A power distribution cabinet fault intelligent detection method and system

Abstract

The present application relates to the technical field of power system state monitoring, in particular to a power distribution cabinet fault intelligent detection method and system, the method comprising: setting a to-be-measured point and a reference point in the power distribution cabinet, and constructing a feature tuple of the current moment by the load current of the to-be-measured point and the instantaneous thermal resistance ratio; obtaining the actual joint probability density value and the theoretical joint probability density value of the feature tuple of the current moment through two-dimensional joint probability density estimation based on space-time self-adaptation; determining the anomaly score of the feature tuple of the current moment through anomaly analysis, determining the fault risk index of the current moment according to the relative deviation of the instantaneous thermal resistance ratio of the to-be-measured point of the current moment and the theoretical instantaneous thermal resistance ratio, and the anomaly score, and calculating the early warning index according to the continuous change trend of the fault risk index to judge whether the power distribution cabinet has a fault. The method improves the accuracy of power distribution cabinet fault detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system state monitoring. Specifically, it relates to a power distribution cabinet fault intelligent detection method and system. BACKGROUND

[0002] The power distribution cabinet is a key hub equipment of the power system, and the electrical connection points such as the switch stub head and the bus connection in the power distribution cabinet are the core components for ensuring the safe and stable operation of the system. In the long-term operation, these connection points are prone to early faults such as poor contact due to factors such as vibration, corrosion, and overload, which leads to an increase in contact resistance. According to Joule's law, the increased resistance will cause local abnormal heating when the current passes through. If this thermal fault is not discovered in time, it may gradually worsen and eventually cause equipment burnout or even fire, posing a serious safety hazard. Therefore, it is crucial to accurately and real-time monitor the health status of the key connection points in the power distribution cabinet.

[0003] In the prior art, a commonly used online monitoring scheme is to use an anomaly detection technology based on data driving. This technology collects the operating parameters of the equipment, usually constructs a two-dimensional feature vector composed of temperature rise and square of load current. Then, a local kernel density estimation-based outlier detection algorithm is used to train a normal behavior density model on a large amount of historical health operation data. In the monitoring process, this model is used to evaluate the local relative density of newly collected data points. When the density of a certain data point is significantly lower than the general density of healthy data points, it is determined to be in an abnormal or fault state.

[0004] However, the above-mentioned traditional method has inherent defects. The core of the problem is that a globally fixed kernel function bandwidth is usually used, but this cannot adapt to the non-uniform distribution characteristics of the power distribution cabinet health data in the feature space: the data is highly dense in the low load area, while it becomes sparse in the high load area. This mismatch causes the fixed bandwidth to be too large in scale and unable to effectively identify the subtle data deviation caused by early minor faults in the low load area, resulting in missed reporting of faults. In the high load area, the bandwidth is relatively too small in scale, often misjudging normal sparse data points as low-density outliers, thus generating a large number of false alarms, leading to inaccurate fault detection results of the power distribution cabinet. SUMMARY

[0005] To solve the problem in the prior art that the fault detection method based on fixed neighborhood scale cannot adapt to the non-uniform distribution characteristics of the power distribution cabinet data, resulting in missed reporting in the low load area and false alarms in the high load area, the present application provides a power distribution cabinet fault intelligent detection method and system.

[0006] In a first aspect, the present application provides a power distribution cabinet fault intelligent detection method, comprising:

[0007] Setting a to-be-measured point and a reference point in the power distribution cabinet, collecting load currents, temperatures of the to-be-measured point and the reference point, and an ambient temperature of the power distribution cabinet; at any time as a current time when the power distribution cabinet is running, determining an instantaneous thermal resistance ratio of the to-be-measured point based on the load currents, the temperatures and the ambient temperature, and constructing a feature tuple of the current time by the load current and the instantaneous thermal resistance ratio;

[0008] Pre-collecting a health sample set composed of feature tuples of the to-be-measured point at multiple times under a healthy state of the power distribution cabinet, and performing two-dimensional joint probability density estimation based on space-time adaptation based on the health sample set and the feature tuple of the current time to obtain an actual joint probability density value of the feature tuple of the current time;

[0009] Determining a theoretical instantaneous thermal resistance ratio corresponding to the load current of the to-be-measured point of the current time through the health sample set, and taking a joint probability density value of a feature tuple composed of the load current and the theoretical instantaneous thermal resistance ratio as a theoretical joint probability density value of the feature tuple of the current time;

[0010] Determining an anomaly score of the feature tuple of the current time through anomaly analysis, calculating a relative deviation of the instantaneous thermal resistance ratio of the to-be-measured point of the current time from the theoretical instantaneous thermal resistance ratio, taking a product of the anomaly score and the relative deviation as a fault risk index of the current time, and performing fault detection on the power distribution cabinet according to the fault risk index.

[0011] The technical scheme first constructs a feature tuple composed of a load current and an instantaneous thermal resistance ratio, eliminates interference of ambient temperature, load instantaneous fluctuation and heat dissipation difference at different positions, and provides a reliable physical quantity basis for fault detection; then, two-dimensional joint probability density estimation based on space-time adaptation is adopted, which can dynamically adjust the bandwidth of kernel density estimation according to the non-uniform distribution characteristics of the power distribution cabinet operation data, i.e., low load density and high load sparsity, to capture small abnormalities with narrow bandwidth for low load and avoid misjudgment of normal sparse points with wide bandwidth for high load, and to accurately quantify the statistical position of the current state in the health distribution by combining the time weight of the sample to adapt to the slow drift of the health baseline; then, the theoretical instantaneous thermal resistance ratio under different load currents is fitted through the health sample, and the theoretical joint probability density is calculated to provide a clear reference point for anomaly determination; finally, the relative deviation of the instantaneous thermal resistance ratio reflects the degradation of physical performance, and the anomaly score of the joint probability density reflects the deviation of the statistical distribution, and the fault risk index calculated by fusing the two can accurately reflect the health state of the power distribution cabinet, thereby improving the accuracy of fault detection of the power distribution cabinet.

