresponding to each element value in the corresponding column vector, the abnormal column is coefficient amplified to form a local anomaly standardization matrix.

[0026] Preferably, the abnormal factor is determined, comprising:

[0027] The clustering analysis is performed on each row vector to obtain the difference coefficient of each clustering result and the corresponding standard range, and the abnormal factor of the corresponding clustering result is determined according to the data trend of the corresponding clustering result.

[0028] Preferably, the environmental sensor is used to collect environmental parameters in real time, and the machine learning model trained based on historical data is used to dynamically adjust the anomaly detection threshold, comprising:

[0029] The environmental parameters are mapped to a fuzzy set based on the fuzzy logic layer of the machine learning model, the fuzzy membership degree is calculated by a gradient membership function, and an environmental gradient factor is introduced.

[0030] Based on the fault probability distribution output by the neural network layer input model of the machine learning model, and combined with the device aging factor and the environmental gradient factor, the dynamic threshold of the corresponding index is calculated;

[0031] Based on FPGA, fuzzy reasoning and neural network parallel operation are realized, and threshold conflict detection formula Yc trigger priority arbitration is integrated:

[0032]

[0033] Wherein, Th1, Th2 are historical threshold, real-time dynamic threshold respectively; δ is conflict judgment threshold;

[0034] When the threshold is adjusted, the front-end acquisition module is automatically adapted, the ADC switching anti-aliasing filter mode is switched under high interference, and the amplifier gain is dynamically adjusted based on priority arbitration:

[0035] G(t)=G0·exp(-k·EMI(t))·ln(e-Yc)

[0036] Wherein, EMI(t) is the intensity of electromagnetic interference; G0 is the initial gain of the amplifier; G(t) is the real-time gain of the amplifier; k is the attenuation coefficient; e is a constant, and the value is 2.7;

[0037] Based on the real-time gain warning, the dynamic warning formula of the local display screen is triggered Encode the fault level, wherein Co is the RGB color value; ΔTh is the threshold offset; RGB() is a three-channel color function.

[0038] Preferably, compared with the dynamic threshold, an abnormal signal is generated, including:

[0039] From the local power parameter feature vector and the global power parameter feature vector, the feature pair of the same index is extracted, the feature pair is numerically analyzed to obtain a numerical pair, and the numerical pair is compared with the corresponding dynamic threshold respectively;

[0040] Yx={|VJ-Th2|>ε1,|VQ-Th2|>ε2}

[0041] Wherein, VJ, VQ are local feature value, global feature value in the feature pair respectively; ε1, ε2 are local conflict threshold, global conflict threshold matched with the numerical pair respectively; Yx is the comparison result;

[0042] The comparison result is matched with the result-index-signal table to obtain the abnormal signal of the corresponding index.

[0043] Preferably, according to the type and level of the abnormal signal, the relay control circuit triggers an audible and light alarm, and sends a warning instruction to the cloud platform through a physically isolated communication module, including:

[0044] According to the monitoring type to which the index belongs and the signal level of the abnormal signal The relay control circuit outputs a differential audible and light signal, wherein min is a minimum value symbol; max is a maximum value symbol; and δ0 is a basic classification threshold value; is a floor symbol;

[0045] Based on a sub-instruction including the monitoring type to which the index belongs, the signal level, the occurrence time stamp, and the associated environmental parameters;

[0046] According to the instruction structure and the instruction triggering condition of the sub-instruction, a enable state is determined, and according to the enable state, a clock enable signal of the corresponding warning instruction is determined;

[0047] According to the monitoring type to which the sub-index belongs, a display address on the cloud platform is determined, and a frequency conversion segment is generated according to the clock enable signal of the corresponding sub-instruction to generate a check signal, and the sub-instruction is checked for errors in dependence on the check signal;

[0048] When there is an error, the sub-instruction is cleared and a new enable signal is determined in dependence on the check signal;

[0049] In dependence on the new enable signal, the clock enable signal of the corresponding sub-instruction is adjusted to obtain a replacement enable signal, and the clock enable signal of the corresponding sub-instruction is adjusted to obtain a replacement enable signal;

[0050] If there is no error, the clock enable signal of the corresponding sub-instruction is retained;

[0051] The last clock enable signal is clocked in sequence to determine the execution probability of each clock to obtain a warning instruction, and the encrypted warning instruction is sent to the cloud platform through a physically isolated communication module.

[0052] Compared with the prior art, the beneficial effects of the present application are:

[0053] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the present application. The objects and other advantages of the present application can be achieved and obtained by the structure particularly pointed out in the written description and the accompanying drawings.

[0054] The technical solutions of the present application will be further described in detail below with the help of the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0055] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and are used to explain the application without restricting it. In the drawings:

[0056] Figure 1 A flow chart of an edge computing-based real-time power meter data analysis and abnormality early warning method in an embodiment of the application;

[0057] Figure 2 A structural diagram of a reconfigurable amplifier in an embodiment of the application. DETAILED DESCRIPTION

[0058] The preferred embodiments of the application are described below in conjunction with the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to explain and illustrate the application, and do not limit the application.

[0059] The application provides an edge computing-based real-time power meter data analysis and abnormality early warning method, as shown in Figure 1 , comprising:

[0060] Step 1: Collecting multi-modal data of the power meter based on the multi-channel analog-to-digital converter integrated with the edge computing node, eliminating high-frequency noise by using the hardware filter circuit, and performing self-adaptive adjustment of signal gain based on the reconfigurable amplifier to generate a standardized data stream;

[0061] Step 2: Performing fast Fourier transform on the global standardized matrix and the local abnormality standardized matrix constructed by the standardized data stream by using the digital signal processing accelerator, and generating corresponding local power parameter feature vectors and global power parameter feature vectors by combining the multi-core processor for operation analysis;

[0062] Step 3: Real-time collection of environmental parameters based on the environmental sensor, dynamic adjustment of the abnormality detection threshold value based on the machine learning model trained by the historical data, comparison of the local power parameter feature vectors and the global power parameter feature vectors with the dynamic threshold value, and generation of an abnormality signal;

[0063] Step 4: Triggering the audible-light alarm by the relay control circuit according to the type and level of the abnormality signal, sending the early warning instruction to the cloud platform through the physically isolated communication module, and starting the redundant power module to control the power meter to continuously run.

[0064] In this embodiment, the edge computing node is a localized computing unit deployed near the power meter, responsible for real-time data processing (reducing cloud transmission delay), such as an industrial-grade edge gateway (e.g., an edge controller with an ARM Cortex-A53 processor), specifically: integrated Ethernet, RS485 interface, supports real-time operating system (e.g., FreeRTOS), wide temperature (-40~85℃) design suitable for power scenarios.

[0065] The multi-channel analog-to-digital converter (ADC) is a hardware module that simultaneously acquires multiple analog signals and converts them to digital signals, specifically: integrated in the hardware layer of the edge computing node, with channels connected to the voltage, current, power, etc. analog output terminals of the power meter.

