Method for identifying safe and abnormal electricity consumption of electric energy meter

Abstract

The invention discloses a method for identifying safe and abnormal electricity consumption of an electric energy meter, and aims to solve the problems that temperature rise of a wire and a contact is difficult to identify in time and the early stage of fire is easy to miss under the condition that a temperature sensor is not added. According to the method, voltage and current are collected synchronously, the ratio of voltage variation to current variation is extracted based on current step to obtain equivalent series resistance and confidence, baseline resistance and environment proxy temperature are estimated in a low-load section, and a temperature sensor including wire temperature and temperature sensor temperature is constructed. A triple physical constraint ordinary differential equation of contact temperature and damage factors is adopted, assimilation observation is carried out, contact temperature slope and resistance drift are fused, and online change point identification and dynamic threshold grading early warning are carried out, so that advanced identification of an abnormal electricity-induced temperature rise trend is realized only by depending on voltage and current data of an electricity meter; the early warning advance is improved; the false alarm is reduced; and the robustness and the interpretability are enhanced.

Description

Technical Field

[0001] This invention relates to the field of electrical safety monitoring technology, and in particular to a method for identifying abnormal electricity usage in electricity meters. Background Technology

[0002] With the widespread adoption of smart grids and smart meters, electricity meters are equipped with the ability to synchronously collect voltage and current data under a unified sampling clock, and are gradually taking on the function of monitoring electricity safety. Existing technologies mainly include: installing temperature sensors at the meter or distribution node and alarming based on over-temperature thresholds or temperature rise rates; using devices such as arc fault detection and leakage protection for safety protection; and monitoring power, harmonics, and waveform characteristics based on voltage and current data, or estimating the equivalent resistance of the line by the ratio of voltage to current changes during current step events and identifying anomalies using statistical change points or machine learning methods.

[0003] However, the above solutions still have shortcomings:

[0004] First, the solution of adding a temperature sensor is costly, difficult to modify, and has a delayed alarm, which is often triggered only when the temperature has exceeded the safety threshold, making it difficult to detect early signs of slow heating of wires and contacts in a timely manner.

[0005] Second, sensorless resistance estimation methods are sensitive to nonlinear loads, harmonics, noise and power grid fluctuations, and lack robust modeling of ambient temperature and baseline drift, resulting in unstable estimations and a large number of false alarms and false misses.

[0006] Third, existing algorithms mostly use static thresholds, pure statistical variable points, or general machine learning models, lacking thermal and electrical coupling mechanisms, thermal hysteresis and energy consistency constraints, and lack observation-based online correction and dynamic thresholds that adapt to load intensity and degradation degree, making it difficult to stably provide graded early warnings in complex scenarios.

[0007] Therefore, a method for identifying abnormal electricity usage in electricity meters that can overcome the shortcomings of the existing technology is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of this invention is to propose a method for identifying abnormal electricity usage in electricity meters. Addressing the problems of existing technologies, such as difficulty in timely identification of slow temperature rises in conductors and contacts without adding temperature sensors, susceptibility to noise and environmental drift interference, static thresholds, and high false alarm / missed alarm rates, this invention proposes a technical solution that involves synchronously acquiring voltage and current under a unified clock; estimating the equivalent series resistance based on the ratio of voltage change to current change using current step changes and constructing confidence and step strength; establishing baseline resistance and ambient proxies at low load levels; assimilating the data using a neural network constant differential equation incorporating conductor temperature, contact temperature, and damage factors, and applying monotonicity, energy conservation, and hysteresis consistency constraints; and fusing contact temperature slope and resistance drift for online change point identification and dynamic threshold-based early warning. This invention achieves the technical effects of early identification of temperature rise trends without adding sensors, improving the reliability and interpretability of early warnings, and adapting to complex loads.

[0009] A method for identifying abnormal electricity consumption in an electricity meter according to an embodiment of the present invention includes:

[0010] S1. Synchronously acquire and align voltage and current sequences under a unified sampling clock, and establish corresponding time indices;

[0011] S2. Identify current step segments based on the current sequence, calculate the local equivalent series resistance value in each segment according to the ratio of voltage change to current change, form an observation resistance value sequence in time order, and construct the observation confidence sequence and load step intensity sequence corresponding to the segment.

[0012] S3. Determine the low load period based on the current sequence, obtain the baseline resistance sequence by statistical analysis, subtract the baseline resistance from the observed resistance to obtain the observed resistance drift sequence, and smooth the baseline resistance with a low-pass filter to obtain the ambient temperature sequence.

[0013] S4. Drive the neural ordinary differential equation model containing the hidden states of conductor temperature, contact temperature and damage factor with the current sequence and the environmental proxy temperature sequence. Its hidden state derivative consists of the heat generation term of the square of the current, the heat dissipation term relative to the environmental proxy temperature and the thermal memory kernel convolution term. Map the hidden state to the equivalent series resistance value to obtain the model calculated resistance value sequence, and apply monotonic, energy conservation and hysteresis consistency constraints. Assimilate according to the observed resistance value sequence and its confidence level, and update the above hidden state and the model calculated resistance value sequence.

[0014] S5. The contact temperature slope sequence is obtained by differential and smoothing from the hidden state of the contact temperature. The contact temperature slope sequence is then adaptively weighted according to the load step intensity and aligned and fused with the observed resistance drift sequence to form a joint trend evidence sequence.

[0015] S6. Calculate the posterior sequence of the change point based on the joint trend evidence sequence, and dynamically adjust the threshold by combining the load step intensity and the latent state of the damage factor. When the posterior continuously exceeds the threshold in multiple adjacent sliding windows, output the graded early warning result sequence.

[0016] Optionally, step S1 specifically includes:

[0017] Synchronous sampling is performed based on the unified sampling clock of the electricity meter. Voltage timing is collected at fixed sampling intervals to form a voltage sequence, and current timing is collected to form a current sequence.

[0018] A time index sequence is generated based on the fixed sampling interval, so that the time index sequence corresponds one-to-one with the voltage sequence and the current sequence;

[0019] Data quality control and preprocessing are performed on voltage and current sequences, including DC bias correction, range saturation check, noise reduction, and interpolation completion for single-point missing samples caused by loss.

[0020] Output preprocessed voltage, current and time index sequences.

[0021] Terminology definition:

[0022] The unified sampling clock is provided by the energy meter and simultaneously drives the two sampling channels of voltage and current. Its sampling period remains constant, so that the sampling trigger of the two channels shares the same timing reference.

[0023] The synchronous acquisition refers to sampling voltage and current at the same trigger moment of a unified sampling clock, so that the two samples correspond to each other in time, which can also be described as synchronous sampling.

[0024] The alignment process involves registering the voltage and current sequences according to their time indices, including eliminating channel start offsets and fixed delays so that the same time index corresponds to the voltage and current at the same physical moment.

[0025] The time index is a time coordinate that corresponds one-to-one with each sampling point. It is generated by incrementing at a fixed sampling interval from the start time and is used to identify the sampling time and support sequence alignment.