[0012] Preferably, the instantaneous thermal resistance ratio of the to-be-tested point is determined based on the following manner: the reference point is set to be composed of a plurality of physical reference points; for each physical reference point, a coefficient of variation of thermal resistance thereof within a preset historical time window is determined, and an inverse of the coefficient of variation is determined as a stability index of the physical reference point; a weight of each physical reference point is obtained by dividing the stability index of the physical reference point by a sum of stability indexes of all physical reference points; an equivalent reference thermal resistance is obtained by weighted average of thermal resistances of all physical reference points at the current time instant, using the weight of each physical reference point; and a ratio of the thermal resistance of the to-be-tested point to the equivalent reference thermal resistance is determined as the instantaneous thermal resistance ratio of the to-be-tested point.

[0013] Preferably, the theoretical joint probability density value of the feature tuple at the current time instant and the actual joint probability density value of the feature tuple at the current time instant are both obtained based on a spatio-temporal adaptive two-dimensional joint probability density estimation.

[0014] Preferably, the actual joint probability density value of the feature tuple at the current time instant satisfies the following relationship:

[0015]

[0016] In the formula, f (I, R) is the actual joint probability density value of the feature tuple at the current time instant, is the feature tuple at the current time instant, is the actual joint probability density value of the feature tuple at the current time instant, and I and R are respectively a load current and an instantaneous thermal resistance ratio of the to-be-tested point at the current time instant, and n and N are respectively a serial number and a total number of the feature tuples in the healthy sample set, is a time weight of the nth feature tuple in the healthy sample set, is a Gaussian kernel function, and I n and R n are respectively a load current and an instantaneous thermal resistance ratio of the nth feature tuple, is an adaptive bandwidth for obtaining the load current, is an adaptive bandwidth for obtaining the instantaneous thermal resistance ratio, is an adaptive bandwidth for is an adaptive bandwidth for

[0017] ​​​​​​​This technical solution achieves precise adaptation to the health data characteristics of the distribution cabinet through a triple adaptive mechanism. The time weighting makes the contribution of recent health samples to the joint probability density estimation at the current moment greater, effectively adapting to the slow drift of the equipment health baseline over time. The spatiotemporal adaptive bandwidth of the load current dimension can be dynamically adjusted according to the magnitude of the load current. In low load dense areas, the bandwidth is automatically reduced to focus on local samples to capture minor anomalies. In high load sparse areas, the bandwidth is automatically expanded to avoid the distortion of probability density estimation caused by the scarcity of samples. The spatial adaptive bandwidth of the instantaneous thermal resistance ratio dimension adjusts the neighborhood range according to the distribution characteristics of the thermal resistance ratio to ensure the consistency of density calculation under different thermal resistance states. The final output can truly reflect the statistical position of the current state in the health distribution.

[0018] Preferably, the distance between the time of each feature tuple in the healthy sample set and the current time is calculated, and the distance is used as a negative indicator of the weight to construct a relationship of weight decay with distance, so as to calculate the time weight of each feature tuple in the healthy sample set.

[0019] Preferably, the adaptive bandwidth of the load current and the adaptive bandwidth of the instantaneous thermal resistance ratio are determined based on the following methods:

[0020] For each load current, through The nearest neighbor matching algorithm finds the closest sample to the load current in the healthy sample set. Each load current value is calculated. The furthest distance between a load current value and the load current value is determined as the adaptive bandwidth of the load current. For each instantaneous thermal resistance ratio, a local range is set with the load current corresponding to the instantaneous thermal resistance ratio as the center. All feature tuples whose load current falls within the local range are screened in the healthy sample set, and the standard deviation of the instantaneous thermal resistance ratio corresponding to all screened feature tuples is calculated. The standard deviation is determined as the adaptive bandwidth of the instantaneous thermal resistance ratio.

[0021] This technical solution is achieved through The nearest neighbor matching method addresses the macroscopic problem of uneven global data distribution, while the local standard deviation method addresses the microscopic problem of variations in local data characteristics. This anisotropic adaptive mechanism ensures that every step of the probability density estimation process closely matches the physical and statistical characteristics of real data, thus providing an accurate and reliable foundation for the entire fault detection system.

[0022] Preferably, the theoretical instantaneous thermal resistance ratio corresponding to the load current of the test point at the current moment is determined in the following way: a two-dimensional feature space is constructed with the load current as the horizontal axis and the instantaneous thermal resistance ratio as the vertical axis. Each feature tuple of the healthy sample set corresponds to a feature point in the two-dimensional feature space. All feature points are fitted by polynomial regression to obtain a health center line. The instantaneous thermal resistance ratio corresponding to the load current of the test point at the current moment on the health center line is obtained as the theoretical instantaneous thermal resistance ratio corresponding to the load current of the test point at the current moment.

