Metering box early warning method and system based on federated learning and multi-modal data fusion

Abstract

This invention belongs to the field of machine learning technology, specifically relating to a metering box early warning method and system based on federated learning and multimodal data fusion. The method includes: acquiring the original sampling sequence and splitting it into a transient electromechanical parameter sequence and a slowly varying environmental parameter sequence according to the medium's conduction velocity; extracting transient historical change features and constructing a cumulative fatigue damage degree based on the fatigue background sequence; calculating the physical compensation margin of the slowly varying parameters and superimposing the real sampled values ​​to obtain environmental feature values; concatenating the dimensionally aligned environmental feature values ​​with the transient parameters to obtain a spatiotemporally aligned feature vector to train a local defect identification model; extracting the discrete standard deviation of the physical compensation margin under steady-state conditions to calculate the aggregation weight coefficient, and uploading it to a central coordination system for weighted aggregation. This invention eliminates spatiotemporal misalignment, suppresses interference from aging node contamination, and improves early warning accuracy.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology. More specifically, this invention relates to a metering bin early warning method and system based on federated learning and multimodal data fusion. Background Technology

[0002] As a key terminal device in the power distribution network, the operational stability of smart metering boxes directly affects electricity safety and the accuracy of electricity metering. In complex outdoor operating environments, metering boxes not only face the effects of fluctuating temperature and humidity but also the thermal effects of drastic changes in line current and the mechanical stress waves caused by sudden external impacts. Existing early warning systems often utilize various sensors to acquire multimodal state data, determine equipment anomalies through fusion technology, and combine this with a federated learning framework. This allows each computing node to train a model locally using multimodal data, and then uploads the parameters to a central server for aggregation and updates, thereby constructing a distributed early warning network while protecting privacy.

[0003] However, the aforementioned existing technologies have shortcomings in practical applications: current fusion methods often forcibly align multimodal data based on the same system clock, ignoring the differences in propagation rates of different physical parameters and sensor response hysteresis when anomalies occur. For example, mechanical stress waves propagate extremely fast, while heat conduction or air convection processes are extremely slow. Forced alignment can cause severe spatiotemporal tearing of feature dimensions, leading to serious misjudgments of initial faults. In addition, since existing early warning models mostly use fixed prior parameters for calibration, the system cannot detect the microscopic fatigue damage to the metering box caused by environmental stress. This easily leads to significant long-term drift errors in the later stages of service, resulting in distorted output metering box early warning results and failing to guarantee high accuracy and reliability of early warning throughout the equipment's entire life cycle. Summary of the Invention

[0004] To address the technical problem that the data alignment logic and model calibration parameters of existing early warning systems are deviated from the actual physical evolution of the equipment, resulting in severe distortion of the full life cycle state assessment, this invention provides solutions in the following aspects.

[0005] In a first aspect, the present invention provides a metering bin early warning method based on federated learning and multimodal data fusion, comprising:

[0006] The original sampling sequence of the intelligent metering box sensor is obtained, and the original sampling sequence is split into transient electromechanical parameter sequence and slowly changing environmental parameter sequence according to the medium conduction velocity, and a unified timestamp is assigned.

[0007] Extract the historical variation characteristics of transient electromechanical parameter sequences, and construct the cumulative fatigue damage degree based on the fatigue background sequence when the enclosure is determined to be in its effective service life.

[0008] Based on the historical variation characteristics of transient electromechanical parameter sequences and cumulative fatigue damage, the physical compensation margin of the slowly changing environmental parameter sequences at the current sampling time is calculated.

[0009] The physical compensation margin is superimposed on the real sampled values ​​of the slowly varying environmental parameter sequence to obtain environmental feature values. The expanded and aligned environmental feature values ​​are then concatenated with the transient electromechanical parameter sequence to obtain a spatiotemporally aligned feature vector. This vector is then combined with the defect classification label to generate a local fusion training set.

[0010] The isomorphic defect identification model is trained and deployed using a local fusion training set. The discrete standard deviation of the physical compensation margin under the steady state of the transient electromechanical parameter sequence is extracted to calculate the aggregate weight coefficient. The local model parameters and aggregate weight coefficients are then uploaded to the central coordination system for weighted aggregation and update.

[0011] Preferably, the step of extracting historical variation characteristics of transient electromechanical parameter sequences and constructing cumulative fatigue damage based on fatigue background sequences when determining that the housing is within its effective service life includes:

[0012] Extract the historical data window, extract the absolute value of the difference between all sampled values ​​of the transient electromechanical parameter sequence within the historical data window and the actual sampled value at the start time, and divide the largest absolute value among the absolute values ​​by the time span length to obtain the first-order rate of change of the transient electromechanical parameter sequence.

[0013] The sampling amplitude of transient electromechanical parameter sequences, sensor communication status, and first-order rate of change of slowly changing environmental parameter sequences are monitored in real time. When the preset over-limit conditions are met, it is determined that the enclosure has macroscopically broken and an emergency physical cut-off and alarm mechanism is triggered. When the preset over-limit conditions are not met, it is determined that the enclosure is in its effective service life. Transient electromechanical parameter sequences with peak values ​​lower than the normal business operation threshold are extracted as fatigue background sequences to construct the cumulative fatigue damage degree.

[0014] Preferably, the step of constructing the cumulative fatigue damage degree based on the fatigue background sequence includes:

[0015] The discrete statistical variance of the fatigue background sequence within the sliding time window is calculated as the instantaneous disturbance power. The instantaneous disturbance power is accumulated and integrated over the equipment service time axis to obtain the cumulative disturbance energy value.

[0016] Substituting the accumulated disturbance energy value into the normalized fatigue mapping function, the accumulated fatigue damage degree is obtained: In the formula, Indicates the cumulative fatigue damage degree. This represents the cumulative disturbance energy value. For reference, the cumulative disturbance energy value, This is a preset damage upper limit constant.