[0026] The data quality control and preprocessing is a quality check and standardization process implemented on the sampled sequence to remove or correct anomalies and suppress noise, including at least DC bias correction, range saturation check, noise reduction processing and interpolation completion for single-point missing samples.

[0027] The DC bias correction involves estimating the average bias of a voltage or current sequence within a selected time window and subtracting it from the sequence to eliminate the DC component introduced by the measurement link.

[0028] The range saturation check is used to detect whether the sampled value reaches or approaches the upper or lower limit of the measurement range and to identify the resulting clipping distortion. Saturated segments can be marked or removed.

[0029] The denoising process is a process that suppresses random noise and narrowband interference while preserving the key components of the signal. It can be achieved by low-pass filtering, band-stop / notch filtering, or adaptive filtering.

[0030] The single-point missing sample is a missing sampling point that appears isolated on the time index, with a missing continuous length of 1 and valid values ​​in the adjacent samples on both sides.

[0031] The interpolation completion is a process of estimating the value of the missing point based on the time index and numerical relationship of the valid samples before and after the missing point. It can be implemented by linear interpolation, spline interpolation or polynomial interpolation.

[0032] Optionally, step S2 specifically includes:

[0033] Based on the current sequence and time index sequence, current step segments are identified according to the absolute value of the current difference between adjacent samples exceeding a preset step threshold and lasting for no less than a preset holding time in consecutive samples. A start index and an end index are determined for each current step segment.

[0034] Within each current step segment, the voltage and current sequences are averaged in short windows near the start and end indices to obtain the start voltage, start current, end voltage, and end current.

[0035] The difference between the ending voltage and the starting voltage is used as the voltage change, and the difference between the ending current and the starting current is used as the current change. When the absolute value of the current change is lower than the minimum effective change threshold, the current step segment is removed.

[0036] For the retained current step segment, the local equivalent series resistance value of the segment is obtained by dividing the voltage change by the current change value, and the center time of the segment is mapped to the time index sequence and arranged in time order to form the observed resistance value sequence.

[0037] For each observed resistance value, the observation confidence level is constructed based on the absolute value of the current change in that segment, the segment length, and the fitting residuals of the voltage and current samples within the segment relative to the start and end points of the linear connection, thus forming an observation confidence level sequence.

[0038] The absolute value of the current change is used as the load step intensity to form a load step intensity sequence;

[0039] Output the observed resistance sequence, the observed confidence sequence, and the load step intensity sequence.

[0040] Terminology definition:

[0041] The current step segment is a continuous sample segment in time index, in which the absolute value of the current difference between adjacent samples within the segment continuously exceeds a preset step threshold for at least a preset holding time.

[0042] The preset step threshold is an amplitude threshold used to determine whether the current difference between adjacent samples constitutes a step. It is measured in current units and can be set fixedly or adaptively according to the noise level.

[0043] The preset holding time is the shortest continuous duration used to confirm the continuity of a step, which can be represented by the number of sampling points or the length of time, in order to avoid instantaneous spikes being misidentified as step jumps;

[0044] The starting index is the starting position of the current step segment on the time index, referring to the first time index contained in the segment;

[0045] The end index is the end position of the current step segment on the time index, referring to the last time index contained in the segment;

[0046] The short window is a preset window set around the neighborhood of the start index or end index for estimating the average value of the start / end voltage and current. Its length is relatively small compared to the segment length to suppress noise and transient effects.

[0047] The minimum effective change threshold is a lower limit set for the absolute value of the current change. When the absolute value of the current change is less than this threshold, it is considered that the change is insufficient for reliable resistance estimation and is therefore rejected.

[0048] The local equivalent series resistance is an estimated value of the equivalent series resistance obtained by dividing the voltage change by the current change within a single current step segment, and is used to characterize the overall series resistance of the line at the center moment of the segment.

[0049] The center time of the segment represents the position of the current step segment moment, which can be the arithmetic mean of the times corresponding to the start index and the end index or an equivalent center positioning method.

[0050] The observed resistance sequence is a sequence formed by mapping the local equivalent series resistance values ​​calculated from each current step segment to a time index according to their center time and arranging them in chronological order.

[0051] The segment length is the duration of the current step segment coverage, which can be expressed as the number of sample points or the time length.

[0052] The starting and ending points of the linear connection are straight line reference trajectories established with the mean of the starting window and the mean of the ending window as endpoints, used to describe the ideal linear changes within the segment;

[0053] The fitting residual is a measure of the deviation of the voltage and current samples within the segment relative to the straight line reference. It can be expressed as absolute deviation, squared deviation, or robust statistics. The larger the value, the stronger the nonlinearity or noise within the segment.

[0054] The observation confidence level is a weighted index that measures the reliability of the observation of the local equivalent series resistance value. It increases monotonically with the increase of the absolute value of the current change and the segment length, and decreases monotonically with the increase of the fitting residual. It can be normalized as needed for subsequent weighting processing.

[0055] The observation confidence sequence is a sequence of observation confidence values ​​that corresponds one-to-one with the observation resistance values ​​and is arranged in chronological order.

[0056] The load step strength is an index characterizing the magnitude of load change, and is taken as the absolute value of the corresponding current change.

[0057] The load step intensity sequence is a sequence of load step intensity values ​​arranged in chronological order of the center time of the current step segment.

[0058] Optionally, step S3 specifically includes:

[0059] Based on the observed resistance sequence and the observed confidence sequence, as well as the current sequence and the time index sequence, the low load period is determined. The low load period is a time index segment in which the current sequence is lower than a preset low load threshold and the continuous time is not less than a preset minimum duration. Time index segments adjacent to current step segments are excluded to avoid transient interference.

[0060] During each low-load period, sliding statistics are performed on the observed resistance value sequence, and the weighted median or weighted mean is calculated using the observed confidence sequence as the weight to obtain the baseline resistance value for that period.

[0061] The baseline resistance values ​​of each time period are concatenated into a baseline resistance value sequence according to the time index sequence. For time points that are not in low load periods, the corresponding baseline resistance values ​​are obtained by interpolation through the time index based on the baseline resistance values ​​of the adjacent low load periods, so that the baseline resistance value sequence is continuous in the entire time index sequence.

[0062] The observed resistance drift sequence is formed by subtracting the baseline resistance sequence for the same time period from the observed resistance sequence.

[0063] Based on low-pass smoothing of the baseline resistance sequence by applying a sliding window moving average, the slow-change component is extracted and defined as an environmental proxy temperature sequence by time index sequence;

[0064] Output baseline resistance sequence, observed resistance drift sequence, and ambient proxy temperature sequence.

[0065] Terminology definition:

[0066] The low-load period is a continuous sample in the time index, in which the amplitude of the current sequence does not exceed the preset low-load threshold and the duration is not less than the preset minimum duration, and time index segments adjacent to the current step segment are excluded to avoid transient interference.

[0067] The preset low load threshold is the upper limit of the current amplitude used to determine the low load state. It is measured in current units and can be set fixedly or adaptively based on background noise and base load level.

[0068] The preset minimum duration is the shortest continuous duration used to confirm the validity of the low load state, which can be represented by the number of sampling points or the time length.