[0023] Preferably, the anomaly score of the feature tuple at the current time is determined based on the following method:

[0024] In the statistical health sample set, the number of characteristic tuples whose load current falls within the local range of the load current at the test point at the current moment is denoted as . Calculate the anomaly score of the feature tuple at the current time:

[0025]

[0026] In the formula, The feature tuple at the current time. Abnormal scores, and They are respectively The actual joint probability density value and the theoretical joint probability density value, It is a natural exponential function. Let V be the variance of the instantaneous thermal resistance ratio of all characteristic elements whose load current is within this local range. To find the maximum value function, To prevent parameters with a denominator of 0.

[0027] This technical solution, through the collaborative design of core statistical deviation quantification and local data characteristic dynamic calibration, relies on the health sample statistics of the current load locality. It focuses on the actual data environment of the current working condition and achieves abrupt calibration through the smoothing characteristics of the exponential function, ensuring that the abnormal score can accurately map the true degree of abnormality across the entire load range, while taking into account both detection sensitivity and anti-interference capability.

[0028] Preferably, fault detection of the power distribution cabinet based on the fault risk index is performed as follows: All feature tuples within a preset time window prior to the current moment are obtained as the historical feature tuple set for the current moment, and the early warning index is calculated.

[0029]

[0030] In the formula, As the early warning indicator for the current moment, a total number of feature tuples in the historical feature tuple set, a failure risk index of the i-th feature tuple in the historical feature tuple set, a failure risk index of the i-th feature tuple in the historical feature tuple set, a failure risk index of the i-th feature tuple in the historical feature tuple set, is an indicator function, the function value is 1 when the condition in the bracket is true, and the function value is 0 when the condition in the bracket is not true; if exceeds the preset warning index threshold, it is determined that the power distribution cabinet fails at the current moment; if does not exceed the preset warning index threshold, it is determined that the power distribution cabinet does not fail at the current moment.

[0031] The technical scheme analyzes the relative size of the failure risk index in the historical time window and the failure risk index at the current moment, and constructs a cumulative effect function, effectively avoiding the problem that the failure risk index at a single moment is disturbed by accidental fluctuations, improving the identification ability of the gradual failure, and improving the accuracy of the failure detection.

[0032] In a second aspect, the present application also provides a power distribution cabinet failure intelligent detection system, which comprises a memory and a processor, the memory stores a computer program, and the processor executes the computer program to realize the steps of any one of the power distribution cabinet failure intelligent detection methods.

[0033] The present application has the following effects:

[0034] The present application provides a high-reliability physical quantity basis for failure detection by constructing feature tuples, based on a two-dimensional joint probability density estimation of space-time adaptation, dynamically matching the non-uniform distribution characteristics of data, ensuring the sensitivity of small abnormalities in the low load area, avoiding false abnormalities in the high load area, realizing the dual consideration of physical degradation nature and statistical verification through the failure risk index, and finally determining the failure detection result through cumulative effect analysis of the historical time window, which can accurately identify the gradual failure of the power distribution cabinet, and improve the accuracy and reliability of the power distribution cabinet failure detection. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 is a method flowchart of the present application. DETAILED DESCRIPTION

[0036] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0037] Referring to Figure 1 , a power distribution cabinet failure intelligent detection method specifically comprises the following steps:

[0038] S1: Construct a feature tuple of the current time according to the real-time collected multi-dimensional data.

[0039] This step eliminates the direct influence of ambient temperature by setting the test point and reference point and calculating the temperature rise. The thermal resistance of the test point and the reference point is calculated, and the thermal resistance ratio of the test point is calculated by comparing the thermal resistance of the test point with a highly stable equivalent reference thermal resistance. The instantaneous thermal resistance ratio of the test point is obtained, and the load current of the test point is combined to form a feature tuple.

[0040] In the fault detection of the power distribution cabinet, the load current of the test point at the current time and the instantaneous thermal resistance ratio of the test point are selected to construct the feature tuple, which is based on the reasonable selection of the physical mechanism of the power equipment fault and the characteristics of the monitoring data. Because the physical representation of the core fault of the power distribution cabinet (such as the switch stub, the bus connection contact) is the abnormal increase of the contact resistance, according to the Joule law, the increase of the resistance will cause local abnormal heating, and the thermal resistance is a physical quantity that directly reflects the correlation characteristics of the current and the temperature rise. The instantaneous thermal resistance ratio is calculated by comparing the thermal resistance of the test point with the equivalent reference thermal resistance, which can further eliminate the environmental interference such as the overall temperature field fluctuation in the cabinet and the individual error of the sensor. Compared with a single temperature or temperature rise parameter, it can focus on the resistance abnormality of the test point itself. At the same time, the load current is a key variable that affects the thermal resistance representation. Under the same resistance, different currents will cause different temperature rises. If only the thermal resistance is used and the current is ignored, it will not be able to distinguish between the normal thermal resistance under normal current and the normal thermal resistance under high current, which will lead to fault misjudgment. Therefore, the combination of the two can form the dynamic correlation characteristics of the current and the thermal resistance, and completely map the health status of the test point under the current load working condition, providing accurate physical quantity basis for subsequent fault analysis.

[0041] Specifically, this step includes the following operations:

[0042] S11: Real-time acquisition of multi-dimensional data.

[0043] In the power distribution cabinet, select the electrical connection points (such as switch stubs and bus lap joints) that need to be monitored as test points. At the same time, select reference points, which are composed of one to several physical reference points away from heat sources and stable structures. Install temperature sensors and current sensors (or obtain current data through the bus) for each test point and each physical reference point, and install an environmental temperature sensor in a suitable position in the power distribution cabinet. The data is collected synchronously at a frequency of 1 second per second. For example, select the bus lap joint as a test point, and select 5 physical reference points away from heat sources and stable structures.