[0066] The power meter is a device for measuring power system parameters, including smart meters, switch cabinet monitoring terminals, etc., such as a three-phase smart meter (measuring voltage, current, active power, power factor), specifically: communicates with the edge node through the Modbus protocol, outputs 4~20mA analog signals or digital signals.

[0067] Multi-modal data is power parameter data of different types and dimensions, such as voltage waveform (time domain data), harmonic content (frequency domain data), and switch state (state quantity data).

[0068] The hardware filter circuit is a filter circuit composed of resistors, capacitors, inductors, etc., which suppresses specific frequency noise, such as a second-order RC low-pass filter (cutoff frequency 1kHz) + notch circuit (suppresses 50Hz power frequency interference).

[0069] High-frequency noise is an interference signal with a frequency higher than the power system fundamental frequency (50Hz), such as 10kHz electromagnetic interference generated by motor commutation and wideband noise generated by arc discharge.

[0070] The reconfigurable amplifier is an amplifier circuit that can dynamically adjust the gain through software or hardware, specifically: the gain is switched through the GPIO of the edge node, and the gain is automatically switched according to the input signal amplitude (e.g., when the voltage is less than 2V, the gain is automatically switched to 8 times).

[0071] Signal gain adaptive adjustment is automatically adjusting the amplifier gain according to the input signal strength to avoid signal saturation or insufficient amplitude, such as when the current signal suddenly changes from 5A (corresponding to 1V output) to 50A (corresponding to 10V output), the amplifier gain decreases from 8 times to 1 times.

[0072] Standardized data stream is a digital data sequence with uniform format, range, and sampling rate, such as converting voltage (0~10kV) and current (0~100A) to normalized values between 0 and 1, and packaging them in the format "timestamp + parameter ID + normalized value".

[0073] Digital signal processing accelerator is a special hardware unit that accelerates signal processing operations such as Fourier transform and filtering, such as DSP slices in FPGA (e.g. DSP48E1 unit in Xilinx Artix-7).

[0074] In this embodiment, the global standardization matrix is a two-dimensional matrix containing the standardization data of all power parameters (row = time point, column = parameter type), such as an 8-row (8 parameters such as voltage, current, power) x 100-column (100 sampling time) matrix.

[0075] In this embodiment, the local abnormality standardization matrix is a strengthened matrix constructed for suspected abnormal local parameters (such as distortion of a phase voltage), such as extracting the abnormal segment of the A-phase voltage harmonic column in the global matrix and expanding it into a 1-row x 50-column local matrix.

[0076] Fast Fourier Transform (FFT) is an algorithm for converting time-domain signals to frequency-domain, used to extract features such as harmonics and resonance, such as performing FFT on the voltage signal to obtain the amplitude ratio of the 50Hz fundamental and the 150Hz third harmonic.

[0077] Multi-core processor is a multi-core CPU that supports parallel computing, such as a 4-core ARM Cortex-A72 processor (clock speed 2.0GHz), which is specifically assigned 1 core to process the global matrix and 1 core to process the local matrix, with data interaction through shared memory.

[0078] Local power parameter feature vector is a vector representing the characteristics of local abnormal parameters (such as harmonic content and fluctuation amplitude of a certain device), such as [third harmonic content 5%, instantaneous voltage fluctuation ±2%, distortion duration 0.5s].

[0079] Global power parameter feature vector is a feature vector representing the overall grid state (such as grid frequency, power angle, line loss rate), such as [frequency 50.02Hz, power angle 10°, line loss rate 2.3%].

[0080] Environmental sensor is a device that collects environmental parameters such as temperature, humidity, and electromagnetic interference, such as temperature and humidity sensor (SHT30) and electromagnetic interference sensor (EMC-100).

[0081] Environmental parameter is an environmental physical quantity that affects power equipment, such as temperature 35℃, humidity 60%RH, and electromagnetic interference intensity 80dBμV / m.

[0082] Machine learning model trained on historical data is a prediction model trained on historical normal / abnormal data (such as a fuzzy logic + neural network hybrid model), which is specifically trained on power parameters and environmental data from the past year to output abnormality probability.

[0083] In this embodiment, the dynamic adjustment of the abnormality detection threshold is a dynamically updated abnormality determination threshold according to the environment and device aging, for example, when the temperature is high (40°C), the voltage fluctuation threshold is relaxed from ±5% to ±7% (adapt to the change of device characteristics).

[0084] The dynamic threshold is a threshold that changes with the environment and device state (different from the fixed threshold), for example, the dynamic threshold of A-phase voltage = 220V ± (0.01 × temperature + 0.005 × running time).

[0085] The abnormal signal is an electrical signal representing the abnormality of the power parameter (such as level jump, pulse signal), for example, when the A-phase current feature vector exceeds the dynamic threshold, the edge node outputs a 3.3V high-level abnormal signal.

[0086] The type of abnormal signal is an abnormal category classified by parameter type, for example, voltage abnormality, current harmonic exceeding, frequency deviation.

[0087] The level of abnormal signal is classified by severity (such as 1-3 levels).

[0088] The relay control circuit is an electronic circuit that controls the action of the relay, specifically: a solid-state relay circuit driven by a triode, receiving a 3.3V control signal.

[0089] The acousto-optic alarm is an alarm sent through sound and light signals, for example: level 1 = green LED flicker + 1kHz buzzer; level 3 = red LED flashing + 8kHz continuous buzzer, specifically: the relay controls the LED drive circuit and buzzer respectively, the alarm mode is bound to the level (hardware lookup table implementation).

[0090] The physically isolated communication module is a module that realizes electrical isolation between the power grid side and the communication side through photoelectric and electromagnetic isolation, for example, a 4G module with a photoelectric coupler (such as Huawei ME909s), with an isolation voltage ≥2500V.

[0091] The cloud platform is a cloud server platform for remote monitoring and data analysis, specifically: deployed in public cloud / private cloud, supporting MQTT protocol to receive edge node data, providing visual dashboard.

[0092] The warning instruction is an abnormal notification instruction sent to the cloud, transmitted through physically isolated communication module encryption (AES-128 encryption), and the instruction retransmission mechanism ensures delivery.

[0093] The redundant power module is a backup power module that automatically switches when the main power fails.

[0094] The beneficial effects of the above technical solutions are: through edge side hardware acceleration, dynamic threshold adaptation, secure communication and hierarchical response, real-time analysis and accurate early warning of power meter data are realized, the abnormal detection precision is improved, and through physical isolation and redundant power supply, the system reliability is ensured, and the high real-time and high security requirements of complex power environment are adapted.

[0095] The application provides a power meter data real-time analysis and abnormal early warning method based on edge computing, before a hardware filter circuit is used to eliminate high-frequency noise, comprising:

[0096] A plurality of sets of historical data are obtained and input into an initial filter circuit in sequence to obtain historical filter data, a double array of noise elimination-noise elimination of the same index is obtained, and the double array is input into a linear analysis model to determine a linear coefficient of noise elimination, at the same time, filter difference data of each historical filter data is determined and input into a difference analysis model, and a difference decomposition coefficient of each index is obtained in combination with the mutual dependence relationship between indexes;

[0097] According to the linear coefficient and the difference decomposition coefficient, the failure problem of the initial filter circuit is analyzed;

[0098] According to the failure problem, an improved structure of the initial filter circuit is determined, the improved structure and the initial filter circuit are fused to obtain a hardware filter circuit.