[0069] The time index segment is a continuous interval formed by adjacent sampling points on the time index sequence, used to identify a continuous time period;

[0070] The segment adjacent to the current step segment is a segment that is directly connected to or shares a boundary with the current step segment in the time index, and is used to define the adjacent area that needs to be excluded.

[0071] The sliding statistics are a process of setting a statistical window during low-load periods and moving it step by step along the time index, and calculating the representative value of the observed resistance sample in each window according to a specified statistical method.

[0072] The weighted median is a median estimate with the observed confidence sequence as the weight, and its value is such that the cumulative weight of samples not greater than this value and samples not less than this value does not exceed half of the total weight.

[0073] The weighted mean is an arithmetic mean estimate with the observed confidence sequence as weights, which is equal to the sum of the products of the sample and its weights divided by the sum of the weights.

[0074] The baseline resistance is the equivalent series resistance benchmark value obtained by weighted median or weighted mean during a single low-load period, used to characterize the resistance level under background and environmental conditions.

[0075] The baseline resistance sequence is obtained by splicing the baseline resistance values ​​of each low-load period according to their corresponding time index, and filling the time points that are not in the low-load period by interpolation of the time index, so that a continuously defined sequence is obtained on the entire time index.

[0076] The time index interpolation is an interpolation method that estimates the baseline resistance value for time points not within low-load periods based on the baseline resistance value and time position of adjacent low-load periods. It can be implemented using linear interpolation or spline interpolation.

[0077] The observed resistance drift sequence is a difference sequence obtained by subtracting the baseline resistance sequence from the observed resistance sequence over the same time period, used to characterize the resistance change relative to the baseline.

[0078] The sliding window moving average is a smoothing method that averages the sequence within a fixed or adaptive length time window and moves it gradually with the time index.

[0079] The low-pass smoothing is a smoothing process that suppresses high-frequency fluctuations in the sequence while preserving slow-changing trends. It can be achieved through sliding window moving average, etc.

[0080] The slow-change component is a low-frequency trend part obtained by applying low-pass smoothing to the baseline resistance sequence, which is used to reflect the background influence of slow changes.

[0081] The environmental proxy temperature sequence is a sequence defined by time indexing based on the slow variation component of the baseline resistance sequence, and is used as a proxy quantity for the influence of environmental temperature to participate in subsequent model driving and compensation.

[0082] Optionally, step S4 specifically includes:

[0083] Based on the thermoelectric dynamics driven by the current sequence, time index sequence, and environmental proxy temperature sequence, a neural ordinary differential equation model is established, which includes the hidden state sequence of conductor temperature, the hidden state sequence of junction temperature, and the hidden state sequence of damage factor.

[0084] The time derivative of the hidden state is represented as a combination of a heat-generating term corresponding to the square of the current sequence, a heat-dissipating term relative to the ambient proxy temperature sequence, and a convolution term of the thermal memory kernel function, wherein the thermal memory kernel function is composed of a set of exponentially decaying kernels to characterize thermal hysteresis.

[0085] The resistance sequence is calculated based on the mapping relationship from the hidden state to the equivalent series resistance. The mapping relationship includes the base resistance, the temperature coefficient term related to the hidden state sequence of the conductor temperature, and the contact resistance term related to the hidden state sequence of the contact temperature and the hidden state sequence of the damage factor.

[0086] The model is subject to three physical consistency constraints, including a monotonic constraint that the calculated resistance sequence does not decrease when the hidden sequence of contact temperature or damage factor increases; an energy conservation constraint that the cumulative input of the square of the current sequence remains consistent with the increase of the hidden sequence of conductor temperature and contact temperature within any time window; and a hysteresis consistency constraint that the peak values ​​of the calculated resistance sequence and the hidden sequence of contact temperature lag reasonably with respect to the peak value of the square of the current sequence after the load increases.

[0087] Based on the assimilation of the observed resistance sequence and the observed confidence sequence on the time index sequence, the residuals of the resistance sequence and the observed resistance sequence are calculated by the weighted minimization model to update the hidden sequence of conductor temperature, hidden sequence of contact temperature and hidden sequence of damage factor, and the model is updated to calculate the resistance sequence simultaneously.

[0088] The output conductor temperature hidden sequence, the contact temperature hidden sequence, the damage factor hidden sequence, and the model calculated resistance sequence.

[0089] Terminology definition:

[0090] The aforementioned ordinary differential equation model is a model that uses trainable parameters to characterize the continuous evolution of the hidden state over time. Its inputs are the current sequence and the environmental proxy temperature sequence, and its outputs are the hidden state sequence and the resistance value sequence calculated by the model.

[0091] The latent states are state variables that are not directly observable within the model but are driven by inputs and can be corrected online through assimilation, including the latent states of conductor temperature, contact temperature, and damage factors.

[0092] The hidden state sequence of conductor temperature is a hidden state trajectory that describes the change of conductor body temperature over time, and is used to reflect the heat generation and heat dissipation behavior of the conductor.

[0093] The contact temperature latent state sequence is an implicit state trajectory that depicts the temperature change of electrical contact points over time, used to reflect the local thermal behavior of the contact interface.

[0094] The hidden state sequence of the damage factor is a dimensionless state trajectory that characterizes the degree of deterioration of the material or contact interface, and its value is monotonically unchanging as the deterioration intensifies.

[0095] The hidden derivative is the rate of change of the hidden state with respect to time, which is calculated from the right-hand side of the model.

[0096] The heat generation term is a heat source driven by a heat source that is proportional to the square of the current sequence, and the square of the current is weighted by coefficients as the heat input.

[0097] The heat dissipation term is a heat exchange driver relative to the ambient proxy temperature, and the negative feedback term is formed by weighting the difference between the latent state and the ambient proxy temperature using the heat dissipation coefficient.

[0098] The relative temperature to the environment is the difference between the hidden state and the environmental temperature, used to characterize the temperature deviation relative to the environment;

[0099] The thermal memory kernel function is a family of kernel functions used to characterize thermal hysteresis and multi-timescale response, preferably composed of a weighted combination of several exponentially decaying kernels;

[0100] The convolution term is an additional driver obtained by convolving the hot memory kernel function with the historical input or hidden state difference over time, used to reflect the lag effect;

[0101] The exponential decay kernel is a kernel function of the form exp(−t / τ), where τ is a time constant. The thermal response at different time scales is characterized by a weighted combination of multiple τs.

[0102] The mapping relationship from the hidden state to the equivalent series resistance is obtained by adding the base resistance, the temperature coefficient term related to the hidden state of the conductor temperature, and the contact resistance term related to the hidden state of the junction temperature and the hidden state of the damage factor to obtain a parameterizable function for calculating the resistance value of the model.

[0103] The base resistance value is the equivalent series resistance reference term under zero temperature rise and no damage conditions;

[0104] The temperature coefficient term is a resistance increment that is implicitly associated with the temperature of the conductor and weighted by the temperature coefficient, and is used to describe the effect of temperature on the resistance of the conductor.