[0017] Preferably, the physical compensation margin satisfies the expression:

[0018] ;

[0019] In the formula, This represents the physical compensation margin corresponding to the current sampling time; The dimensionless preset spatiotemporal coupling sensitivity coefficient; This indicates the maximum full-scale span value of the slowly changing environmental parameters inside the intelligent metering box of this model; This represents the first-order rate of change of a transient electromechanical parameter sequence; This represents the initial reference tolerance rate constant specified by the manufacturer. Indicates the cumulative degree of fatigue damage; This represents the hyperbolic tangent function.

[0020] Preferably, before calculating the physical compensation margin of the slowly varying environmental parameter sequence at the current sampling time, the method further includes:

[0021] A first-in-first-out circular queue is established as a time alignment buffer for the parameter sequence of a slowly changing environment. The queue depth of the time alignment buffer is configured as follows: ,in The queue depth for the time alignment buffer. This is the single thermal equilibrium recovery time constant of the intelligent metering box. The sampling interval for the slowly changing environmental parameter sequence, The rounding up symbol;

[0022] When the time alignment buffer is full, discard the oldest data at the head of the queue, and extract the actual sampled value of the slowly varying environmental parameter sequence residing in the time alignment buffer at the current sampling time to calculate the physical compensation margin.

[0023] Preferably, the step of concatenating the dimensionally aligned environmental feature values ​​with the transient electromechanical parameter sequence to obtain the spatiotemporally aligned feature vector includes:

[0024] The zero-order hold interpolation method is used to expand the array length of environmental feature values ​​upward to match the array length of the real-time sampled values ​​of the transient electromechanical parameter sequence. The expanded environmental feature values ​​and the real-time sampled values ​​of the transient electromechanical parameter sequence are then concatenated to obtain the spatiotemporally aligned feature vector.

[0025] Preferably, the step of calculating the aggregated weighting coefficient based on the discrete standard deviation of the physical compensation margin in the stationary state of the extracted transient electromechanical parameter sequence includes:

[0026] When the sampling amplitude of the transient electromechanical parameter sequence continuously lies between the upper and lower boundary lines symmetrically constructed based on the positive and negative values ​​of the normal business operation threshold during the model evaluation period, the statistical standard deviation of multiple consecutive physical compensation margins during the model evaluation period is calculated; the factory-calibrated background hardware noise threshold is obtained; and the reciprocal of the sum of the statistical standard deviation of the physical compensation margin and the background hardware noise threshold is used as the aggregation weight coefficient.

[0027] Preferably, the normal business operation threshold is dynamically updated using an exponential moving average.

[0028] Preferably, uploading the local model parameters and aggregated weight coefficients to the central coordination system for weighted aggregation and updating includes:

[0029] Upload the local model parameters and aggregate weight coefficients to the central coordination system;

[0030] In response to the detection that the statistical standard deviation of the physical compensation margin calculated by all nodes participating in this round of aggregation exceeds the preset failure baseline. If a node is in a cold start period, a backup mechanism to prevent paralysis is triggered: the central coordination system forcibly suspends the aggregation weight coefficient calculated based on the statistical standard deviation, degenerates the aggregation strategy of this round into an arithmetic mean weight strategy with uniform distribution across computing nodes, and uses the arithmetic mean of the local model parameters uploaded by all local computing nodes participating in this round of aggregation as the global model parameters; wherein, the failure baseline Satisfying the formula: In the formula, The preset chaos tolerance coefficient is used to characterize the standard deviation shrinkage ratio under extreme and harsh operating conditions; The dimensionless preset spatiotemporal coupling sensitivity coefficient; This indicates the maximum full-scale span value of the slowly changing environmental parameters inside the intelligent metering box of this model;

[0031] In response to the failure to trigger the anti-paralysis backup mechanism, the central coordination system uses the aggregate weight coefficient of each node as a multiplier to multiply the corresponding local model parameters, sums all the product results and divides them by the sum of the aggregate weight coefficients of all nodes to obtain the global model parameters.

[0032] The acquired global model parameters will be distributed and updated.

[0033] Secondly, the present invention provides a meter box early warning system based on federated learning and multimodal data fusion, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned meter box early warning method based on federated learning and multimodal data fusion is implemented.

[0034] By adopting the above technical solution, the meter box early warning method based on federated learning and multimodal data fusion is generated into a computer program and stored in a memory for loading and execution by a processor. This allows for the creation of a terminal device based on the memory and processor, making it convenient to use.

[0035] The beneficial effects of this invention are as follows: This invention decomposes the original sensing data into transient electromechanical parameters and slowly varying environmental parameters based on the medium conduction velocity. It then combines the historical variation characteristics of the transient electromechanical parameters with the cumulative fatigue damage degree to perform physical compensation on the slowly varying environmental parameters, eliminating the spatiotemporal misalignment caused by the difference in propagation delay of heterogeneous physical quantities. This ensures that the feature alignment logic aligns with the actual physical evolution law of the underlying equipment. Simultaneously, it trains a defect identification model based on spatiotemporal aligned feature vectors and extracts the discrete standard deviation of the physical compensation margin in a stable state of the transient electromechanical parameter sequence to calculate the aggregation weight coefficient. This objectively quantifies the stability of the local sensing hardware, thereby adaptively weakening the weight of aging or damaged nodes during federated weighted aggregation. This effectively suppresses the interference of model parameters uploaded by unstable nodes on the global network, improving the accuracy and robustness of the equipment's full lifecycle early warning assessment. Attached Figure Description

[0036] Figure 1 This is a flowchart of the meter box early warning method based on federated learning and multimodal data fusion in this invention;

[0037] Figure 2 This is a schematic diagram illustrating the changes in the transient electromechanical parameter sequence and the normal service operation threshold in this invention;

[0038] Figure 3 This is a schematic diagram comparing the actual sampled values ​​of the slowly changing environmental parameters with the compensated environmental characteristic values ​​in this invention;

[0039] Figure 4 This is a schematic diagram illustrating the change of the aggregation weight coefficient over time in this invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0042] This invention discloses a metering bin early warning method based on federated learning and multimodal data fusion, referring to... Figure 1 This includes steps S1-S4:

[0043] S1: Obtain the original sampling sequence of the smart metering box sensor, split the original sampling sequence into transient electromechanical parameter sequence and slowly changing environmental parameter sequence according to the medium conduction speed, and assign a unified timestamp.