[0099] In this embodiment, the plurality of sets of historical data are original power parameter data collected by the power meter under different working conditions (such as normal operation, load mutation, fault state), which are used to analyze the performance of the filter circuit, for example, 100 sets of data in the past 3 months are selected, each set contains 24 hours of voltage (0-10kV) and current (0-100A) sampling values (sampling rate 1kHz).

[0100] The initial filter circuit is a prototype filter circuit that has not been optimized, which is used to preliminarily filter the noise in the power signal, for example, a first-order RC low-pass filter (cut-off frequency about 1.6kHz) composed of a 1kΩ resistor and a 100nF capacitor.

[0101] The historical filter data is denoising data obtained by processing the historical data through the initial filter circuit, for example, the noise amplitude of the filtered data is reduced from 500mV to 100mV after the original voltage data (containing 10kHz high-frequency noise) is processed by the initial RC filter.

[0102] Noise-uncancelled pair of the same index is a pair of arrays composed of "unfiltered raw data" and "filtered denoised data" for the same power parameter (such as A-phase voltage), such as the A-phase voltage pair = [{"raw data": [220.1, 220.3,..., 219.8]}, {"filtered data": [219.9, 220.0,..., 219.7]}], each group containing 1000 sampling points.

[0103] Linear analysis model is a mathematical model for analyzing the linear relationship between raw data and filtered data to determine the linear attenuation characteristics of the filtering circuit, such as a linear regression model: y = ax + b, where x is the raw data, y is the filtered data, a is the linear coefficient, and b is the intercept.

[0104] Linear coefficient of noise cancellation is a parameter representing the linear attenuation ability of the filtering circuit to noise, the closer the value is to 0, the better the noise cancellation effect, such as the linear coefficient k = 0.2 of 10 kHz noise, indicating that the original noise amplitude is attenuated to 20% after filtering.

[0105] Filtering difference data of each historical filtered data is the difference between the raw data and the historical filtered data at the same sampling time, reflecting the amount of noise remaining that the filtering circuit has not eliminated, such as the raw voltage value 220.5V, filtered to 219.8V, filtering difference data = 0.7V (including uneliminated high-frequency noise), specifically: calculate the difference of the corresponding elements of the pair of arrays to generate the filtering difference array.

[0106] Difference analysis model is a model for decomposing the source of filtering difference data, identifying whether noise residue is caused by circuit parameter defects or external interference, specifically: using principal component analysis (PCA) model to decompose the filtering difference data into "circuit bandwidth deficiency", "component tolerance", "external interference" and other principal components.

[0107] Interdependence between indicators is the correlation between different power parameters (such as voltage and current, harmonic and power factor), which is a cross factor affecting the filtering effect, such as when the current harmonic increases, the voltage waveform distortion intensifies, and the correlation coefficient between the two is 0.8 (strong positive correlation), specifically: the dependence relationship is quantified through Pearson correlation coefficient calculation or causal inference model (such as Granger causality test), and the result is stored in matrix form.

[0108] Difference decomposition coefficient of each indicator is the coefficient of filtering difference data decomposed into each parameter according to the dependence relationship between indicators, reflecting the contribution of a certain parameter to the overall noise residue, such as the difference decomposition coefficient of the voltage indicator = 0.6 (indicating that 60% of the filtering difference is caused by voltage-related factors), and the current indicator = 0.4.

[0109] The failure problem of the initial filter circuit is the defect of the initial filter circuit in noise elimination and adaptability, which is analyzed by linear coefficient and difference decomposition coefficient, for example, the low linear coefficient (such as k=0.5) indicates that the high-frequency noise attenuation is insufficient;

[0110] The difference decomposition coefficient shows that the current harmonic causes 30% of the filtering difference, which indicates that the circuit has weak harmonic suppression ability.

[0111] The improved structure of the initial filter circuit is a supplementary circuit structure designed for the failure problem, which is used to enhance the filtering performance, for example, for the insufficient high-frequency noise attenuation: adding a second-order Butterworth low-pass filter (cut-off frequency 500Hz); for the weak harmonic suppression: series 50Hz trap circuit (Q value = 10).

[0112] Specifically, the frequency response of the improved structure is simulated by using Multisim software, and whether the-3dB attenuation point and the stopband attenuation meet the standard (such as 10kHz noise attenuation≥40dB) is verified.

[0113] The fusion of the improved structure and the initial filter circuit is to combine the improved structure and the initial circuit through hardware switching or cascading to form a composite circuit with basic filtering and targeted optimization, for example, initial RC filter (first level) → improved second-order Butterworth filter (second level) → trap circuit (third level), and the structure switching under different working conditions is realized through a relay, specifically: reserving the interface of the improved structure in the PCB design, and controlling the relay switching through the GPIO of the edge node (such as automatically enabling the trap circuit when the harmonic is detected to be over-standard).

[0114] The hardware filter circuit is the final filter circuit that fuses the initial circuit and the improved structure, and has self-adaptive noise elimination capability, for example, cascaded composite circuit (RC filter + Butterworth filter + trap circuit), which supports manual / automatic switching of filter mode, specifically: the overall performance is verified by a circuit tester to ensure that the noise suppression meets the standard in the frequency band of 10Hz~1MHz, and the response time is ≤50μs.

[0115] The beneficial effects of the above technical solutions are: through the circuit optimization driven by historical data (linear coefficient quantization filtering capability + difference decomposition coefficient positioning failure point), the self-adaptive upgrade of the hardware filter circuit is realized by combining structure fusion, the high-frequency noise attenuation capability of the initial circuit is improved, the harmonic suppression effect is improved, and the noise characteristics of different power parameters can be dynamically adapted, which significantly reduces the noise interference of subsequent data analysis, and lays a foundation for accurate collection of power meter data.

[0116] The application provides a power meter data real-time analysis and abnormal early warning method based on edge computing, and the reconfigurable amplifier, as shown in Figure 2 includes:

[0117] The gain intelligent control module is composed of a signal feature extraction unit and a dynamic decision circuit, the reconfigurable cascade amplification array is composed of a four-stage gain cooperative regulation architecture, the environment adaptive filtering module is composed of a temperature drift compensation type active band-pass filter and an intelligent notch circuit, the intelligent analog-to-digital conversion module is composed of a double-channel synchronous sampling ADC, and the data standardization processing module is composed of a multi-level cache and a high-reliability transmission unit.

[0118] The gain intelligent control module, the reconfigurable cascade amplification array, the environment adaptive filtering module, the intelligent analog-to-digital conversion module and the data standardization processing module are sequentially connected.