[0105] The contact resistance term is a resistance increment formed in association with the latent state of the contact temperature and the latent state of the damage factor, used to describe the additional impedance of the contact interface.

[0106] The resistance value sequence calculated by the model is the equivalent series resistance value sequence obtained by calculating point by point on the time index based on the above mapping relationship;

[0107] The physical consistency constraints are a set of three types of constraints imposed on the model, including monotonic constraints, energy conservation constraints, and hysteresis consistency constraints.

[0108] The monotonic constraint is a constraint that the calculated resistance of the model does not decrease when the hidden state of the contact temperature or the hidden state of the damage factor increases.

[0109] The energy conservation constraint is a constraint that the cumulative input of the square of the current sequence within any time window must maintain a consistent relationship with the increment of the hidden state of the conductor temperature and the hidden state of the junction temperature, and the error must not exceed a preset tolerance.

[0110] The hysteresis consistency constraint is a constraint that, after the load increases, the peak value of the hidden state of the resistance and contact temperature calculated by the model appears positive relative to the square peak value of the current and is within a preset range.

[0111] The reasonable lag is a time difference that is not less than zero and does not exceed a preset lag upper limit;

[0112] The assimilation is a process of online correction of the model based on the observed resistance sequence and its observation confidence. The hidden state and related parameters are updated by calculating the residual between the resistance and the observed resistance through weighted minimization of the model.

[0113] The weighted minimization is an estimation process that uses the observation confidence level as the sample weight to construct a loss function and then minimizes it.

[0114] The residual is the difference or error measure between the calculated resistance value and the observed resistance value.

[0115] The thermoelectric dynamics described here is a coupling mechanism that describes the interaction between current-induced heat generation, heat dissipation from the environment, and thermal hysteresis, which together affect the circuit resistance.

[0116] Optionally, step S5 specifically includes:

[0117] Based on the hidden state sequence of junction temperature and the time index sequence, the first difference is calculated according to the interval between adjacent time indices and divided by the corresponding time index interval to obtain the unsmoothed junction temperature slope sequence.

[0118] The unsmoothed junction temperature slope sequence is smoothed by applying a moving average of a sliding window to obtain the junction temperature slope sequence.

[0119] Based on the observed resistance drift sequence and aligned with the junction temperature slope sequence according to the time index sequence, interpolation is performed on time points where time indices are inconsistent to make the two sequences correspond at the same time index.

[0120] A fusion weight sequence is generated based on the load step intensity sequence, and the fusion weight sequence increases monotonically with the increase of load step intensity. The value of the fusion weight sequence is kept between zero and one by amplitude constraint.

[0121] At each time index, the sum of the fusion weight corresponding to the junction temperature slope and the fusion weight corresponding to the observed resistance drift is set to one, and the junction temperature slope sequence and the observed resistance drift sequence are weighted and summed according to the above fusion weight to form a joint trend evidence sequence.

[0122] Output the junction temperature slope sequence and the joint trend evidence sequence.

[0123] Terminology definition:

[0124] The adjacent time index interval is the time difference between two adjacent sampling points in the time index sequence, which is used for difference and normalization calculations.

[0125] The first-order difference is a calculation of the difference between the hidden values ​​of the junction temperature at adjacent time indices, and is normalized by the interval between adjacent time indices to approximate the time derivative.

[0126] The unsmoothed junction temperature slope sequence is a rate of change sequence obtained by dividing the junction temperature hidden state sequence by the first-order difference and the adjacent time index interval, and has not yet undergone smoothing and noise reduction processing.

[0127] The contact temperature slope sequence is a smoothed rate of change sequence obtained by applying a sliding window moving average to the unsmoothed contact temperature slope sequence, and is used to characterize the time variation trend of contact temperature.

[0128] The adaptive weighting is a weighting strategy that takes the load step intensity as input and determines the fusion weights through a preset monotonic function, so that the weights increase monotonically as the load step intensity increases.

[0129] The alignment and fusion process involves aligning the junction temperature slope sequence and the observed resistance drift sequence according to the time index, and then combining them based on the fusion weight to form a single trend evidence.

[0130] The fusion weights are non-negative weights assigned to the junction temperature slope or the observed resistance drift at a certain time index, used to control the contribution of the two types of evidence to the fusion result;

[0131] The fusion weight sequence is a sequence of fusion weight values ​​arranged according to time index, and its values ​​are constrained by amplitude and driven by the load step intensity sequence.

[0132] The amplitude constraint restricts the range of values ​​for the fusion weights to be within the closed interval [0,1].

[0133] The weight normalization constraint is a constraint that sets the sum of the fusion weight of the junction temperature slope and the fusion weight of the observed resistance drift to one at each time index.

[0134] The weighted summation is an operation in which the junction temperature slope and the observed resistance drift are multiplied by the corresponding fusion weights at each time index and then summed to obtain the fusion result.

[0135] The joint trend evidence sequence is a univariate sequence obtained by weighting and summing the docking point temperature slope sequence and the observed resistance drift sequence on a common time index according to the fusion weight, and is used to comprehensively characterize the abnormal temperature rise trend.

[0136] Optionally, step S6 specifically includes:

[0137] Based on the joint trend evidence sequence and calculated point by point according to the time index sequence, a sliding window containing the preceding and following segments is set at each time index. The mean difference, slope difference, and variance difference of the joint trend evidence between the preceding and following segments are compared, and the above differences are combined into a change point statistical score according to the preset weights.

[0138] Based on the monotonic mapping, the variable point statistical score is normalized into a variable point posterior sequence between 0 and 1, so that the variable point posterior sequence increases monotonically as the variable point statistical score increases.

[0139] Based on the load step intensity sequence and the damage factor latent state sequence, a hazard function is constructed, which is the product of the baseline hazard value, the load step intensity gain, and the damage factor gain. The gain is set to be a function that monotonically increases with the increase of the load step intensity and the damage factor latent state, so as to dynamically adjust the threshold sensitivity of the change point posterior sequence.

[0140] Set a confidence threshold and a minimum duration window. When the posterior sequence of the change point continuously exceeds the confidence threshold and meets the minimum duration window on multiple adjacent time indices, generate a graded early warning result sequence based on the magnitude and duration of the exceedance, and suppress false alarms caused by short-term fluctuations by maintaining the time parameter.

[0141] Output the posterior sequence of the change point and the hierarchical early warning result sequence.

[0142] Terminology definition:

[0143] The sliding window is a finite-length time window centered on the current time index, containing two sub-windows: one before the current time index and one after it, for local statistical comparison.

[0144] The front segment and the back segment are two sub-windows within the aforementioned sliding window, divided relative to the current time index. The front segment is located in the past interval of the time axis, and the back segment is located in the future interval of the time axis.

[0145] The mean difference is the difference obtained by subtracting the mean of the joint trend evidence samples from the mean of the later joint trend evidence samples;

[0146] The slope difference is the difference obtained by subtracting the linear fitting slope of the first segment from the linear fitting slope of the second segment sample.

[0147] The variance difference is the difference between the variance of the later sample and the variance of the earlier sample.