[0044] It should be noted that in smart metering boxes, changes in line current or mechanical vibrations caused by impacts to the box propagate as electromagnetic waves or mechanical stress waves. Sensors respond extremely quickly and require high-frequency sampling to capture transient spikes. However, in a sealed operating state without macroscopic cracking, the internal temperature and humidity changes of a smart metering box primarily depend on natural weather changes in the external environment and the slow-state heat conduction of the box's outer shell. Its thermodynamic evolution is extremely slow, making low-frequency sampling suitable. Therefore, this invention separates these two types of data with different response speeds at the hardware acquisition source and assigns timestamps using a unified hardware clock, providing underlying data support for subsequent alignment across physical time differences.

[0045] Specifically, the system acquires the raw digital signals output by various physical sensors installed inside the smart metering box via analog-to-digital converters, and reads the hardware clock signal of the device's main control chip. A unique timestamp is added to each raw digital signal to generate a raw sampling sequence carrying a time tag. Based on the medium conduction velocity of the physical quantities monitored by the sensors, the raw sampling sequence carrying the time tag is divided into a transient electromechanical parameter sequence and a slowly varying environmental parameter sequence. The transient electromechanical parameter sequence includes line current data collected by a current sensor and mechanical vibration data collected by a vibration sensor, with a sampling frequency of not less than 1000 Hz. The slowly varying environmental parameter sequence includes temperature and humidity data collected by a temperature and humidity sensor, with a sampling frequency of not more than 1 Hz.

[0046] S2: Extract the historical change characteristics of the transient electromechanical parameter sequence, and when the enclosure is determined to be in its effective service life, construct the cumulative fatigue damage degree based on the fatigue background sequence. Based on the historical change characteristics of the transient electromechanical parameter sequence and the cumulative fatigue damage degree, calculate the physical compensation margin of the slowly changing environmental parameter sequence at the current sampling time.

[0047] It should be noted that conventional signal alignment methods involve capturing data from different sensors at the same system timestamp. However, this ignores the delay effect of physical anomalies propagating in space. To address this spatiotemporal misalignment, this invention utilizes the early burst energy contained in transient electromechanical parameters to deduce the actual change in slowly varying environmental parameters after overcoming physical diffusion delays. Furthermore, in long-term outdoor service scenarios, intelligent metering boxes continuously endure wind-driven micro-vibrations. This long-term environmental parasitic stress can induce irreversible material fatigue and micro-yielding, causing a nonlinear decay in the box's true physical limit against external damage. Adhering rigidly to the factory-calibrated static reference constant will inevitably result in a severe undercompensation for aging equipment. To this end, this invention is based on the classic principle of fatigue cumulative damage. Without adding additional structural health monitoring hardware, it extracts transient electromechanical parameter sequences with peak values ​​below the normal business operation threshold during historical normal operation to characterize the small electromechanical disturbance energy regarded as obsolete noise, constructs parasitic fatigue attenuation constraints, and implements dynamic real-time contraction of the benchmark tolerance rate of change constant throughout the entire life cycle of the equipment. This ensures that the pre-compensation logic for timing misalignment still maintains high physical precision and high reliability under extreme hardware aging conditions.

[0048] Specifically, based on the current sampling time, a historical data window is extracted by tracing back in the historical direction. The time span of the historical data window is determined by the implementer through an acoustic testing instrument to pre-determine the stress wave characteristic decay time constant of the metering box of this model. The specific method for obtaining the stress wave characteristic decay time constant is as follows: a standard pulse excitation is applied to the metering box. Since the energy dissipation of the stress wave in the solid medium follows the objective physical law of natural exponential decay, the time required for the amplitude of the electromechanical response signal to decay to the inverse of the natural constant, i.e., 37%, of the initial peak value is recorded using an acoustic testing instrument as the stress wave characteristic decay time constant. All real-time sampled values ​​of the transient electromechanical parameter sequence within the historical data window ending at the current sampling time are extracted, and the absolute values ​​of the transient electromechanical parameter sequence corresponding to the start time of the historical data window are subtracted from each of them to obtain the transient historical background value. The maximum absolute value of the transient historical background value is divided by the time span of the historical data window to obtain the first-order rate of change of the transient electromechanical parameter sequence.

[0049] Furthermore, before performing material fatigue accumulation tracking and subsequent physical compensation, the underlying embedded system responsible for the basic control of the equipment needs to prioritize real-time judgment of whether the enclosure has experienced macroscopic fracture. Specifically, the triggering mechanism for judging macroscopic fracture of the enclosure includes at least one of the following: First, an absolute value hard limit judgment mechanism, that is, real-time comparison of the absolute value of the sampling amplitude of the transient electromechanical parameter sequence. When the absolute value of the sampling amplitude is greater than the preset physical fracture limit absolute threshold, the mechanism is triggered to judge fracture. The physical fracture limit absolute threshold is calibrated by the implementer conducting a destructive impact test on the same model enclosure and extracting the maximum absolute value of the sensor response amplitude at the moment when the enclosure shell undergoes substantial fracture. Second, a hardware physical disconnection judgment mechanism, that is, real-time monitoring of the communication status of each sensor through the hardware watchdog of the main control chip. When the internal sensor core is detected, the mechanism will determine whether the fracture has occurred. The mechanism is triggered when a data packet is lost or the data bus level is abnormally floating, thus indicating that the internal physical wiring is broken due to severe tearing of the enclosure structure. The third is the thermodynamic sealing model collapse judgment mechanism, which calculates the first-order rate of change of the slowly changing environmental parameter sequence in real time. When the first-order rate of change exceeds the preset physical limit of conventional heat conduction in the sealed space, the mechanism is triggered to determine the collapse, thus indicating that the temperature and humidity change is not natural and instantaneous due to the direct connection and convection of the internal and external air after the enclosure is damaged. The physical limit of conventional heat conduction in the sealed space is constrained by the implementer testing the maximum natural heating or cooling rate that the metering box can achieve in a perfectly sealed state in a climate alternation test chamber.