[0119] In this embodiment, the signal flow direction is: power meter analog signal→ gain intelligent control (on-demand gain adjustment) → reconfigurable cascade amplification (wide-range amplification) → environment adaptive filtering (anti-temperature drift and interference elimination) → intelligent analog-to-digital conversion (synchronous sampling and digitization) → data standardization processing (caching, isolation and normalization) → output to an edge computing node.

[0120] The above technical solution has the beneficial effects that a high signal-to-noise ratio, high synchronism and high anti-interference standardized digital stream are finally provided for real-time analysis of power meter data, and accurate anomaly detection of an edge computing layer is supported.

[0121] The application provides an edge computing-based real-time analysis and anomaly early warning method for power meter data, and a local anomaly standardized matrix is constructed based on a standardized data stream, including the following steps.

[0122] A power system monitored by a power meter is determined, a standardized matrix is constructed according to standardized data of the power system at each monitoring time point, and the remaining sampling points except the first sampling point and the tail sampling point in each row of the standardized matrix are sequentially selected as second sampling points, wherein the standardized data stream is composed of standardized data of multiple monitoring time points.

[0123] The first difference absolute value is obtained by combining the element values of the two adjacent columns of the second sampling point, and the second difference absolute value is obtained by determining the nearest cycle boundary point from the second sampling point and locking the element value of the corresponding row according to the set state working cycle of the power system.

[0124] The sum of all first difference absolute values and second difference absolute values is obtained, and the step coefficient of the corresponding second sampling point is obtained in combination with the set state working cycle.

[0125] When the step coefficient of the second sampling point is greater than a preset coefficient, a to-be-verified label is set to the second sampling point.

[0126] Otherwise, a normal label is set to the second sampling point.

[0127] According to the label setting result, determine the to-be-verified column proportion of each column vector in the corresponding standardization matrix and the to-be-verified row proportion of each row vector, and compare them with the preset proportion to lock the abnormal row and the abnormal column respectively;

[0128] According to the first number of abnormal indicators corresponding to the abnormal rows involved in each abnormal column, the second number of to-be-verified labels involved in the corresponding abnormal column, and the abnormal factor corresponding to each element value in the corresponding column vector, the abnormal column is subjected to coefficient amplification to form a local abnormal standardization matrix.

[0129] Preferably, the abnormal factor is determined, comprising:

[0130] The clustering analysis is performed on each row vector to obtain the difference coefficient of each clustering result and the corresponding standard range, and the abnormal factor of the corresponding clustering result is determined according to the data trend of the corresponding clustering result.

[0131] In this embodiment, the power system monitored by the power meter is a specific power network (such as a factory power distribution network or a residential area power distribution network) monitored by the power meter in real time, and contains multiple monitoring parameters. For example, the monitoring indicators include A-phase voltage, B-phase voltage, C-phase voltage, A-phase current, B-phase current, C-phase current, total active power, and power factor (a total of 8 items, corresponding to 8 rows of the matrix), which are specifically collected by the sensor array (voltage transformer, current transformer, etc.) of the power meter, with a monitoring period of 1 minute / time and a total of 100 time points (corresponding to 100 columns of the matrix).

[0132] The standardized data of each monitoring time point is the standardized value of 8 power parameters at a single time point (eliminating the dimension and uniformly mapping to the range of 0-1), for example, the standardized data of the 50th time point is [0.92(A-phase voltage), 0.91(B-phase voltage), 0.93(C-phase voltage), 0.75(A-phase current), 0.73(B-phase current), 0.76(C-phase current), 0.88(total active power), 0.95(power factor)].

[0133] The standardized matrix is a two-dimensional matrix of 100 columns by 8 rows, with column index as the monitoring time point (1-100) and row index as the power parameter indicator (1-8), and the element as the standardized data. For example, in the matrix M[8x100], M[3,2] = 0.89 represents the standardized value of C-phase voltage at 2 time points is 0.89.

[0134] The selection of the remaining sampling points (second sampling points) in each column, excluding the first sampling point and the last sampling point, means that from the 100 columns of the matrix 100, the first column (head) and the 100th column (tail) are excluded, and the remaining 98 columns (2-99 columns) are selected as the sample points to be analyzed based on the row and column selection points. That is, each element in the remaining 98 columns and 8 rows of the matrix is analyzed as a sample point for subsequent analysis.

[0135] In this embodiment, the absolute value of the difference between the normalized values of the parameters of the two adjacent columns of the second sampling point (reflecting the fluctuation difference of the adjacent indicators at the same time point) is calculated, such as the first difference absolute value of the 49th column and the 50th column in the corresponding row, and the first difference absolute value of the 50th column and the 51st column in the corresponding row.

[0136] The set state working period of the power system is a typical operation period of the power system (such as a production shift, a load fluctuation period), which is used to locate the period boundary point, for example, the set period of the distribution network is 20 minutes (corresponding to 20 columns of the matrix, because 1 column = 1 minute), that is, every 20 columns is a period boundary (20th, 40th, 60th, 80th, 100th column), specifically: according to the actual working condition of the power system (such as shift change time, equipment start-stop rule), the period is preset and stored as the configuration parameter of the edge node.

[0137] The lock element value (second difference absolute value) of the nearest period boundary point to the remaining sampling point and the corresponding row is to find the nearest period boundary column (such as the 20th column) in the column where the second sampling point is located, and calculate the absolute value of the difference between the normalized values of the parameters in the same row (reflecting the deviation from the period steady state), such as the second sampling point being the 25th column and the nearest boundary point being the 20th column, and the second difference absolute value of the 25th column and the second 29th column of row 1 (A phase voltage) is calculated.

[0138] The step coefficient corresponding to the remaining sampling point obtained in combination with the set state working period is the ratio of the total fluctuation sum to the period, which quantifies the degree of deviation of the sampling point from the steady state (the longer the period, the step coefficient threshold can be appropriately relaxed), and the step coefficient = (average value of first difference sum + second difference sum) / period column number. It should be noted that there may be one nearest period boundary point or two nearest period boundary points corresponding to the second sampling point, therefore, the average value of the second difference sum needs to be determined (only for one second sampling point).

[0139] In this embodiment, the preset coefficient is the critical value for judging whether the sampling point is abnormal, which is obtained from historical abnormal data statistics, and the preset value of the step coefficient is set to 0.02 based on historical fault data (exceeding this value is considered to be possibly abnormal).

[0140] In this embodiment, the label to be verified / normal label is the label of the second sampling point, i.e., when the step coefficient > preset coefficient, the label to be verified (possible anomaly) is marked, otherwise, the normal label is marked, for example, the step coefficient of the 25th column and the 1st row = 0.0175 < 0.02 → normal label; the step coefficient of the 75th column and the 2nd row = 0.025 > 0.02 → label to be verified.

[0141] In this embodiment, the proportion of the column to be verified in each column vector is that the number of the label to be verified contained in each column accounts for the proportion of the total number of rows (8 rows) in the column (reflecting the overall abnormal tendency of the column index) in 100 columns.