[0044] S12: Determine the temperature rise of the test point and the temperature rise of the reference point.

[0045] The definition of temperature rise is the temperature of the electronic components above the ambient temperature. During the operation of the power distribution cabinet, the temperature of the measured point (such as the head of the switch post and the connection of the bus) should be higher than the ambient temperature due to the Joule heat generated by the current. The temperature of the reference point (a physical reference point away from the heat source) should also be close to or slightly higher than the ambient temperature. The temperature rise is usually positive.

[0046] At any moment of the operation of the power distribution cabinet, the temperature rise of the measured point and the temperature rise of the reference point are determined based on the temperature and the ambient temperature at the current moment, i.e. the temperature of the measured point is subtracted from the ambient temperature to obtain the temperature rise of the measured point, and the temperature of the reference point (each physical reference point) is subtracted from the ambient temperature to obtain the temperature rise of the reference point (each physical reference point).

[0047] S13: Calculate the thermal resistance of the measured point and the reference point based on the temperature rise of the measured point and the current, and the temperature rise of the reference point and the current.

[0048] At the current moment, the temperature rise of the measured point is divided by the square of its load current to obtain the thermal resistance of the measured point. Similarly, the temperature rise of the reference point (each physical reference point) is divided by the square of its load current to obtain the thermal resistance of the reference point (each physical reference point).

[0049] This way is to reverse the thermal resistance by the temperature rise and the square of the load current in the thermal equilibrium state. Specifically, when the load current increases, if the measured point is not faulty, the temperature rise will increase by the square of the load current, and the calculated thermal resistance will remain stable. When the measured point fails, the temperature rise increases, even if the current remains unchanged, the temperature rise will increase, and the calculated thermal resistance will abnormally increase.

[0050] The thermal resistance calculated in this way is a fault feature quantity, which can remain stable when the device is healthy and can change significantly when the device fails.

[0051] S14: Determine the instantaneous thermal resistance ratio of the measured point.

[0052] In order to further eliminate the interference of the overall temperature fluctuation in the power distribution cabinet and the individual differences of the sensors, this step adopts multiple physical reference points to construct an equivalent reference thermal resistance with stronger robustness, and then calculates an instantaneous thermal resistance ratio based on the thermal resistance of the measured point and the equivalent reference thermal resistance.

[0053] Specifically, it includes:

[0054] Select 100 time points before the current moment as the historical time period of the current moment;

[0055] For any physical reference point, calculate all thermal resistances in the historical time period, calculate the coefficient of variation of all thermal resistances, the coefficient of variation is the ratio of the standard deviation to the mean, which is a dimensionless relative fluctuation index, and it is very scientific to select it as the stability criterion because it eliminates the influence of different measurement point thermal resistance mean size, and can fairly compare the fluctuation degree of different physical reference points at their respective reference level, the smaller the coefficient of variation, the more stable the physical characteristics, therefore, the reciprocal of the coefficient of variation is taken as the stability index of the physical reference point. The greater the stability index of a physical reference point, the more reliable its data, and the greater its contribution in calculating the equivalent reference thermal resistance. Therefore, the stability index of each physical reference point is divided by the sum of the stability indexes of all physical reference points to obtain the weight of each physical reference point.

[0056] At the current time, a weighted average operation is performed according to the thermal resistance of each physical reference point and the weight of each physical reference point, to obtain an equivalent reference thermal resistance. The ratio of the thermal resistance of the to-be-measured point to the equivalent reference thermal resistance is taken as the instantaneous thermal resistance ratio of the to-be-measured point, which physically reflects the deviation of the to-be-measured point from a stable and healthy reference, effectively eliminating the influence of environmental temperature and load fluctuation.

[0057] S15: The load current of the to-be-measured point and the instantaneous thermal resistance ratio constitute the feature tuple of the current time.

[0058] The load current of the to-be-measured point at the current time is denoted as , the instantaneous thermal resistance ratio of the to-be-measured point is denoted as , and the feature tuple of the current time is obtained .

[0059] S2: Based on the spatio-temporal adaptive two-dimensional joint probability density estimation, the actual joint probability density value and the theoretical joint probability density value of the feature tuple of the current time are calculated.

[0060] The calculation of the actual and theoretical joint probability density values based on the spatio-temporal adaptive two-dimensional joint probability density estimation is the core link between the construction of the feature tuple and the fault risk judgment. The foregoing steps construct the feature tuple through the load current and the instantaneous thermal resistance ratio, eliminating the interference of environmental temperature and load instantaneous fluctuation.

[0061] This step further considers that in the long-term use process of the power distribution cabinet, as the use time continues to lengthen, the cabinet body of the power distribution cabinet will produce heat dissipation efficiency fluctuation, component thermal property slow change caused by normal aging of equipment, and heat dissipation capacity attenuation caused by dust accumulation of ventilation filter, etc., leading to slow drift of the feature tuple distribution under healthy state (i.e. healthy baseline).

[0062] If the fixed health model is used, such as only based on the initial health sample to construct a static density model for fault detection, it will lead to misjudgment of the health state after drift as abnormal, or unable to identify early minor faults in the baseline drift background, affecting the detection accuracy.