[0050] When any of the above triggering mechanisms are met, it is determined that the enclosure has actually suffered a macroscopic rupture due to extreme damage, and its internal sealed physical heat conduction medium has been destroyed. At this time, the underlying embedded system directly takes over the global hardware control of the device, and outputs a fatal alarm by triggering the highest priority hardware interrupt program. Specifically, the fatal alarm drives the local relay to perform an emergency physical cutoff of the main circuit, and simultaneously controls the communication module to send an urgent short message containing the confirmed status of physical damage to the enclosure to the power dispatch center.

[0051] If none of the above triggering mechanisms are met, it is determined that the enclosure has not experienced macroscopic cracking and is still within its effective service life. At this point, the system enters the microscopic fatigue tracking and proactive compensation phase. Specifically, this involves: constructing upper and lower boundary lines symmetrically based on the positive and negative values ​​of the normal operating threshold; and extracting transient electromechanical parameter sequences whose sampling amplitudes have consistently remained between the upper and lower boundary lines since the equipment went online as the fatigue background sequence. The normal operating threshold is dynamically updated using an exponential moving average. ,in This represents the normal business operation threshold for the current period. Based on the normal business operation threshold of the previous cycle, the system sets the normal business operation threshold for the initial cycle for the cold start phase of the device's first online deployment. The standard operating load peak constant as specified by the manufacturer. To extract the 95th percentile values ​​of the transient electromechanical parameter sequence peaks during the historical fault-free operation cycle of this node, The forgetting factor is obtained by the implementers based on the historical long-term load volatility measurement of the service environment in which the smart metering box is located. The specific measurement method is as follows: calculate the ratio of the standard deviation to the mean of the transient electromechanical parameter sequence of the node during the historical fault-free operation period as the historical long-term load volatility, and then linearly normalize the reciprocal of the load volatility value, truncating it to an empirical value range of 0.01 to 0.05, which is used as the forgetting factor. The specific value of the threshold is determined by introducing an exponential moving average mechanism based on a forgetting factor. This invention ensures that the dynamic normal business operation threshold drifts slowly with environmental load characteristics, avoiding misjudgments due to sudden changes.

[0052] For example, Figure 2 This diagram illustrates the changes in transient electromechanical parameter sequences and normal operating thresholds. It includes the fluctuation curves of the transient electromechanical parameter sequences, as well as upper and lower boundary lines constructed symmetrically from the normal operating threshold values. During normal equipment operation, the fluctuation amplitude of the transient electromechanical parameter sequences is consistently limited to the range defined by these two boundary lines. This data is extracted as a fatigue background sequence for cumulative damage calculation. When the equipment experiences an external impact, the transient electromechanical parameter sequences exhibit significant spikes, with their instantaneous amplitude exceeding either the upper or lower boundary line. The upper and lower boundary lines show a synchronous and slow adjustment trend over time, demonstrating the effectiveness of using an exponential moving average mechanism to ensure that the threshold boundaries evolve smoothly with environmental load characteristics, effectively avoiding misjudgments caused by instantaneous signal mutations.

[0053] Based on the principle of energy accumulation, the degree of microscopic physical damage to a solid material under alternating loads is proportional to the accumulated external energy it absorbs and dissipates. That is, the degree of damage is determined by the proportion of currently dissipated energy to the yield critical energy. Therefore, this invention constructs a sliding time window using the characteristic decay time constant of the stress wave as the window length, calculates the discrete statistical variance of the fatigue background sequence within each sliding time window, and uses this discrete statistical variance as the instantaneous disturbance power for the corresponding sliding time window. The instantaneous disturbance power is then accumulated and integrated over the equipment's service time axis to obtain the accumulated disturbance energy value. This characterizes the accumulation process of fatigue work within the structure.

[0054] Accumulated disturbance energy value Substituting into the normalized fatigue mapping function, we obtain the cumulative fatigue damage:

[0055]

[0056] In the formula, It represents the cumulative fatigue damage degree that characterizes the physical fatigue degradation state of a material; The reference cumulative disturbance energy value for the same type of metering box when it reaches the structural yield critical point in accelerated fatigue test is obtained as follows: take at least three sets of the same type of metering box prototypes, apply a step alternating load higher than normal working conditions using a high-frequency fatigue testing machine to simulate the accelerated aging process, and simultaneously monitor the change of the natural frequency of the prototype structure. The moment when the natural frequency first shows an irreversible shift of more than 5% is determined as the structural yield critical point. Subsequently, use vibration sensors to collect the transient electromechanical parameter sequence of the prototype from the start of the test to the yield critical point, calculate the instantaneous disturbance power in each sliding time window and perform cumulative integration on the time axis, and set the average integral value of multiple sets of prototypes at the critical point as the reference cumulative disturbance energy value. To predetermine the upper limit of damage constant, since micro-fatigue cracks in materials in engineering practice usually enter the unstable propagation stage before the theoretical cumulative damage degree reaches 1, it is necessary to impose a truncation constraint to ensure that the cumulative fatigue damage degree is always less than 1 before the macroscopic fracture of the box occurs. The empirical range of the predetermined upper limit of damage constant is 0.7 to 0.9. In specific implementation, the implementers can determine its specific value based on the experimental evaluation results of the specific fatigue safety margin of the material of this type of metering box.