[0142] The proportion of the row to be verified in each row vector is that the number of the label to be verified contained in each row accounts for the proportion of the total number of columns (100 columns) in the row (reflecting the overall abnormal tendency of the row index) in 8 rows.

[0143] The preset proportion (column / row) is the critical proportion for judging whether the column / row is abnormal (set by system reliability requirements), for example, the preset row proportion = 10%, the preset column proportion = 30%, the column whose proportion of the column to be verified > the preset column proportion is an abnormal column, and the row whose proportion of the row to be verified > the preset row proportion is an abnormal row.

[0144] The first number of the abnormal index corresponding to the abnormal row involved in the abnormal column is the number of the intersection of the abnormal column and the abnormal row (i.e., the column intersection point of the abnormal row), and the second number of the label to be verified involved in the abnormal column is the total number of all the labels to be verified in the abnormal column (including the labels to be verified in the non-abnormal row).

[0145] The abnormal factor corresponding to each element value in the corresponding column vector is the degree of deviation of each element in the column from the normal range, which is obtained by cluster analysis, specifically: using K-means clustering (k = 2, normal / abnormal class), calculating the Euclidean distance between the element and the center of the normal class as the abnormal factor.

[0146] The abnormal column array is obtained by amplifying the coefficient of the abnormal column, and the abnormal column data is weighted and amplified by the first number, the second number and the abnormal factor, so as to highlight the abnormal characteristics, for example: the amplification formula of the abnormal column 1 = element value × (1 + first number × 0.01 + second number × 0.02 + 0.1 × the abnormal factor corresponding to the element of the corresponding column).

[0147] The local abnormal standardization matrix is a submatrix composed of the amplified results of all abnormal columns, focusing on the parameters and time points with high abnormal risk, for example, the abnormal column = [1, 3], and the amplified values of these two columns are retained in each row → the local matrix is 100 rows × 2 columns.

[0148] In this embodiment, the data in each row vector is subjected to cluster analysis, that is, the data involved in the corresponding row is classified, since the corresponding row is the same index and the standard range under the same index is the same, the difference coefficient can be directly calculated, and the difference coefficient is the deviation degree of the cluster center and the preset standard range (such as the standardized range of normal voltage 0.8-1.0), wherein the absolute difference between the cluster center and the midpoint of the standard range / the midpoint of the standard range is calculated as the difference coefficient.

[0149] According to the data trend of the corresponding clustering result, the abnormal factor of the corresponding clustering result is determined by adjusting the difference coefficient combined with the data trend (upward / downward / fluctuation) to obtain the final abnormal factor, for example, the abnormal group data presents a downward trend (from 0.8 to 0.6) → abnormal factor = difference coefficient x 1.2 (trend intensification weight) = 0.2 x 1.2 = 0.24, specifically: the trend slope of the clustering data is fitted by linear regression, and when the absolute value of the slope is greater than 0.01, the weight (such as x 1.2) is increased.

[0150] The beneficial effects of the above technical solutions are: through the fine analysis (second-order difference calculation, step coefficient screening, abnormal factor amplification) of the standardized matrix, the local abnormal standardized matrix constructed can accurately focus on high-risk parameters and time points, reduce redundant data compared with global matrix analysis, and improve more accurate data basis.

[0151] The application provides a power meter data real-time analysis and abnormal early warning method based on edge computing, which comprises the following steps of:

[0152] A fuzzy logic layer based on the machine learning model maps the environmental parameters into a fuzzy set, calculates the fuzzy membership degree through a gradient membership function, and introduces an environmental gradient factor.

[0153] A neural network layer based on the machine learning model inputs the fault probability distribution output by the model, and combines a device aging factor and the environmental gradient factor to solve the dynamic threshold of the corresponding index.

[0154] FPGA is used to realize fuzzy reasoning and neural network parallel operation, and a threshold conflict detection formula Yc is used to trigger priority arbitration.

[0155]

[0156] Wherein, Th1 and Th2 are historical threshold value and real-time dynamic threshold value respectively; and delta is a conflict determination threshold value.

[0157] When the threshold value is adjusted, the front-end acquisition module is automatically adapted, the ADC switches the anti-aliasing filter mode under high interference, and the amplifier gain is dynamically adjusted based on priority arbitration.

[0158] G(t) = G0·exp(-k·EMI(t))·ln(e-Yc)

[0159] Where, EMI(t) is the electromagnetic interference intensity; G0 is the initial gain of the amplifier; G(t) is the real-time gain of the amplifier; k is the attenuation coefficient; e is a constant, and the value is 2.7;

[0160] The dynamic warning formula of the local display screen is triggered based on the real-time gain warning Encode the failure level, where Co is the RGB color value; ΔTh is the threshold offset; RGB() is a three-channel color function.

[0161] In this embodiment, the fuzzy logic layer processes (environmental parameters → fuzzy set + gradient factor):

[0162] The environmental sensor is a sensing device for collecting environmental quantities such as temperature, humidity, and electromagnetic interference (EMI).

[0163] The environmental parameter → fuzzy set is to map the continuous environmental value (such as 25℃) to the fuzzy language such as “low temperature / medium / high temperature”, and the membership degree is used to describe the attribution, for example, the temperature fuzzy set is defined as [20, 25, 30] (the medium interval), and the membership degree of 28℃ is (30-28) / (30-25) = 0.4 (triangular membership degree formula). Specifically, the fuzzification is realized by Verilog in the FPGA, and the boundary of the membership degree function is pre-stored (such as the temperature interval [0, 20, 30, 50] corresponding to “low temperature / medium / high temperature”).

[0164] The gradient membership function is to calculate the rate of change of the environmental parameter (such as dT / dt = 5℃ / min), and integrate the membership degree to reflect the change trend, for example, the gradient formula ω1 = 1-exp(-0.5·dT / dt), and when dT / dt = 5℃ / min, ω1 ≈ 0.918 (the greater the gradient, the higher the membership degree).

[0165] In this embodiment, the environmental gradient factor is a coefficient (0~1) quantifying the rate of change of the environment, and the greater the gradient, the closer the factor to 1, for example, dT / dt = 5℃ / min corresponds to a factor of 0.8, and dT / dt = 1℃ / min corresponds to a factor of 0.2.

[0166] The neural network layer solves (failure probability + aging factor → dynamic threshold):

[0167] The failure probability distribution is a failure probability vector of each power indicator (voltage, current, etc.) output by a machine learning model (such as LSTM+CNN), for example, the model outputs [A-phase voltage abnormal probability 0.3, B-phase current abnormal probability 0.2].

[0168] The device aging factor is a coefficient (0-1) calculated according to the running time and the number of failures (0-1). The more serious the aging, the closer the factor to 1. For example, running for 10 years and failing 5 times, the formula factor = (10 / 20) x 0.7 + (5 / 10) x 0.3 = 0.5 (design life 20 years, design failure 10 times).