[0063] Therefore, by pre-constructing a health sample set and introducing a spatiotemporal adaptive two-dimensional joint probability density estimation, on the one hand, the historical data of the health sample set can be used to provide a dynamically updated health reference point for the current time feature tuple, and through the adaptive adjustment of the time dimension, the recent health sample is given a higher weight to adapt to the slow drift of the health baseline, avoiding the time bias of the static model. On the other hand, the two-dimensional joint probability density estimation can simultaneously depict the cooperative distribution characteristics of the load current and the instantaneous thermal resistance ratio in two dimensions. Compared with single-dimensional density analysis, it can more comprehensively reflect the statistical rules of the feature tuple under the health state, and the spatiotemporal adaptive mechanism can dynamically adjust the kernel function bandwidth according to the non-uniform distribution of the feature tuple in different load intervals, ensuring that the probability density value can be accurately calculated in the full load range, and providing a reliable quantitative standard for distinguishing between normal baseline drift and abnormal fault deviation.

[0064] Specifically, the method comprises the following steps:

[0065] S21: Collect a health sample set composed of feature tuples of the to-be-measured point at multiple time points under the health state of the power distribution cabinet. For example, 1000 feature tuples of the to-be-measured point at multiple time points under the health state of the power distribution cabinet are collected in advance to form a health sample set, so as to ensure rich data samples.

[0066] S22: Set the time weight of each feature tuple in the health sample set.

[0067] Calculate the distance between the time of each feature tuple in the health sample set and the current time; take the distance as a negative index of the weight, and construct a relationship formula in which the weight decays with the distance, so as to calculate the time weight of each feature tuple in the health sample set. For example, the reciprocal of the distance between the time of each feature tuple and the current time is taken as the weight of the feature tuple.

[0068] S23: Determine the actual joint probability density value of the feature tuple at the current time.

[0069] Through a weighted anisotropic adaptive kernel density estimation process, the probability density of the current device state point in the health model is calculated based on a large number of historical health samples, that is, the joint probability density value of the two-dimensional index represented by the feature tuple at the current time in the historical health sample set. The higher the joint probability density value is, the more similar the current state is to the health state, and vice versa.

[0070] Specifically, the actual joint probability density value of the feature tuple at the current time satisfies the following relationship:

[0071]

[0072] In the formula, is the actual joint probability density value of the feature tuple at the current time , which quantifies the normality of the current device state. The larger the actual joint probability density value, the healthier the power distribution cabinet at the current time, and the less likely it is to fail. Conversely, the smaller the actual joint probability density value, the more likely it is to fail. The summation operation indicates that the final joint probability density value is the superposition of the joint contributions of all healthy samples. Each healthy sample has an impact on the actual joint probability density value of the feature tuple at the current time . and are the load current and the instantaneous thermal resistance ratio of the point to be tested at the current time, respectively, and are the serial number and the total number of the feature tuples in the healthy sample set, is the time weight of the th feature tuple in the healthy sample set, which is the key to achieving time adaptation. It gives different importance to each healthy sample. The closer the sample to the current time, the greater the weight, and the greater the contribution. In the calculation process, it can focus on the recent device state. is a Gaussian kernel function, which is essentially a distance and weight converter, the closer is, the greater its value is. and are the load current and the instantaneous thermal resistance ratio of the th feature tuple, respectively, is the adaptive bandwidth for obtaining the load current, is the adaptive bandwidth for obtaining the instantaneous thermal resistance ratio, is the adaptive bandwidth of , which dynamically adapts according to the load current , is the adaptive bandwidth of , which dynamically adapts according to the instantaneous thermal resistance ratio . The load current and the instantaneous thermal resistance ratio use different adaptive bandwidth methods, reflecting the anisotropy in these two dimensions.

[0073] In this formula, is a weighted normalization factor that ensures the total integral of all weighted kernel functions is 1, so that the final calculated is a strict probability density value, which is the standard form of weighted kernel density estimation.

[0074] In this formula, Part of the calculation involves kernel density estimation along the load current dimension. Its core task is to accurately assess the density of the current load current relative to historical healthy samples. A larger value in this part indicates a greater similarity between the current feature tuples and those in the healthy sample set along the load current dimension. The Gaussian function is characterized by a larger output value the smaller the absolute value of its input; its maximum value is achieved when the input is 0. Here, the input is... To make this input approach 0 means that the molecule... The smaller the difference, the better. and The closer. This is due to the adaptive bandwidth adjustment mechanism. Unlike the instantaneous thermal resistance ratio, the load current is an independent variable determining the equipment's operating condition, and its health data exhibits a typical globally non-uniform distribution: highly concentrated in low-load areas and extremely sparse in high-load areas. To address this characteristic, the adaptive bandwidth of the load current is determined as follows: for each load current, the k-nearest neighbor matching algorithm is used to find the nearest neighbor to that load current in the set of healthy samples. Each load current value and calculate The furthest distance between a load current value and that load current value is determined as the adaptive bandwidth of that load current.

[0075] Thus, in low-load areas with dense data, the bandwidth is automatically narrowed, enabling the model to distinguish details and identify minor anomalies; in high-load areas with sparse data, the bandwidth is automatically widened, taking into account normal but distant operating points, thereby effectively avoiding false alarms. This is used to scale the Gaussian kernel function, ensuring that the sum of the probability densities contributed by a single healthy sample is always 1, regardless of whether the bandwidth is narrowed or widened. This eliminates the direct influence of bandwidth size itself on the density value calculation, guaranteeing the fairness and effectiveness of density comparisons across the entire load range.