[0057] Furthermore, a first-in-first-out circular queue is established as a time alignment buffer for the slowly changing environmental parameter sequence, and the queue depth of the time alignment buffer is configured as follows: ,in The single thermal equilibrium recovery time constant of the intelligent metering box is given. The heat transfer process of the intelligent metering box is thermodynamically approximated as a first-order linear time-invariant system. According to Newton's law of cooling, its temperature difference decays exponentially with time. Therefore, the single thermal equilibrium recovery time constant is determined by the implementers through environmental step response testing of the intelligent metering box, taking the time required for the internal temperature response curve to decrease to the reciprocal of the natural constant of the initial temperature difference (approximately 37%). The sampling interval for the slowly changing environmental parameter sequence, As the rounding sign, when the time alignment buffer is full, the oldest data at the head of the queue is automatically discarded to maintain real-time rolling alignment.

[0058] Extract the actual sampled values ​​of the slowly varying environmental parameter sequence residing in the time alignment buffer at the current sampling time. Based on the first-order rate of change of the transient electromechanical parameter sequence and the cumulative fatigue damage, calculate the physical compensation margin of the slowly varying environmental parameter sequence at the current sampling time. The physical compensation margin satisfies the expression:

[0059]

[0060] In the formula, This represents the physical compensation margin corresponding to the current sampling time; This is a dimensionless, preset spatiotemporal coupling sensitivity coefficient used to adjust the relative mapping weight of the influence of electromechanical transients on the characteristics of the thermal environment. Its actual value is obtained through least-squares fitting of experimental data. The specific process is as follows: take multiple units in their undamaged factory condition (i.e., set the cumulative fatigue damage degree). A prototype of the same model intelligent metering box was subjected to multiple sets of transient mechanical and electrical impact excitations with different energy gradients in a controlled environment laboratory. For each set of excitations, the reference first-order rate of change test value of the transient electromechanical parameter sequence at the moment of impact was extracted simultaneously using sensors. And continuously monitor the time constant of a single thermal equilibrium recovery. Subsequently, the maximum passive response deviation amplitude of the slowly changing environmental parameters was measured. Divide the deviation test amplitude obtained from each test group by the maximum full-scale span value of the parameter. To obtain the actual evolution ratio Simultaneously, the corresponding first-order rate of change test value for each group is divided by the initial benchmark tolerance rate of change constant calibrated at the factory. Substituting this into the hyperbolic tangent function, we obtain the theoretical impact mapping independent variable. Collect multiple sets of incentives For the data pairs, a least squares method was used for intercept-free linear fitting, with the objective of minimizing the sum of squared residuals. The slope of the resulting regression line is the dimensionless spatiotemporal coupling sensitivity coefficient. The specific fitting formula is as follows: In the formula, and Representing the first The theoretical impact mapping independent variable and the actual evolution ratio obtained from the group experiment.

[0061] This indicates the maximum full-scale span value of the slowly changing environmental parameters inside the intelligent metering box of this model. It is obtained by the implementers through pre-measuring or retrieving the historical highest and lowest operating extreme values ​​of the equipment under a complete natural environmental cycle in a specific service climate area and calculating the difference between the two. This represents the first-order rate of change of a transient electromechanical parameter sequence; This represents the initial reference tolerance rate constant specified by the manufacturer. This represents the cumulative fatigue damage degree, characterizing the physical fatigue degradation state of the material, and is the cumulative fatigue damage degree before the box-type structure experiences macroscopic fracture. Always less than 1; It is a dimensionless fatigue reduction factor used as a constant representing the rate of change of the initial reference tolerance as the equipment service life extends. Dynamic real-time contraction is performed to restore the true critical physical limit of the enclosure against rupture at the current moment. The first-order rate of change of the transient electromechanical parameter sequence is used. The ratio of the actual physical limit after contraction is a dimensionless equivalent transient impact coefficient. To avoid the extreme amplitude of electromechanical transients causing the compensation amount to deviate numerically from the limits of the actual physical environment, this invention utilizes... The hyperbolic tangent function nonlinearly maps the equivalent transient impulse coefficient to Interval, first-order rate of change of transient electromechanical parameter sequence Larger or cumulative fatigue damage The higher the value, the greater the equivalent transient impact coefficient, leading to a more dramatic passive evolution of the slowly changing environmental parameters in the future, and thus a decrease in the physical compensation margin. This is followed by a positive, proportional scaling-up. Initial baseline tolerance rate constant. The data was obtained by the implementers through a destructive impact calibration experiment on a brand-new, sealed intelligent metering box. Specifically, a standard drop hammer impact tester was used to apply a vertical impact to the new box, with the impact energy increasing in a stepwise manner. High-frequency response data from the vibration sensor was collected simultaneously. After filtering out circuit noise using a Butterworth low-pass filter, the global peak value of the first derivative of the signal at the moment of impact was extracted. This peak value is the initial reference tolerance rate of change constant. .

[0062] S3: The physical compensation margin is superimposed on the real sampled values ​​of the slowly varying environmental parameter sequence to obtain environmental feature values. The expanded and aligned environmental feature values ​​are then concatenated with the transient electromechanical parameter sequence to obtain a spatiotemporally aligned feature vector. This vector is then combined with the defect classification label to generate a local fusion training set.