[0148] The preset weight is a non-negative weighting coefficient used to combine the mean difference, slope difference and variance difference into a single score, and can be set fixedly or adaptively.

[0149] The change point statistical score is a single value obtained by weighting the mean difference, slope difference and variance difference according to preset weights, and is used to measure the strength of evidence that there is a change point at the time index.

[0150] The monotonic mapping is a function that maps the variable point statistical score to the closed interval [0,1] and increases monotonically as the score increases;

[0151] The posterior sequence of the change point is a normalized sequence of values ​​obtained by monotonic mapping of the statistical scores of the change points at each time index, which is used to represent the posterior confidence of the occurrence of the change point.

[0152] The hazard function is a monotonic function used to adjust the threshold sensitivity, and its output is the product of the baseline hazard value, the load step intensity gain, and the damage factor gain.

[0153] The baseline hazard value is a constant or slowly varying baseline used for threshold adjustment without considering the effects of load and damage;

[0154] The load step strength gain is a gain function that monotonically increases with the increase of the load step strength, and is used to improve the output of the hazard function;

[0155] The damage factor gain is a gain function that monotonically increases with the increase of the hidden state of the damage factor, and is used to improve the output of the hazard function;

[0156] The threshold sensitivity dynamic adjustment is a process of changing the decision threshold or its equivalent sensitivity in real time based on the output of the hazard function, so that the decision is more sensitive when the risk increases and more conservative when the risk decreases.

[0157] The confidence threshold is a threshold parameter used to determine the posterior sequence of the change point, and its value is located in the closed interval [0,1].

[0158] The minimum duration window number is the minimum count required for the posterior of the change point to exceed the confidence threshold on the continuous-time index, used to suppress occasional fluctuations;

[0159] The excess amplitude is the difference between the posterior value of the change point and the confidence threshold, which is used to measure the excess intensity.

[0160] The duration is the length of continuous coverage that exceeds the confidence threshold in the posterior of the change point, which can be expressed as the number of sampling points or the length of time.

[0161] The graded early warning result sequence is a discrete graded alarm value sequence output by time index, and its grade is determined according to the combination rules of the exceedance amplitude and duration.

[0162] The hold time parameter is the minimum hold duration parameter used to maintain or delay alarm clearance. It is required to maintain the alarm continuously after the warning is reached to avoid false alarms or frequent jumps caused by short-term drop.

[0163] The beneficial effects of this invention are:

[0164] First, it has strong early identification capabilities: based on the observation of the equivalent series resistance of the current step segment, the low load baseline and the ambient proxy temperature, combined with the multi-hidden-state neural ordinary differential equation with thermal memory kernel and the change point detection of joint trend evidence, it can identify the slow heating trend of the conductor and the connection, and realize early warning before the over-temperature threshold strategy.

[0165] Second, it exhibits high robustness and low false alarm and false negative rates: Through observation confidence weighting, baseline drift suppression, monotonic consistency and energy conservation and hysteresis consistency triple physical constraints and observation assimilation, it improves stability under noise, harmonic, nonlinear load and grid fluctuation scenarios; combined with the dynamic threshold of load step strength and damage factor, it reduces false alarm and false negative rates.

[0166] Third, it requires no new hardware and is easy to implement: relying only on the existing voltage and current sampling data of the electricity meter, it runs online under a unified sampling clock and outputs hierarchical early warnings, making it suitable for large-scale deployment at the meter or edge devices, reducing transformation costs and maintenance complexity. Attached Figure Description

[0167] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0168] Figure 1 This is a flowchart of a method for identifying abnormal electricity usage in an electricity meter, as proposed in this invention. Detailed Implementation

[0169] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0170] refer to Figure 1 A method for identifying abnormal electricity usage in an electricity meter, comprising:

[0171] S1. Synchronously acquire and align voltage and current sequences under a unified sampling clock, and establish corresponding time indices;

[0172] S2. Identify current step segments based on the current sequence, calculate the local equivalent series resistance value in each segment according to the ratio of voltage change to current change, form an observation resistance value sequence in time order, and construct the observation confidence sequence and load step intensity sequence corresponding to the segment.

[0173] S3. Determine the low load period based on the current sequence, obtain the baseline resistance sequence by statistical analysis, subtract the baseline resistance from the observed resistance to obtain the observed resistance drift sequence, and smooth the baseline resistance with a low-pass filter to obtain the ambient temperature sequence.

[0174] S4. Drive the neural ordinary differential equation model containing the hidden states of conductor temperature, contact temperature and damage factor with the current sequence and the environmental proxy temperature sequence. Its hidden state derivative consists of the heat generation term of the square of the current, the heat dissipation term relative to the environmental proxy temperature and the thermal memory kernel convolution term. Map the hidden state to the equivalent series resistance value to obtain the model calculated resistance value sequence, and apply monotonic, energy conservation and hysteresis consistency constraints. Assimilate according to the observed resistance value sequence and its confidence level, and update the above hidden state and the model calculated resistance value sequence.

[0175] S5. The contact temperature slope sequence is obtained by differential and smoothing from the hidden state of the contact temperature. The contact temperature slope sequence is then adaptively weighted according to the load step intensity and aligned and fused with the observed resistance drift sequence to form a joint trend evidence sequence.

[0176] S6. Calculate the posterior sequence of the change point based on the joint trend evidence sequence, and dynamically adjust the threshold by combining the load step intensity and the latent state of the damage factor. When the posterior continuously exceeds the threshold in multiple adjacent sliding windows, output the graded early warning result sequence.

[0177] In this specific embodiment, S1 specifically refers to:

[0178] Driven by a unified sampling clock, the electricity meter synchronously samples voltage and current, and completes alignment and preprocessing. First, it generates a time index sequence according to a fixed sampling interval, which corresponds one-to-one with the voltage and current sequences. The time index satisfies the formula:

[0179] ;

[0180] in Indicates the first Time index of each sampling point, Indicates the sampling start time, Sample index is a non-negative integer. With a fixed sampling interval, voltage and current sequences were acquired on this time index sequence, denoted as follows: and Furthermore, alignment processing eliminates channel start-up offset and fixed delay, ensuring that the same Voltage and current samples corresponding to the same physical moment;

[0181] Then on and Implement data quality control and preprocessing, specifically including DC bias correction, which involves estimating the average bias of voltage and current within a sliding window and... and The process involves point-by-point subtraction to obtain the zero-mean AC component, range saturation checking (identifying samples approaching or reaching the upper or lower limits of the range and marking saturated segments for subsequent removal or replacement), noise reduction (applying linear phase low-pass or band-stop / notch filtering to both channels while maintaining synchronous phase relationship to suppress random noise and narrowband interference), and interpolation completion for single-point missing samples (when an isolated missing point exists at a certain time index, interpolation is performed based on its adjacent valid samples). and Perform linear or spline interpolation and retain the interpolation markers for subsequent weighting processing;

[0182] After completing the above steps, the preprocessed voltage and current sequences are output, denoted as follows: and and the corresponding one-to-one aligned time index sequence This provides synchronized, aligned, and quality-controlled data input for subsequent steps.