[0076] In this formula, Part of this involves kernel density estimation along the instantaneous thermal resistance ratio dimension. Its core task is to accurately assess the instantaneous thermal resistance ratio of the feature tuples at the current moment. The larger this value is relative to the density of healthy samples, the more similar the feature tuples at the current moment are to the feature tuples in the set of healthy samples in terms of instantaneous thermal resistance ratio. The determination method for health data points differs from that for load current; it reflects the inherent dispersion of health data points under the current load conditions.

[0077] The adaptive bandwidth of the instantaneous thermal resistance ratio is determined as follows: for each instantaneous thermal resistance ratio, a local range is set with the load current corresponding to the instantaneous thermal resistance ratio as the center. For example, the lower limit of the local range of the load current is set to be the load current minus 5 , and if it is negative, the minimum is limited to 0, and the upper limit is the load current plus 5 . When the power distribution cabinet is running, small fluctuations in the load current can be regarded as similar working conditions, and the thermal characteristics (such as thermal resistance and heat dissipation behavior) in this range have continuity and homogeneity, which can ensure that the local statistics reflect the true health benchmark under this working condition. Then, all feature tuples in the health sample set whose load currents fall within the local range are filtered out, and the standard deviation of the instantaneous thermal resistance ratio corresponding to all filtered feature tuples is calculated, and the standard deviation is determined as the adaptive bandwidth of the instantaneous thermal resistance ratio.

[0078] In this way, when is small, it indicates that the healthy instantaneous thermal resistance ratio value under this working condition is very stable and concentrated. At this time, a smaller bandwidth will make the kernel function become sharp, and even if deviates slightly from the health center line, the probability density value will also decrease sharply, thereby reflecting the high detection accuracy of the model under this stable working condition. Similarly, part of the function is to perform scale normalization on the kernel function, ensuring that the change in bandwidth does not affect the total amount of probability contributed by each sample, ensuring fairness and mathematical rigor when performing density estimation in different discrete degree regions.

[0079] In addition, since the adaptive bandwidth of the load current is essentially a distance value, the adaptive bandwidth of the instantaneous thermal resistance ratio is essentially a standard deviation. In order to prevent or from being 0, resulting in an unreasonable situation in the formula where the denominator is 0, it is set that if or is 0 in an extreme case, a very small positive number is added to or to avoid the interference caused by this extreme unreasonable situation. For example, this very small positive number is set to .

[0080] In summary, the joint probability density value calculated by this formula is the weighted sum of the contributions of all healthy samples. Therefore, if many healthy samples have feature sets (corresponding load current and instantaneous thermal resistance ratio) that are very close to the feature sets (corresponding load current and instantaneous thermal resistance ratio) at the current moment, a high joint probability density value will be obtained after weighting and summing them all. This increases the probability that the distribution cabinet is in a healthy state at the current moment, and vice versa.

[0081] S24: Determine the theoretical joint probability density value of the feature tuples at the current time.

[0082] The theoretical instantaneous thermal resistance ratio corresponding to the load current of the test point at the current moment is determined by using a set of healthy samples. This includes: constructing a two-dimensional feature space with the load current on the horizontal axis and the instantaneous thermal resistance ratio on the vertical axis; obtaining a feature point in the two-dimensional feature space for each feature tuple of the healthy sample set; performing polynomial regression fitting on all feature points to obtain a health center line; and taking the instantaneous thermal resistance ratio corresponding to each load current on the health center line as the theoretical instantaneous thermal resistance ratio corresponding to that load current. Thus, the corresponding theoretical instantaneous thermal resistance ratio can be found in the two-dimensional space based on the load current of the test point at the current moment.

[0083] The characteristic element consisting of the load current and the theoretical instantaneous thermal resistance ratio of the test point is denoted as the theoretical characteristic element at the current moment. The joint probability density value of the theoretical characteristic element is calculated according to the method in step S23, and is used as the theoretical joint probability density value of the characteristic element at the current moment.

[0084] S3: Determine the relative deviation between the instantaneous thermal resistance ratio of the test point and the theoretical instantaneous thermal resistance ratio, as well as the abnormal score of the characteristic tuple at the current moment.

[0085] S31: Determine the anomaly scores of the feature tuples at the current moment through anomaly analysis.

[0086] In the statistical health sample set, the number of characteristic tuples whose load current falls within the local range of the load current at the test point at the current moment is denoted as . The variance of the instantaneous thermal resistance ratio of all characteristic elements with load current within this local range is denoted as... ;

[0087] Calculate the anomaly score of the feature tuple at the current time step:

[0088]

[0089] In the formula, The feature tuple at the current time. Abnormal scores, and They are respectively The actual joint probability density value and the theoretical joint probability density value, It is a natural exponential function. To find the maximum value function, To prevent parameters with a denominator of 0, set to .

[0090] This formula combines a basic statistical bias measure with a dynamic, data-quality-based calibration factor to generate an anomaly score that is both sensitive and robust. It can be understood by breaking it down into two core components.