[0063] Specifically, the physical compensation margin at the current sampling moment Algebraically superimposed onto the actual sampled values ​​of the slowly varying environmental parameter sequence at the current sampling moment, the compensated environmental feature values ​​are obtained. Real-time sampled values ​​of the transient electromechanical parameter sequence at the current sampling moment are extracted. Due to the difference in array length between high and low frequency samples, zero-order hold interpolation is used to expand the array length of the compensated environmental feature values ​​upwards to match the array length of the real-time sampled values ​​of the transient electromechanical parameter sequence. The expanded environmental feature values ​​and the real-time sampled values ​​of the transient electromechanical parameter sequence are then concatenated as feature vectors to obtain a spatiotemporally aligned feature vector that eliminates the time difference in physical medium transmission. The power operation and maintenance management platform is accessed through a secure encrypted interface to obtain the on-site manual inspection record report of the smart metering box within the corresponding historical time period. The diagnosed metering box defect classification labels are extracted from the on-site manual inspection record report. The spatiotemporally aligned feature vectors of multiple consecutive sampling periods are combined one-to-one with the extracted metering box defect classification labels to generate a local fusion training set for use by the local computing node.

[0064] For example, Figure 3 This diagram illustrates the comparison between the actual sampled values ​​of the gradually varying environmental parameters and the compensated environmental characteristic values. During the period before a significant impact on the transient electromechanical parameters, the two curves essentially overlap, indicating that substantial compensation is unnecessary. At the moment of impact, the compensated environmental characteristic values ​​respond earlier and exhibit a significant peak, while the actual sampled values ​​of the gradually varying environmental parameters remain stable due to the thermal conduction hysteresis effect. In the subsequent period after the impact, the actual sampled values ​​of the gradually varying environmental parameters begin to rise and gradually approach the compensated characteristic values. This diagram demonstrates the effect of calculating the physical compensation margin using the first-order rate of change of the transient electromechanical parameter sequence and the cumulative fatigue damage, and then algebraically superimposing the compensation margin onto the actual sampled values ​​to eliminate the time lag in physical medium conduction.

[0065] S4: Use the local fusion training set to train the deployed isomorphic defect identification model, extract the discrete standard deviation of the physical compensation margin in the transient electromechanical parameter sequence under the steady state to calculate the aggregate weight coefficient, and upload the local model parameters and aggregate weight coefficient to the central coordination system for weighted aggregation update.

[0066] It should be noted that, in the traditional approach of federated learning architectures to perform global parameter aggregation, local model parameters from each computing node are typically received indiscriminately based on data volume or a fixed communication ratio. However, in real-world power scenarios, some nodes operate in harsh environments for extended periods, causing sensor baselines to experience temperature drift or aging, resulting in severe and irregular oscillations in the actual output physical compensation margin. To avoid misinterpreting normal operational load surges as sensor aging, this invention introduces a conditional filtering mechanism. This mechanism extracts the degree of discrete fluctuation in the compensation margin only when excluding real high-frequency electromechanical shocks, automatically reducing the aggregation weight of physically unstable nodes. By leveraging the stability of the front-end real signal, the communication aggregation process at the back end is constrained, improving the anti-interference capability under complex operating conditions.

[0067] Specifically, it is ensured that each local computing node deploys a homogeneous defect identification model with the same network topology. The defect identification model is an end-to-end nonlinear classifier, such as a one-dimensional convolutional neural network or a multilayer perceptron, which are deep learning models suitable for processing one-dimensional feature vectors. The local fusion training set is used as the input data source, and the spatiotemporally aligned feature vector is used as the input feature of the defect identification model. The output defect probability prediction value and the corresponding metering box defect classification label are substituted into the cross-entropy loss function to calculate the loss value. Based on the loss value, backpropagation is performed through the gradient descent algorithm to update the parameters of the local model after training.

[0068] Using the single thermal balance recovery time of the intelligent metering box as the model evaluation cycle, this invention uses this cycle to align with the independent physical rhythm of the box's heat exchange. Within the model evaluation cycle, when the sampling amplitude of the transient electromechanical parameter sequence continuously lies between the upper and lower boundary lines symmetrically constructed based on the positive and negative values ​​of the normal operating threshold, the continuous physical compensation margins calculated within that model evaluation cycle are statistically analyzed. Statistical standard deviation It should be noted that the normal business operation threshold has a unified numerical source in this invention. In step S2, the normal business operation threshold serves as the upper limit truncation boundary of historical data, used to strip away and extract the pure daily fatigue background sequence. In step S4, the normal business operation threshold serves as a real-time conditional filtering threshold, used to shield the interference of real destructive attacks on the assessment of the underlying hardware aging status of the device.

[0069] The aggregate weight coefficient of the local computing node is calculated based on the statistical standard deviation of the physical compensation margin within the model evaluation period. The aggregate weight coefficient satisfies the following expression:

[0070]

[0071] In the formula, This represents the aggregation weight coefficient of the local compute node; Indicates physical compensation margin The statistical standard deviation; This represents the factory-calibrated background hardware noise threshold of the local computing node, used to characterize the natural physical thermal noise of the device in an absolutely static state, and to prevent statistical standard deviation from causing noise when the system is in an absolutely ideal steady state. Risk of division-by-zero crash due to a value of zero. The aforementioned background hardware noise threshold. The calibration method is as follows: Place a reference metrology box of the same model equipped with sensors from the same batch in a standard laboratory environment with constant temperature and vibration isolation. When the system reaches physical stillness and thermal equilibrium, continuously collect the output signals of the slowly changing environmental parameter sensors and calculate the standard deviation of its basic white noise. Use the standard deviation of the basic white noise as the background hardware noise threshold. The dimensions of the background hardware noise threshold and the physical compensation margin Consistent. The aggregated weight coefficient aims to directly and equivalently map the data dispersion of the underlying hardware of the node to the communication trust of the top-level federated learning. When a node is in a harsh working condition that causes sensor aging or severe temperature drift, the physical compensation margin will fluctuate violently and cause the statistical standard deviation to increase sharply. At this time, by taking the reciprocal operation, the weight of the unstable node can be adaptively and significantly weakened, thereby effectively isolating the contaminated gradient generated by the damaged hardware from damaging the global model.