[0183] In this specific embodiment, S2 specifically refers to:

[0184] Based on time index sequence With the preprocessed current sequence Identifying current step segments involves determining the starting index of each current step segment based on whether the absolute value of the current difference between adjacent samples exceeds a preset step threshold and remains for at least a preset holding time within consecutive samples. End Index ,in Indicates the first Time index of each sampling point, For sample index, and They represent the first The start and end indices of each current step segment Indicates the index of the current step segment;

[0185] Short windows are set near the start index and the end index respectively for the preprocessed voltage sequence. With the preprocessed current sequence The starting voltage, starting current, ending voltage, and ending current are obtained by averaging, and a reference trajectory within the segment is constructed using linear interpolation for subsequent residual calculation.

[0186] In the For a given current step segment, the local equivalent series resistance is observed according to the formula:

[0187] ;

[0188] in Indicates the first Observed values ​​of the local equivalent series resistance of each current step segment Indicates the starting index Average voltage within the nearby short window Indicates the end of the index Average voltage within the nearby short window Indicates the starting index Average current within the nearby short window Indicates the end of the index The average current within a short window, with the underlined symbol indicating the window averaging operator. and These represent the start and end positions of the segment, respectively.

[0189] The current step segment is removed when the absolute value of the current change is lower than the minimum effective change threshold.

[0190] For the retained current step segment, the observation confidence is constructed based on the absolute value of the current change, the segment length, and the fitting residual of the linear trajectory of the voltage and current samples within the segment relative to the mean of the starting window and the mean of the ending window. The observation confidence increases monotonically with the increase of the absolute value of the current change and the segment length, and decreases monotonically with the increase of the fitting residual.

[0191] Simultaneously, the absolute value of the current change is taken as the load step intensity, and the arithmetic mean of the times corresponding to the start and end indices is taken as the segment center time. The observed resistance sequence is formed by mapping the central time to the time index sequence and arranging them in chronological order. At the same time, the corresponding observed confidence sequence and load step intensity sequence are also formed for use in subsequent steps.

[0192] In this specific embodiment, S3 specifically refers to:

[0193] Based on time index sequence With the preprocessed current sequence Identify low-load periods, among which Indicates the first Time index of each sampling point, For sample index, Indicates time index For current samples at a given location, time index segments with current amplitudes below a preset low-load threshold and a continuous duration of at least a preset minimum duration are identified during low-load periods. Time index segments adjacent to current step segments are excluded to avoid transient interference. The current step segment is indexed by its start and end. and Logo, For fragment indexing;

[0194] The local equivalent series resistance observations are mapped to time indices according to the segment center time to form an observation resistance sequence. ,in Indicates time index The observed equivalent series resistance at the location, Indicates the first Observed values ​​of the local equivalent series resistance of a current step segment;

[0195] During each low-load period, for Sliding statistics are performed, and the weighted median or weighted mean is calculated using the observation confidence level as the weight to estimate the baseline resistance for that period. The obtained baseline resistance values ​​are then concatenated into a baseline resistance value sequence according to the time index. ,in Indicates time index Baseline resistance at the location;

[0196] For time points not falling within low-load periods, the baseline resistance value of adjacent low-load periods is interpolated using a time index to make the result complete. Defined continuously over the entire time index;

[0197] Therefore, the observed resistance drift sequence is defined as the observed resistance minus the baseline resistance, and the calculation formula is as follows:

[0198] ;

[0199] in Indicates time index Observational resistance drift at the location Indicates time index The observed equivalent series resistance at the location, Indicates time index Baseline resistance at the location, For sample index;

[0200] Finally, for Low-pass smoothing using a sliding window moving average is applied to extract the slow-change component, which is then defined by time index as an environmental proxy temperature sequence. ,in Indicates time index The ambient temperature is used to generate a baseline resistance sequence. Observation of resistance drift sequence With environmental agent temperature sequence .

[0201] In this specific embodiment, S4 specifically refers to:

[0202] Based on time index sequence Preprocessed current sequence With environmental agent temperature sequence A physically constrained neural ordinary differential equation model is constructed to describe the continuous evolution of the hidden state over time, where Indicates the first Time index of each sampling point, For sample index, Indicates time index Current sample at the location, Indicates time index The ambient temperature at the location;

[0203] The model includes the hidden state sequences of conductor temperature, junction temperature, and damage factor, denoted as follows: and The driving term of the hidden derivative is composed of the heat generation term corresponding to the square of the current and the relative... The heat dissipation term is combined with the thermal memory convolution term composed of multiple time constant exponential kernels. The specific solution is achieved by an ordinary differential equation solver and trainable parameters, and the hidden trajectory is output continuously on the time axis.

[0204] To link the hidden states with the equivalent series resistance, an additive mapping is used to obtain the resistance sequence calculated by the model. The mapping relationship is as follows:

[0205] ;

[0206] in Indicates time index The model calculates the equivalent series resistance at the location. Indicates the basic resistance value under zero temperature rise and no damage conditions. The temperature coefficient representing the effect of conductor temperature on resistance. Indicates time index The conductor temperature is in a hidden state at the location. The coefficient representing the effect of contact temperature on the contact resistance term. Indicates time index The junction temperature is hidden. This represents the coefficient of the damage factor for the contact resistance term. Indicates time index The damage factor is in a latent state at the location;

[0207] To ensure interpretability and robustness, a triple physical consistency constraint is imposed on the model, namely, the monotonic constraint requires that when... or When increasing Without decreasing, the energy conservation constraint requires that within any finite time window, the energy must be reduced by... The heat corresponding to the square of the cumulative input and and The increase in load must remain consistent, and the hysteresis consistency constraint requires that after the load increases... and peak relative to The peak value has a non-negative time lag with an upper limit;

[0208] The assimilation stage forms and maps the observed resistance sequence to the time index in step S2. Its observed confidence sequence As input, perform analysis on the residuals at the time index where observations exist. Perform weighted minimization to update hidden states online. and and related mapping parameters ,in Indicates time index The observed equivalent series resistance at the location, Indicates time index The observation confidence weights at the location are determined by the free evolution of the neural ordinary differential equation and guided by constrained regularization terms for the time index of missing observations.

[0209] Output wire temperature hidden state sequence Hidden state sequence of junction temperature Damage factor hidden sequences With model calculation of resistance sequence .

[0210] In this specific embodiment, S5 specifically includes:

[0211] Based on time index sequence Hidden state sequence of junction temperature With the observed resistance drift sequence A junction temperature slope sequence is generated and aligned and fused to form joint trend evidence. First, under a fixed sampling interval, the unsmoothed junction temperature slope is calculated using the difference quotient between adjacent time indices, and then a sliding window moving average is applied to suppress high-frequency noise. The smoothed junction temperature slope sequence is denoted as... ,in Indicates the first Time index of each sampling point, For sample index, Indicates time index The junction temperature is hidden. Indicates time index The temperature slope at the junction;

[0212] Then and Alignment is performed according to time indices. Time points with inconsistent time indices are padded using time index interpolation to ensure correspondence on the same time index. Subsequently, a fused weight sequence is constructed based on the load step intensity formed and mapped to the time index in step S2, denoted as... and ,in Indicates time index Load step strength at the point Indicates time index The fusion weight assigned to the junction temperature slope is subject to amplitude constraints and lies within the range. and follow Increases monotonically, and at each time index, the fusion weight corresponding to the observed resistance drift is set to... To satisfy the weight normalization constraint;

[0213] Based on this, the joint trend evidence sequence is defined as a weighted sum:

[0214] ;

[0215] in Indicates time index Evidence of joint trends at the site Indicates time index Assigned to Fusion weights Indicates time index Temperature slope at the junction, Indicates time index The observed resistance drift at that location;

[0216] Output contact temperature slope sequence With joint trend evidence sequence .