[0091] Part 1 Used for basic anomaly quantification, this part is the core source of the anomaly score, reflecting how far the current state deviates from the normal range. It is a normalized probability ratio that compares the absolute probability density with the theoretically healthiest point under the current operating conditions, rather than directly using the absolute probability density. This eliminates the influence of load variations on the magnitude of the probability density itself, making the comparison fairer. When the equipment is healthy, this ratio is close to 1; when an anomaly occurs, the ratio approaches 0. (Logarithmic function) Its function is to amplify differences. For example, when the probability ratio drops from 0.1 to 0.01, its logarithm changes from -1 to -2, a significant change. This makes the module highly sensitive to even small probability decays. This structure converts logarithmic values ​​(usually negative or 0) into a positive, increasing score. The more anomalous the state, the smaller the probability ratio, and the more negative the logarithmic value, the higher the final score. It constitutes a sensitive statistical bias detector, capable of transforming small probability deviations into significant scores. The function here guarantees that regardless of the input, the final output of this part will not be less than 0. It focuses the model's core attention on detecting state deterioration anomalies where the probability ratio decreases, while non-fault deterioration cases with a probability ratio much greater than 1 are uniformly considered as no anomalies, i.e., the output is 0. for If the value is 10 times or more (which is usually unlikely, but to prevent such an extreme case), then the part is negative and is restricted to 0, thus avoiding logical confusion and calculation problems caused by negative numbers.

[0092] Part Two It is equivalent to a confidence-based calibration factor. A smaller value indicates that there are very few healthy samples in the entire healthy sample set with a load current similar to the current moment. This suggests that the current distribution cabinet is operating under an unusual condition, such as an extremely high or unstable load range. The model lacks sufficient historical data to construct a reliable health reference standard, and the confidence level of the basic statistical bias metric needs to be reduced. Conversely, if... The greater, the more the current power distribution cabinet is running in a common working condition, the confidence of the basic statistical deviation measure needs to be improved, the confidence of the basic statistical deviation measure needs to be improved; The greater, the greater the fluctuation of the instantaneous thermal resistance ratio corresponding to the health sample similar to the current load condition in the whole health sample set, the worse the data stability, the confidence of the basic statistical deviation measure needs to be reduced, and vice versa, The smaller, the better the data stability, the confidence of the basic statistical deviation measure needs to be improved.

[0093] Therefore, by Implement the correction logic when The smaller, The greater, the current power distribution cabinet is running in an unusual working condition, and the greater the fluctuation of the instantaneous thermal resistance ratio corresponding to the health sample similar to the current load condition, the unstable data quality, The smaller, the confidence of the basic statistical deviation measure needs to be reduced; otherwise, the current power distribution cabinet is running in a common working condition, and the smaller the fluctuation of the instantaneous thermal resistance ratio corresponding to the health sample similar to the current load condition, the stable data quality, The greater, the confidence of the basic statistical deviation measure needs to be improved.

[0094] In this way, the dynamic sensitivity adjustment based on confidence is realized. The more health samples similar to the current working condition, the more stable the data, the closer the confidence to 1, and the abnormal score is determined by the basic statistical deviation measure, which ensures the maximum detection ability of early minor faults. The fewer health samples similar to the current working condition, the greater the data fluctuation, the closer the confidence to 0, and the basic statistical deviation measure is actively suppressed, thereby effectively avoiding false alarms caused by sparse data.

[0095] In summary, the two parts are fused by multiplication to realize the dual judgment of fault physical nature and statistical verification, avoid the limitations of single index (such as pure physical index is easy to be disturbed by environment, pure statistical index is easy to be disturbed by data noise), and conform to the scientific principle of multi-dimensional cross verification for fault detection.

[0096] S32: Determine the relative deviation of the instantaneous thermal resistance ratio of the to-be-measured point from the theoretical instantaneous thermal resistance ratio.

[0097] According to the classical relative deviation calculation formula:

[0098]

[0099] Among them, is the relative deviation of the instantaneous thermal resistance ratio of the to-be-measured point from the theoretical instantaneous thermal resistance ratio, is the maximum value function, the relative deviation of the instantaneous thermal resistance ratio of the to-be-tested point at the current moment, the theoretical instantaneous thermal resistance ratio of the to-be-tested point at the current moment. The greater the value is, the greater the difference between the current state and the ideal state is, and the more likely the health state of the power distribution cabinet has a problem, and vice versa.

[0100] On the one hand, the classical relative deviation form is adopted to eliminate the numerical and dimensional interference of the instantaneous thermal resistance ratio, so that the abnormality degree under different loads is comparable; on the other hand, by taking the maximum function to only retain the positive deviation, the increase in thermal resistance, which is the core physical direction of the power distribution cabinet failure, is accurately focused on, which can not only sensitively capture the small thermal resistance deviation of early failure, but also filter non-failure fluctuations to reduce false alarms.

[0101] S4: Fusion of relative deviation and abnormal score to determine the fault risk index at the current moment.

[0102] The relative deviation of the instantaneous thermal resistance ratio of the to-be-tested point and the theoretical instantaneous thermal resistance ratio, and the abnormal score of the feature tuple at the current moment are fused in a multiplicative manner to obtain the fault risk index at the current moment. Specifically, the following relationship is satisfied:

[0103]

[0104] In the formula, is the fault risk index at the current moment, is the relative deviation of the instantaneous thermal resistance ratio of the to-be-tested point and the theoretical instantaneous thermal resistance ratio, is the abnormal score of the feature tuple at the current moment.

[0105] In the formula, Focusing on the relative deviation of the instantaneous thermal resistance ratio, the absolute degradation degree of the fault core physical quantity (thermal resistance) is accurately reflected, Based on the statistical distribution anomaly of the feature tuple, the relative abnormality degree under the comparison of healthy samples is reflected.