[0072] For example, Figure 4This is a schematic diagram illustrating the change of aggregation weight coefficients over time. Figure 4 The curve exhibits a stepped adjustment characteristic, reflecting the calculation mechanism that uses the single thermal equilibrium recovery time as the model evaluation period. During periods of small transient electromechanical parameter fluctuations, the statistical standard deviation of the physical compensation margin is small, and the aggregate weight coefficient remains at a high level. During periods of significant electromechanical parameter shocks, the physical compensation margin fluctuates significantly, the statistical standard deviation increases, and the aggregate weight coefficient decreases accordingly. This curve demonstrates the effect of mapping the data dispersion of the node's underlying hardware to the trust level of federated learning communication, and verifies the mechanism of adaptively weakening the weights of unstable nodes by taking the reciprocal operation.

[0073] The local compute node will store the local model parameters and the corresponding aggregate weight coefficients. The data packets are packaged and uploaded to the central coordination system, which parses the data packets from each node and assigns them an aggregate weight coefficient. The corresponding local model parameters are multiplied by the multipliers, all product results are summed, and then divided by the sum of the aggregate weight coefficients of all nodes to obtain the global model parameters after performing the weighted average calculation. These parameters are then distributed and updated to all local computing nodes. Simultaneously, to prevent issues arising during the system's cold start due to the incomplete statistical standard deviation of the first cycle... The parameter gaps that arise from the calculations, and the statistical standard deviation of all nodes under extreme and harsh operating conditions. The rapid increase in the sum of aggregated weights in the system poses a risk of numerical overflow, causing the total aggregated weights to approach zero. This invention introduces a normalization safety net mechanism into the central coordination system: when the statistical standard deviation calculated by all nodes participating in this round of aggregation is detected... If all values ​​exceed the preset failure baseline, or if a node is in a cold start period with no historical data available, the central coordination system will forcibly suspend the calculated aggregate weight coefficients. The aggregation strategy is degraded to an arithmetic average weight strategy with uniformly distributed weights across nodes to ensure the absolute robustness and anti-paralysis capability of the federated learning global model aggregation process under extreme boundary conditions. The specific method for setting the failure baseline is as follows: due to the introduction of a hyperbolic tangent function in the physical compensation margin expression... The theoretical upper limit of its absolute amplitude is mathematically defined, and a failure baseline is constructed accordingly: .in, Indicates the failure baseline. This represents the maximum full-scale span value of the slowly changing environmental parameters. This is the spatiotemporal coupling sensitivity coefficient; The preset chaos tolerance coefficient is used to characterize the standard deviation shrinkage ratio under extreme and severe operating conditions. Its empirical value range is 0.5 to 0.8. The definition of this interval is based on the statistical limit of the bounded oscillation of physical signals: that is, the theoretical standard deviation coefficient of an ideal uniform chaotic distribution is about 0.577, and the theoretical standard deviation coefficient of the limiting simple harmonic resonance is about 0.707. This invention uses the mathematical models of these two types of extreme and severe operating conditions as anchor points, and combines the filtering impedance inertia of the underlying hardware to extend the reasonable engineering margin to both ends, thereby strictly locking the tolerance interval of 0.5 to 0.8 and accurately defining the physical extreme boundary of the system's anti-disturbance.

[0074] The present invention also discloses a meter box early warning system based on federated learning and multimodal data fusion, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the meter box early warning method based on federated learning and multimodal data fusion according to the present invention is implemented.

[0075] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0076] In the description of this specification, "multiple" or "several" means at least two, such as two, three or more, unless otherwise expressly and specifically defined.

[0077] While this specification has shown and described numerous embodiments of the invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of this invention.

Claims

1. A metering bin early warning method based on federated learning and multimodal data fusion, characterized in that, include: The original sampling sequence of the intelligent metering box sensor is obtained, and the original sampling sequence is split into transient electromechanical parameter sequence and slowly changing environmental parameter sequence according to the medium conduction velocity, and a unified timestamp is assigned. Extract the historical variation characteristics of transient electromechanical parameter sequences, and construct the cumulative fatigue damage degree based on the fatigue background sequence when the enclosure is determined to be in its effective service life. Based on the historical variation characteristics of transient electromechanical parameter sequences and cumulative fatigue damage, the physical compensation margin of the slowly changing environmental parameter sequences at the current sampling time is calculated. The physical compensation margin is superimposed on the real sampled values ​​of the slowly varying environmental parameter sequence to obtain environmental feature values. The expanded and aligned environmental feature values ​​are then concatenated with the transient electromechanical parameter sequence to obtain a spatiotemporally aligned feature vector. This vector is then combined with the defect classification label to generate a local fusion training set. The isomorphic defect identification model is trained and deployed using a local fusion training set. The discrete standard deviation of the physical compensation margin under the steady state of the transient electromechanical parameter sequence is extracted to calculate the aggregate weight coefficient. The local model parameters and aggregate weight coefficients are then uploaded to the central coordination system for weighted aggregation and update.

2. The metering bin early warning method based on federated learning and multimodal data fusion according to claim 1, characterized in that, The extraction of historical variation characteristics of transient electromechanical parameter sequences, and the construction of cumulative fatigue damage based on fatigue background sequences when determining that the housing is within its effective service life, include: Extract the historical data window, extract the absolute value of the difference between all sampled values ​​of the transient electromechanical parameter sequence within the historical data window and the actual sampled value at the start time, and divide the largest absolute value among the absolute values ​​by the time span length to obtain the first-order rate of change of the transient electromechanical parameter sequence. The sampling amplitude of transient electromechanical parameter sequences, sensor communication status, and first-order rate of change of slowly changing environmental parameter sequences are monitored in real time. When the preset over-limit conditions are met, it is determined that the enclosure has macroscopically broken and an emergency physical cut-off and alarm mechanism is triggered. When the preset over-limit conditions are not met, it is determined that the enclosure is in its effective service life. The upper limit boundary line and the lower limit boundary line are symmetrically constructed with positive and negative values ​​of the normal business operation threshold. The transient electromechanical parameter sequence with sampling amplitude always between the upper limit boundary line and the lower limit boundary line is extracted as a fatigue background sequence to construct the cumulative fatigue damage degree.