[0217] In this specific embodiment, S6 specifically refers to:

[0218] Based on time index sequence Joint trend evidence sequence Step S2 maps the load step intensity sequence to the time index. and the hidden sequences of damage factors Conduct online change point identification and dynamic threshold-based early warning, indexing at each time point. The location is set up with a sliding window containing a front section and a back section, respectively for... Calculate the mean of the first segment and the mean of the second segment, the slope of the linear fit of the first segment and the slope of the linear fit of the second segment, and the variance of the first segment and the variance of the second segment. Subtract the first segment from the second segment to obtain the difference in mean, slope, and variance. Then, weight the above three types of differences according to the preset non-negative weights to obtain the change point statistical score. ,in Indicates the first Time index of each sampling point, For sample index, Indicates time index Evidence of joint trends at the site Indicates time index Load step strength at the point Indicates time index The latent state of damage factors at the location Indicates time index The change point statistical score is obtained by a weighted combination of the mean difference, slope difference, and variance difference.

[0219] The changepoint statistical score is normalized using a monotonic mapping to be between... The posterior of the changing point is as follows:

[0220] ;

[0221] in Indicates time index Posterior confidence level at the change point This indicates mapping real numbers to... And a normalized function that is monotonically non-decreasing with input, Indicates time index Statistical scoring of change points; further based on and Construct a hazard function and obtain a dynamic threshold sequence based on it. ,make Follow and Increase and monotonically decrease to improve the sensitivity of judgment under dangerous working conditions, and use the benchmark threshold. As a reference under no-load and no-damage conditions;

[0222] Set minimum number of persistent windows With holding time parameter ,when Continuously exceeding multiple adjacent time indices And the continuous length is not less than At that time, a tiered early warning result sequence is generated based on the combination rules of the magnitude and duration. ,in Indicates time index The dynamic judgment threshold used at the location Represents the reference threshold constant, Indicates the minimum number of sustained windows. This indicates the hold-up time parameter used to maintain alarm output and suppress short-term drop-off. Indicates time index The output is a discrete hierarchical early warning result;

[0223] Output the posterior sequence of the variable point With the sequence of graded early warning results .

[0224] 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.

[0225] This invention constructs a causal chain for indirect temperature rise identification under sensorless conditions: First, voltage and current are synchronously acquired under a unified clock. Based on the current step, the local equivalent series resistance is extracted and the confidence level is estimated. A baseline resistance and ambient proxy temperature are established during low-load periods. Then, thermoelectric coupling is described using a neural network constant differential equation including wire temperature, contact temperature, and damage factor. The derivative consists of the heat generated by the square of the current, the heat dissipation relative to the ambient proxy temperature, and a thermal memory kernel term. Weighted assimilation ensures that the model's calculated resistance matches the observations. Finally, the alignment and fusion of the contact temperature slope and resistance drift form joint trend evidence. Combined with online change-point posterior and a dynamic threshold that adapts to the load step intensity and damage factor, a graded early warning for continuous over-limit conditions is output. This chain maps measurable electrical parameters to temperature rise trends, effectively achieving early identification and low false alarms.

[0226] In terms of algorithm structure, this invention proposes targeted improvements to address the technical problems: First, it employs multi-hidden-state neural ordinary differential equations and introduces a thermal memory kernel to explicitly characterize thermal hysteresis and material degradation, making slow heating and progressive damage distinguishable and traceable. Second, it applies triple physical constraints of monotonicity, energy conservation, and hysteresis consistency to enhance interpretability and parameter identifiability, avoiding unreasonable estimates under noise and harmonic conditions. Third, it utilizes low-load baselines and environmental proxy temperatures to compensate for environmental and baseline drift, and improves the effectiveness of resistance observation by weighting observation confidence and step intensity. Fourth, it makes detection sensitive to real risks but insensitive to short-term disturbances by combining trend evidence and dynamic thresholding of the hazard function. These improvements synergistically enhance lead time, stability, and generalization ability, better achieving the effect of early warning technology.

Claims

1. A method for identifying abnormal electricity usage in an electricity meter, characterized in that, Include: S1. Synchronously acquire and align voltage and current sequences under a unified sampling clock, and establish corresponding time indices; S2. Identify current step segments based on the current sequence, calculate the local equivalent series resistance value in each segment according to the ratio of voltage change to current change, form an observation resistance value sequence in time order, and construct the observation confidence sequence and load step intensity sequence corresponding to the segment. S3. Determine the low load period based on the current sequence, obtain the baseline resistance sequence by statistical analysis, subtract the baseline resistance from the observed resistance to obtain the observed resistance drift sequence, and smooth the baseline resistance with a low-pass filter to obtain the ambient temperature sequence. S4. Drive the neural ordinary differential equation model containing the hidden states of conductor temperature, contact temperature and damage factor with the current sequence and the environmental proxy temperature sequence. Its hidden state derivative consists of the heat generation term of the square of the current, the heat dissipation term relative to the environmental proxy temperature and the thermal memory kernel convolution term. Map the hidden state to the equivalent series resistance value to obtain the model calculated resistance value sequence, and apply monotonic, energy conservation and hysteresis consistency constraints. Assimilate according to the observed resistance value sequence and its confidence level, and update the above hidden state and the model calculated resistance value sequence. S5. The contact temperature slope sequence is obtained by differential and smoothing from the hidden state of the contact temperature. The contact temperature slope sequence is then adaptively weighted according to the load step intensity and aligned and fused with the observed resistance drift sequence to form a joint trend evidence sequence. S6. Calculate the posterior sequence of the change point based on the joint trend evidence sequence, and dynamically adjust the threshold by combining the load step intensity and the latent state of the damage factor. When the posterior continuously exceeds the threshold in multiple adjacent sliding windows, output the graded early warning result sequence.

2. The method for identifying abnormal electricity use in an electricity meter according to claim 1, characterized in that, S1 specifically refers to: Synchronous sampling is performed based on the unified sampling clock of the electricity meter. Voltage timing is collected at fixed sampling intervals to form a voltage sequence, and current timing is collected to form a current sequence. A time index sequence is generated based on the fixed sampling interval, so that the time index sequence corresponds one-to-one with the voltage sequence and the current sequence; Data quality control and preprocessing are performed on voltage and current sequences, including DC bias correction, range saturation check, noise reduction, and interpolation completion for single-point missing samples caused by loss. Output preprocessed voltage, current and time index sequences.