[0169] The dynamic threshold calculation is a combination of failure probability, environmental gradient, and aging factor, and outputs real-time threshold through neural network regression layer. For example, the historical threshold is 220V±5%, the environmental factor is 0.7, the aging factor is 0.6, and the model output adjustment coefficient is 1.2. Then the dynamic threshold is 220±6.6V.

[0170] FPGA parallel and threshold conflict (hardware acceleration + arbitration):

[0171] FPGA parallel operation is that FPGA runs fuzzy inference (hardware acceleration) and neural network (IP core) at the same time, which reduces the delay. For example, Xilinx Zynq FPGA, fuzzy inference uses LUT, neural network generates IP core through HLS, and parallel processing environment and power data.

[0172] The threshold conflict formula Yc is to compare the historical threshold (Th1) and the dynamic threshold (Th2). If the deviation is greater than δ, it is marked as conflict (Yc=1). For example, Th1=220±5V (215-225V), Th2=220±8V (212-228V), δ=3V. If the lower limit of Th2 is 210V, |210-215|=5>3, then Yc=1.

[0173] Priority arbitration is Yc=1, which prioritizes filter switching and gain adjustment to ensure threshold reliability.

[0174] Hardware adaptation (ADC + amplifier dynamic adjustment):

[0175] ADC switching anti-aliasing filter is high interference (such as EMI>60dBμV / m), ADC switches to high sampling rate + strong filter mode, such as normal mode 1kSPS, cutoff 1kHz; anti-aliasing mode 10kSPS, cutoff 500Hz (ADS1278 supports programmable configuration), specifically: ADC receives FPGA instructions through SPI to switch mode in real time.

[0176] The amplifier gain formula G(t) is a combination of EMI(t) and Yc, which dynamically calculates the gain to suppress strong interference saturation. For example, the gain is reduced from 20 to 5.53 to avoid signal saturation.

[0177] The dynamic warning is encoded by a threshold offset (Delta Th), the larger the Delta Th, the brighter the red color and the darker the blue color, for example, Delta Th = 5% (dynamic threshold is 5% higher than history), Co = RGB(255, 0, 0) (pure red, severe warning); Delta Th = 2% -> RGB(102, 0, 153) (purple, mild warning).

[0178] The beneficial effects of the above technical solutions are: through fuzzy logic to capture the environment dynamics, neural network to fuse device aging, FPGA parallel acceleration, and hardware level threshold self-adaption, multi-factor intelligent regulation of abnormal threshold is realized, local RGB warning directly maps the fault level, an intelligent monitoring system of environment-device-algorithm-hardware deep cooperation is constructed, and the accuracy and reliability of real-time analysis of power meters are ensured.

[0179] The application provides a power meter data real-time analysis and abnormal warning method based on edge computing, which compares with a dynamic threshold to generate an abnormal signal, and comprises the following steps:

[0180] A feature pair of the same index is extracted from the local power parameter feature vector and the global power parameter feature vector, numerical analysis is performed on the feature pair to obtain a numerical pair, and the numerical pair is compared with a corresponding dynamic threshold;

[0181] Yx = {|VJ-Th2|>epsilon1, |VQ-Th2|>epsilon2}

[0182] Wherein, VJ and VQ are local feature values and global feature values in the feature pair respectively; epsilon1 and epsilon2 are local conflict thresholds and global conflict thresholds matched with the numerical pair respectively; Yx is a comparison result;

[0183] The comparison result is matched with a result-index-signal matching table to obtain an abnormal signal of the corresponding index.

[0184] In this embodiment, the feature pair of the same index is a pair composed of feature values of the same power index in the local and global vectors (such as local A-phase voltage fluctuation vs. global A-phase voltage fluctuation), for example, local A-phase voltage fluctuation 3% (denoted as VJ) and global A-phase voltage fluctuation 1.8% (denoted as VQ) form a feature pair (3%, 1.8%).

[0185] In this embodiment, numerical analysis is to convert physical quantities (such as percentage, Hz) of the feature pair into calculable scalars (such as 3% -> 0.03), for example, the feature pair (3%, 1.8%) is analyzed as (0.03, 0.018), specifically: the edge node software outputs a floating point number pair through string segmentation + unit conversion (such as "3%" -> float("3") / 100).

[0186] In this embodiment, the numerical pair is the numerical pair (VJ, VQ) composed of the analyzed local and global characteristic values.

[0187] In this embodiment, the dynamic threshold (Th2) is the real-time abnormality judgment threshold calculated (such as the maximum voltage fluctuation allowed 2%→0.02, dynamically updated with the environment and device aging).

[0188] In this embodiment, the allowed deviation of local / global characteristic values and dynamic threshold (ε1>ε2, because local device fluctuation is more frequent and global power grid is more sensitive), specifically: ε1=0.005 (0.5%, local allows a larger deviation), ε2=0.003 (0.3%, global deviation is more stringent), set according to historical fault data statistics (such as analyzing 1 year of fault cases to find the threshold with the lowest false positive rate), and stored as a configurable parameter (such as ε1=0.5% corresponding to hardware register value 0x0050).

[0189] The comparison result (Yx) is to determine whether |VJ-Th2|>ε1 or |VQ-Th2|>ε2 is true:

[0190] If true, Yx=1 (abnormal); otherwise, Yx=0 (normal).

[0191] For example: VJ=0.03, Th2=0.02, ε1=0.005→|0.03-0.02|=0.01>0.005→local trigger; VQ=0.018, Th2=0.02, ε2=0.003→|0.018-0.02|=0.002<0.003→global not triggered; because of local trigger, Yx=1.

[0192] Specifically: hardware level: FPGA calculates in parallel through subtractor+comparator, delay≤1μs;

[0193] Software level: edge node CPU uses conditional judgment statement (such as C language if(fabs(VJ-Th2)>ε1||fabs(VQ-Th2)>ε2)Yx=1;).

[0194] The result-index-signal correspondence table is the mapping table of Yx, abnormal type, and signal code, as shown in Table 1, (standardized output, convenient for subsequent module analysis).

[0195] Table 1 Mapping table

[0196] Yx Indicator Type Abnormal Signal Code 1 Voltage Fluctuation Abnormality 0x01 0 Normal 0x00

[0197] In this embodiment, the abnormal signal is the abnormal identification matched by the correspondence table (including type, level, and signal code, used to trigger alarm / upload), such as {"type":"voltage fluctuation abnormality","level":2,"code":0x01}.

[0198] Specifically: hardware output: output 3.3V high level through GPIO (duration is bound with level);

[0199] Data upload: packaged as an MQTT message and sent to the cloud through a physically isolated module.

[0200] The beneficial effects of the above technical solution are: through two-dimensional verification of local details and global trends (differentiated threshold + hardware-level parallel comparison), the scheme breaks through the limitations of traditional single-point feature detection, realizes cross-scale accurate identification of anomalies (coordinated judgment of device-level fluctuations and power grid-level trends), and guarantees output consistency through a standardized signal reference table, provides clear basis for subsequent hierarchical early warning and fault location, and supports intelligent upgrading of the power system from passive response to active early warning.