3. The metering box early warning method based on federated learning and multimodal data fusion according to claim 1 or 2, characterized in that, The construction of cumulative fatigue damage based on fatigue background sequences includes: The discrete statistical variance of the fatigue background sequence within the sliding time window is calculated as the instantaneous disturbance power. The instantaneous disturbance power is accumulated and integrated over the equipment service time axis to obtain the cumulative disturbance energy value. Substituting the accumulated disturbance energy value into the normalized fatigue mapping function, the accumulated fatigue damage degree is obtained: In the formula, Indicates the cumulative fatigue damage degree. This represents the cumulative disturbance energy value. For reference, the cumulative disturbance energy value, This is a preset damage upper limit constant.

4. The metering box early warning method based on federated learning and multimodal data fusion according to claim 2, characterized in that, The physical compensation margin satisfies the expression: ； In the formula, This represents the physical compensation margin corresponding to the current sampling time; The dimensionless preset spatiotemporal coupling sensitivity coefficient; This indicates the maximum full-scale span value of the slowly changing environmental parameters inside the intelligent metering box of this model; This represents the first-order rate of change of a transient electromechanical parameter sequence; This represents the initial reference tolerance rate constant specified by the manufacturer. Indicates the cumulative degree of fatigue damage; This represents the hyperbolic tangent function.

5. The metering bin early warning method based on federated learning and multimodal data fusion according to claim 1, characterized in that, Before calculating the physical compensation margin of the slowly varying environmental parameter sequence at the current sampling time, the following steps are also included: A first-in-first-out circular queue is established as a time alignment buffer for the parameter sequence of a slowly changing environment. The queue depth of the time alignment buffer is configured as follows: ,in The queue depth for the time alignment buffer. This is the single thermal equilibrium recovery time constant of the intelligent metering box. The sampling interval for the slowly changing environmental parameter sequence, The rounding up symbol; When the time alignment buffer is full, discard the oldest data at the head of the queue, and extract the actual sampled value of the slowly varying environmental parameter sequence residing in the time alignment buffer at the current sampling time to calculate the physical compensation margin.

6. The metering box early warning method based on federated learning and multimodal data fusion according to claim 1, characterized in that, The step of concatenating the dimensionally aligned environmental feature values ​​with the transient electromechanical parameter sequence to obtain the spatiotemporally aligned feature vector includes: The zero-order hold interpolation method is used to expand the array length of environmental feature values ​​upward to match the array length of the real-time sampled values ​​of the transient electromechanical parameter sequence. The expanded environmental feature values ​​and the real-time sampled values ​​of the transient electromechanical parameter sequence are then concatenated to obtain the spatiotemporally aligned feature vector.

7. The metering bin early warning method based on federated learning and multimodal data fusion according to claim 2, characterized in that, The calculation of the aggregated weight coefficients for the discrete standard deviation of the physical compensation margin in the extracted transient electromechanical parameter sequence under stationary conditions includes: When the sampling amplitude of the transient electromechanical parameter sequence continuously lies between the upper and lower boundary lines symmetrically constructed based on the positive and negative values ​​of the normal business operation threshold during the model evaluation period, the statistical standard deviation of multiple consecutive physical compensation margins during the model evaluation period is calculated; the factory-calibrated background hardware noise threshold is obtained; and the reciprocal of the sum of the statistical standard deviation of the physical compensation margin and the background hardware noise threshold is used as the aggregation weight coefficient.

8. The metering box early warning method based on federated learning and multimodal data fusion according to claim 2 or 7, characterized in that, The normal business operation threshold is dynamically updated using an exponential moving average.

9. The metering bin early warning method based on federated learning and multimodal data fusion according to claim 4, characterized in that, The step of uploading local model parameters and aggregate weight coefficients to the central coordination system for weighted aggregation and updating includes: Upload the local model parameters and aggregate weight coefficients to the central coordination system; In response to the detection that the statistical standard deviation of the physical compensation margin calculated by all nodes participating in this round of aggregation exceeds the preset failure baseline. If a node is in a cold start period, a backup mechanism to prevent paralysis is triggered: the central coordination system forcibly suspends the aggregation weight coefficient calculated based on the statistical standard deviation, degenerates the aggregation strategy of this round into an arithmetic mean weight strategy with uniform distribution across computing nodes, and uses the arithmetic mean of the local model parameters uploaded by all local computing nodes participating in this round of aggregation as the global model parameters; wherein, the failure baseline Satisfying the formula: In the formula, The preset chaos tolerance coefficient is used to characterize the standard deviation shrinkage ratio under extreme and harsh operating conditions; The dimensionless preset spatiotemporal coupling sensitivity coefficient; This indicates the maximum full-scale span value of the slowly changing environmental parameters inside the intelligent metering box of this model; In response to the failure to trigger the anti-paralysis backup mechanism, the central coordination system uses the aggregate weight coefficient of each node as a multiplier to multiply the corresponding local model parameters, sums all the product results and divides them by the sum of the aggregate weight coefficients of all nodes to obtain the global model parameters. The acquired global model parameters will be distributed and updated.

10. A metering box early warning system based on federated learning and multimodal data fusion, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the metering box early warning method based on federated learning and multimodal data fusion according to any one of claims 1-9.

Images