3. The method for identifying abnormal electricity use in an electricity meter according to claim 1, characterized in that, S2 specifically refers to: Based on the current sequence and time index sequence, current step segments are identified according to the absolute value of the current difference between adjacent samples exceeding a preset step threshold and lasting for no less than a preset holding time in consecutive samples. A start index and an end index are determined for each current step segment. Within each current step segment, the voltage and current sequences are averaged in short windows near the start and end indices to obtain the start voltage, start current, end voltage, and end current. The difference between the ending voltage and the starting voltage is used as the voltage change, and the difference between the ending current and the starting current is used as the current change. When the absolute value of the current change is lower than the minimum effective change threshold, the current step segment is removed. For the retained current step segment, the local equivalent series resistance value of the segment is obtained by dividing the voltage change by the current change value, and the center time of the segment is mapped to the time index sequence and arranged in time order to form the observed resistance value sequence. For each observed resistance value, the observation confidence level is constructed based on the absolute value of the current change in that segment, the segment length, and the fitting residuals of the voltage and current samples within the segment relative to the start and end points of the linear connection, thus forming an observation confidence level sequence. The absolute value of the current change is used as the load step intensity to form a load step intensity sequence; Output the observed resistance sequence, the observed confidence sequence, and the load step intensity sequence.

4. The method for identifying abnormal electricity use in an electricity meter according to claim 1, characterized in that, S3 specifically refers to: Based on the observed resistance sequence and the observed confidence sequence, as well as the current sequence and the time index sequence, the low load period is determined. The low load period is a time index segment in which the current sequence is lower than a preset low load threshold and the continuous time is not less than a preset minimum duration. Time index segments adjacent to current step segments are excluded to avoid transient interference. During each low-load period, sliding statistics are performed on the observed resistance value sequence, and the weighted median or weighted mean is calculated using the observed confidence sequence as the weight to obtain the baseline resistance value for that period. The baseline resistance values ​​of each time period are concatenated into a baseline resistance value sequence according to the time index sequence. For time points that are not in low load periods, the corresponding baseline resistance values ​​are obtained by interpolation through the time index based on the baseline resistance values ​​of the adjacent low load periods, so that the baseline resistance value sequence is continuous in the entire time index sequence. The observed resistance drift sequence is formed by subtracting the baseline resistance sequence for the same time period from the observed resistance sequence. Based on low-pass smoothing of the baseline resistance sequence by applying a sliding window moving average, the slow-change component is extracted and defined as an environmental proxy temperature sequence by time index sequence; Output baseline resistance sequence, observed resistance drift sequence, and ambient proxy temperature sequence.

5. The method for identifying abnormal electricity use in an electricity meter according to claim 1, characterized in that, S4 specifically refers to: Based on the thermoelectric dynamics driven by the current sequence, time index sequence, and environmental proxy temperature sequence, a neural ordinary differential equation model is established, which includes the hidden state sequence of conductor temperature, the hidden state sequence of junction temperature, and the hidden state sequence of damage factor. The time derivative of the hidden state is represented as a combination of a heat-generating term corresponding to the square of the current sequence, a heat-dissipating term relative to the ambient proxy temperature sequence, and a convolution term of the thermal memory kernel function, wherein the thermal memory kernel function is composed of a set of exponentially decaying kernels to characterize thermal hysteresis. The resistance sequence is calculated based on the mapping relationship from the hidden state to the equivalent series resistance. The mapping relationship includes the base resistance, the temperature coefficient term related to the hidden state sequence of the conductor temperature, and the contact resistance term related to the hidden state sequence of the contact temperature and the hidden state sequence of the damage factor. The model is subject to three physical consistency constraints, including a monotonic constraint that the calculated resistance sequence does not decrease when the hidden sequence of contact temperature or damage factor increases; an energy conservation constraint that the cumulative input of the square of the current sequence remains consistent with the increase of the hidden sequence of conductor temperature and contact temperature within any time window; and a hysteresis consistency constraint that the peak values ​​of the calculated resistance sequence and the hidden sequence of contact temperature lag reasonably with respect to the peak value of the square of the current sequence after the load increases. Based on the assimilation of the observed resistance sequence and the observed confidence sequence on the time index sequence, the residuals of the resistance sequence and the observed resistance sequence are calculated by the weighted minimization model to update the hidden sequence of conductor temperature, hidden sequence of contact temperature and hidden sequence of damage factor, and the model is updated to calculate the resistance sequence simultaneously. The output conductor temperature hidden sequence, the contact temperature hidden sequence, the damage factor hidden sequence, and the model calculated resistance sequence.

6. The method for identifying abnormal electricity use in an electricity meter according to claim 1, characterized in that, S5 specifically refers to: Based on the hidden state sequence of junction temperature and the time index sequence, the first difference is calculated according to the interval between adjacent time indices and divided by the corresponding time index interval to obtain the unsmoothed junction temperature slope sequence. The unsmoothed junction temperature slope sequence is smoothed by applying a moving average of a sliding window to obtain the junction temperature slope sequence. Based on the observed resistance drift sequence and aligned with the junction temperature slope sequence according to the time index sequence, interpolation is performed on time points where time indices are inconsistent to make the two sequences correspond at the same time index. A fusion weight sequence is generated based on the load step intensity sequence, and the fusion weight sequence increases monotonically with the increase of load step intensity. The value of the fusion weight sequence is kept between zero and one by amplitude constraint. At each time index, the sum of the fusion weight corresponding to the junction temperature slope and the fusion weight corresponding to the observed resistance drift is set to one, and the junction temperature slope sequence and the observed resistance drift sequence are weighted and summed according to the above fusion weight to form a joint trend evidence sequence. Output the junction temperature slope sequence and the joint trend evidence sequence.

7. The method for identifying abnormal electricity use in an electricity meter according to claim 1, characterized in that, S6 specifically refers to: Based on the joint trend evidence sequence and calculated point by point according to the time index sequence, a sliding window containing the preceding and following segments is set at each time index. The mean difference, slope difference, and variance difference of the joint trend evidence between the preceding and following segments are compared, and the above differences are combined into a change point statistical score according to the preset weights. Based on the monotonic mapping, the variable point statistical score is normalized into a variable point posterior sequence between 0 and 1, so that the variable point posterior sequence increases monotonically as the variable point statistical score increases. Based on the load step intensity sequence and the damage factor latent state sequence, a hazard function is constructed, which is the product of the baseline hazard value, the load step intensity gain, and the damage factor gain. The gain is set to be a function that monotonically increases with the increase of the load step intensity and the damage factor latent state, so as to dynamically adjust the threshold sensitivity of the change point posterior sequence. Set a confidence threshold and a minimum duration window. When the posterior sequence of the change point continuously exceeds the confidence threshold and meets the minimum duration window on multiple adjacent time indices, generate a graded early warning result sequence based on the magnitude and duration of the exceedance, and suppress false alarms caused by short-term fluctuations by maintaining the time parameter. Output the posterior sequence of the change point and the hierarchical early warning result sequence.

Images