[0201] The present application provides a kind of based on edge computing's electric power meter data real-time analysis and abnormal early warning method, according to the type and level of abnormal signal, through relay control circuit triggers sound light alarm, and sends early warning instruction to cloud platform through physically isolated communication module, including:

[0202] According to the monitoring type to which the index of the abnormal signal belongs and the signal level of the abnormal signal Control relay circuit output differentiating sound light signal, wherein min is minimum symbol;Max is maximum symbol;δ0 is basic classification threshold; It is a down rounding symbol;

[0203] Based on the sub-instruction including the monitoring type to which the index belongs, signal level, occurrence timestamp and associated environmental parameters;

[0204] According to the instruction structure and instruction trigger condition of the sub-instruction, determine the enable state, and according to the enable state, determine the clock enable signal of the corresponding early warning instruction;

[0205] According to the monitoring type to which the sub-index belongs, determine the display address on the cloud platform, and generate a verification signal in combination with the frequency conversion segment determined by the clock enable signal of the corresponding sub-instruction, and perform error checking on the sub-instruction relying on the verification signal;

[0206] When there is an error, the sub-instruction is cleared and a new enable signal is determined relying on the verification signal;

[0207] Relying on the new enable signal, adjust the clock enable signal of the corresponding sub-instruction in the order number of all original sub-instructions to obtain a replacement enable signal;

[0208] If there is no error, the clock enable signal of the corresponding sub-instruction is retained;

[0209] The last clock enable signal is clocked to determine the execution probability of each clock to get the early warning instruction, and the encrypted early warning instruction is sent to the cloud platform through the physically isolated communication module.

[0210] In this embodiment, the relay control circuit is composed of a solid-state relay (such as G3MB-202P) and a driving chip (such as ULN2003), which controls the on-off of the sound-light device and differentiates the sound-light:

[0211] Level 1 (mild): green LED (2Hz flashing) + 1kHz buzzer (intermittent sound);

[0212] Level 3 (severe): red LED (10Hz flashing) + 8kHz buzzer (continuous sound).

[0213] Level 2 triggers yellow LED (5Hz flashing) + 3kHz buzzer (for 2 seconds).

[0214] In this embodiment, the composition of the physically isolated communication module is:

[0215] Optical isolation: TLP281 optocoupler (isolation voltage 2500V) is used to cut off the electrical connection between the power side and the communication side;

[0216] Electromagnetic shielding: external ferrite magnetic ring + shielding cover, suppress common mode interference.

[0217] In this embodiment, the sub-instruction construction is:

[0218] Sub-instruction content: includes index type (such as "voltage fluctuation"), signal level (such as 2), timestamp (such as "2025-07-19 10:05:00"), and associated environmental parameters (such as {"temperature": 35℃, "EMI": 60dBμV / m}).

[0219] Enable state is whether the sub-instruction triggers sending (such as level ≥2 enables, level 1 only local alarm), for example, level 2→enable (1), level 1→not enable (0).

[0220] Clock enable signal is a timing signal that controls the timing of instruction sending (frequency and level are bound, level 3→100Hz, level 2→50Hz, level 1→10Hz), specifically: FPGA generates PWM wave (such as level 3 corresponds to 100Hz square wave) as the trigger clock of instruction sending.

[0221] Display address is the display location of the cloud platform (such as "region A-switch B-voltage fluctuation"), for example, through the configuration file mapping: {voltage fluctuation:dashboard / regionA / switchB / voltage}.

[0222] The frequency conversion section is a frequency interval (e.g., level 3 corresponds to 100-200 Hz) of the clock enable signal, and is used to generate a check signal, specifically: the edge node converts the frequency by lookup table method (e.g., level→frequency array: freq_map[3]=150).

[0223] In this embodiment, the check signal is a check code based on the frequency conversion section (e.g., 150 Hz corresponds to check bit 0xAA, generated by XOR operation).

[0224] The error check is a comparison of the check bit of the sub-instruction and the check signal, and if they are inconsistent, it is determined to be an error.

[0225] For example, the sub-instruction check bit is 0xAA, and the actual calculation is 0xAB→determined to be an error.

[0226] The new enable signal is the enable signal regenerated when there is an error (e.g., retransmit after 1 second delay, marked as "retry"), which is specifically implemented by a counter+timer (e.g., retransmission times≤3 times).

[0227] The replacement enable signal is the clock enable signal of the adjusted error instruction (e.g., insert the instruction of level 2 in front of the queue of level 3).

[0228] The execution probability is the transmission probability allocated according to the clock order (level 3 accounts for 70%, level 2 accounts for 20%, and level 1 accounts for 10%), for example, among 10 sub-instructions, 7 are sent preferentially in level 3.

[0229] The encrypted early warning instruction is a sub-instruction encrypted by AES-128 (the key is stored in the security chip of the edge node).

[0230] The beneficial effects of the above technical solutions are: a closed loop is constructed by hierarchical sound and light+secure communication+fault-tolerant mechanism: differential sound and light realizes second-level identification, physical isolation+encryption guarantees anti-interference reliability, the check and retransmission mechanism solves the transmission packet loss, and finally the delivery rate of the early warning instruction is improved, the fault response speed is shortened, and the intelligent early warning capability of the power system operation and maintenance is provided, which is visible, manageable and controllable.

[0231] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application also intends to include these modifications and variations.

Claims

1. A method for real-time analysis and anomaly early warning of power meter data based on edge computing, characterized in that, include: Step 1: The multi-channel analog-to-digital converter integrated on the edge computing node acquires multi-modal data from the power meter, uses a hardware filtering circuit to eliminate high-frequency noise, and performs adaptive signal gain adjustment based on a reconfigurable amplifier to generate a standardized data stream. Step 2: Use a digital signal processing accelerator to perform a fast Fourier transform on the global normalized matrix and the local anomaly normalized matrix constructed from the normalized data stream, and combine it with a multi-core processor to perform calculations and analysis to generate the corresponding local power parameter feature vectors and global power parameter feature vectors. Step 3: Based on real-time collection of environmental parameters by environmental sensors, combined with a machine learning model trained on historical data, the anomaly detection threshold is dynamically adjusted, and the feature vectors of local power parameters and global power parameters are compared with the dynamic threshold to generate anomaly signals; Step 4: Based on the type and level of the abnormal signal, trigger an audible and visual alarm through the relay control circuit, and send a warning command to the cloud platform through the physically isolated communication module. At the same time, start the redundant power supply module to control the power meter to continue operating.

2. The method for real-time analysis and anomaly early warning of power meter data based on edge computing according to claim 1, characterized in that, Before using hardware filtering circuits to eliminate high-frequency noise, the following steps are included: Several sets of historical data are acquired and sequentially input into the initial filtering circuit to obtain historical filtered data. A double array of noise not eliminated and noise eliminated for the same indicator is obtained. The double array is input into the linear analysis model to determine the linear coefficient of noise elimination. At the same time, the filter difference data of each historical filtered data is determined and input into the difference analysis model. Combined with the interdependence between indicators, the difference decomposition coefficient of each indicator is obtained. Based on the linear coefficients and difference decomposition coefficients, the failure problem of the initial filter circuit is analyzed; Based on the failure problem, an improved structure for the initial filter circuit is determined, and the improved structure is integrated with the initial filter circuit to obtain a hardware filter circuit.

3. The method for real-time analysis and anomaly early warning of power meter data based on edge computing according to claim 1, characterized in that, The reconfigurable amplifier includes: The system includes a gain intelligent control module consisting of a signal feature extraction unit and a dynamic decision circuit; a reconfigurable cascaded amplifier array consisting of a four-level gain collaborative adjustment architecture; an environmental adaptive filtering module consisting of a temperature drift compensation active bandpass filter and an intelligent notch filter circuit; an intelligent analog-to-digital conversion module consisting of a dual-channel synchronous sampling ADC; and a data standardization processing module consisting of a multi-level buffer and a high-reliability transmission unit. The intelligent gain control module, the reconfigurable cascaded amplifier array, the environmental adaptive filtering module, the intelligent analog-to-digital conversion module, and the data standardization processing module are connected in sequence.

4. The method for real-time analysis and anomaly early warning of power meter data based on edge computing according to claim 1, characterized in that, Constructing a local anomaly normalization matrix based on normalized data flow, including: The power system monitored by the power meter is determined, and a standardization matrix is ​​constructed based on the standardized data of the power system at each monitoring time point. From each row of the standardization matrix, the remaining sampling points except for the first sampling point and the last sampling point are selected in sequence and regarded as the second sampling points. The standardized data stream is composed of standardized data from multiple monitoring time points. The first absolute value of the difference is taken by combining the two adjacent columns of element values ​​of the second sampling point. At the same time, according to the set state working cycle of the power system, the nearest cycle boundary point to the second sampling point is determined and the locked element value of the corresponding row is locked and the second absolute value of the difference is taken. Obtain the sum of the absolute values ​​of all first and second differences, and combine this with the set state working cycle to obtain the step coefficient of the corresponding second sampling point; When the step coefficient of the second sampling point is greater than the preset coefficient, a verification label is set for the second sampling point. Otherwise, set a normal label for the second sampling point; Based on the label setting results, determine the proportion of columns to be verified and the proportion of rows to be verified in each column vector of the corresponding standardized matrix, and compare them with the preset proportions to lock out abnormal rows and columns. Based on the first number of abnormal indicators corresponding to the abnormal rows involved in each abnormal column, the second number of labels to be verified involved in the corresponding abnormal column, and the abnormal factor corresponding to each element value in the corresponding column vector, the abnormal column is subjected to coefficient amplification to form a local abnormality standardization matrix.

5. The method for real-time analysis and anomaly early warning of power meter data based on edge computing according to claim 4, characterized in that, Identify outliers, including: Cluster analysis is performed on each row vector to obtain the difference coefficient between each cluster result and the corresponding standard range, and the outlier factors of the corresponding cluster result are determined according to the data trend of the corresponding cluster result.

6. The method for real-time analysis and anomaly early warning of power meter data based on edge computing according to claim 1, characterized in that, Based on real-time environmental parameter acquisition by environmental sensors and combined with a machine learning model trained on historical data, the anomaly detection threshold is dynamically adjusted, including: The fuzzy logic layer based on the machine learning model maps environmental parameters to fuzzy sets, calculates fuzzy membership through the gradient membership function, and introduces an environmental gradient factor. Based on the fault probability distribution of the neural network layer input model of the machine learning model, and combined with the equipment aging factor and environmental gradient factor, the dynamic threshold of the corresponding index is calculated. Implement fuzzy inference and parallel neural network computation based on FPGA, and integrate the threshold conflict detection formula Yc to trigger priority arbitration: Where Th1 and Th2 are the historical threshold and the real-time dynamic threshold, respectively; δ is the conflict determination threshold. When adjusting the threshold, the system automatically adapts to the front-end acquisition module. Under high interference, the ADC switches to anti-aliasing filtering mode, and the amplifier gain is dynamically adjusted based on priority arbitration. G(t)=G0·exp(-k·EMI(t))·ln(e-Yc) Where EMI(t) is the electromagnetic interference intensity; G0 is the initial gain of the amplifier; G(t) is the real-time gain of the amplifier; k is the attenuation coefficient; and e is a constant with a value of 2.

7. Dynamic alert formula based on real-time gain warning triggering local display screen Encode fault level, where Co is the RGB color value; ΔTh is the threshold offset; RGB() is the three-channel color function.

7. The method for real-time analysis and anomaly early warning of power meter data based on edge computing according to claim 6, characterized in that, An anomalous signal is generated by comparing it with a dynamic threshold, including: Extract feature pairs of the same index from the local power parameter feature vector and the global power parameter feature vector, perform numerical analysis on the feature pairs to obtain numerical pairs, and compare the numerical pairs with the corresponding dynamic thresholds respectively. Yx={|VJ-Th2|>ε1,|VQ-Th2|>ε2} Where VJ and VQ are the local and global feature values ​​of the feature pair, respectively; ε1 and ε2 are the local and global conflict thresholds for matching the numerical pair, respectively; and Yx is the comparison result. The comparison results are matched with the result-indicator-signal comparison table to obtain the abnormal signals of the corresponding indicators.

8. The method for real-time analysis and anomaly early warning of power meter data based on edge computing according to claim 7, characterized in that, Based on the type and level of the abnormal signal, an audible and visual alarm is triggered via a relay control circuit, and a warning command is sent to the cloud platform via a physically isolated communication module, including: Based on the monitoring type of the abnormal signal and the signal level of the abnormal signal. The control relay circuit outputs differentiated acoustic and optical signals, where min represents the minimum value, max represents the maximum value, and δ0 represents the basic hierarchical threshold. The floor symbol; Based on sub-instructions that include the monitoring type of the indicator, the signal level, the occurrence timestamp, and the associated environmental parameters; The enable state is determined based on the instruction structure and instruction triggering conditions of the sub-instruction, and the clock enable signal of the corresponding warning instruction is determined based on the enable state. The display address on the cloud platform is determined according to the monitoring type of the sub-indicator, and a verification signal is generated by combining the frequency conversion segment determined by the clock enable signal of the corresponding sub-instruction. The sub-instruction is then checked for errors based on the verification signal. When an error occurs, the sub-instruction is cleared and a new enable signal is determined based on the verification signal; Based on the sequence number of the new enable signal in the clock enable signals of all the original sub-instructions, the clock enable signals of the corresponding sub-instructions are adjusted to obtain the replacement enable signal. If there is no error, the clock enable signal of the corresponding sub-instruction will be retained; The final clock enable signal is sorted by clock sequence to determine the execution probability of each clock and obtain an early warning command. The encrypted early warning command is then sent to the cloud platform through a physically isolated communication module.

Images