[0106] By fusing the relative deviation and the abnormal score, the dual-dimensional cross verification of physical degradation and statistical anomaly is realized, avoiding the limitations of a single index (pure physics is susceptible to environmental interference or pure statistics is susceptible to noise misjudgment); the balance between sensitivity and reliability is also achieved. If a small physical deviation is accompanied by significant statistical anomaly, it can be effectively captured to improve the early failure identification capability, while a single-dimensional false fluctuation will be suppressed due to the other dimension not being significant, reducing the false alarm rate. The power distribution cabinet fault risk is comprehensively and accurately quantified, providing a reliable basis for state monitoring and early warning.

[0107] S5: Fault detection of the power distribution cabinet according to the fault risk index.

[0108] obtain all feature tuples in a preset time window (for example, set to 1 minute) before the current moment as a historical feature tuple set of the current moment;

[0109] calculate the early warning index:

[0110]

[0111] wherein, is the early warning index of the current moment, is the total number of feature tuples in the historical feature tuple set, is the failure risk index of the i-th feature tuple in the historical feature tuple set, is the failure risk index of the current moment, is the indicator function, the function value is 1 when the condition in the bracket is true, and the function value is 0 when the condition in the bracket is not true, the more feature tuples in the historical feature tuple set that satisfy the more likely the power distribution cabinet is to fail at the current moment, and vice versa.

[0112] In the first aspect, the technical solution uses the historical feature tuple set in the preset time window to construct a dynamic benchmark, which eliminates the limitations of fixed thresholds being easily disturbed by dynamic working conditions such as load and environment, making the failure determination more in line with the actual operating state of the power distribution cabinet. In the second aspect, the technical solution uses the indicator function to count the proportion of historical failure risk indexes that do not exceed the current failure risk index, quantifying the relative extreme degree of the current risk in the recent history. If this proportion increases, it means that the current risk is significantly higher than most moments in the recent history, and the likelihood of failure is high, and vice versa, which is normal fluctuation. This method based on relative comparison combined with statistical proportion can effectively filter out abnormal risk index spikes caused by transient noise, as very few data in the history will be lower than the current value under noise, which will not cause false positives, and can also sensitively capture the real failure trend of rising risk, balancing the robustness and sensitivity of failure detection. At the same time, setting a 1-minute time window balances real-time performance and data statistical reliability.

[0113] Since the essence is the proportion of historical feature tuples that satisfy , which is a dimensionless proportion value, the value range is [0, 1], where 0 means that the current risk is much lower than all historical moments, and 1 means that the current risk is higher than all historical moments.

[0114] For example, set the early warning index threshold to 0.25 for failure detection, which can filter out most normal fluctuations in healthy states, reducing the frequency of false alarms, and effectively capture abnormalities in the early stages of failure, reducing the frequency of missed alarms. If ​more than 0.25, it is determined that the power distribution cabinet has failed at the current time, and if less than 0.25, it is determined that the power distribution cabinet has not failed at the current time.

[0115] The application further provides a power distribution cabinet fault intelligent detection system, which comprises a memory and a processor, the memory stores a computer program, and the processor executes the computer program to realize the operations of steps S1-S5, thereby realizing accurate fault detection of the power distribution cabinet.

[0116] The above merely describes preferred embodiments of the application and is not intended to limit the application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the application shall be included in the protection scope of the application.

Claims

1. A power distribution cabinet fault intelligent detection method, characterized in that, The application relates to a method for detecting faults of a power distribution cabinet, comprising the following steps: Setting a to-be-measured point and a reference point in the power distribution cabinet, collecting load currents, temperatures of the to-be-measured point and the reference point, and an ambient temperature of the power distribution cabinet; At any moment of the operation of the power distribution cabinet, the moment is the current moment, the instantaneous thermal resistance ratio of the to-be-measured point is determined based on the load currents, the temperatures and the ambient temperature, comprising the following steps: the reference point is set to be composed of a plurality of physical reference points; for each physical reference point, the coefficient of variation of the thermal resistance of the physical reference point in a preset historical time window is determined, and the reciprocal of the coefficient of variation is determined as the stability index of the physical reference point; the stability index of each physical reference point is divided by the sum of the stability indexes of all the physical reference points to obtain the weight of each physical reference point; the thermal resistances of all the physical reference points at the current moment are weighted and averaged by using the weight of each physical reference point to obtain an equivalent reference thermal resistance; the ratio of the thermal resistance of the to-be-measured point to the equivalent reference thermal resistance is determined as the instantaneous thermal resistance ratio of the to-be-measured point; the thermal resistance of the to-be-measured point is the temperature rise of the to-be-measured point divided by the square of the load current of the to-be-measured point, and the thermal resistance of each physical reference point is the temperature rise of each physical reference point divided by the square of the load current of each physical reference point; the temperature rise refers to the temperature of the to-be-measured point or the physical reference point being higher than the ambient temperature; the load current of the to-be-measured point and the instantaneous thermal resistance ratio constitute a feature tuple at the current moment; A health sample set composed of a plurality of feature tuples of the to-be-measured point at different moments under the healthy state of the power distribution cabinet is collected in advance, and two-dimensional joint probability density estimation based on space-time adaptation is carried out based on the health sample set and the feature tuple at the current moment to obtain an actual joint probability density value of the feature tuple at the current moment; The theoretical instantaneous thermal resistance ratio corresponding to the load current of the to-be-measured point at the current moment is determined through the health sample set, and the joint probability density value of the feature tuple composed of the load current and the theoretical instantaneous thermal resistance ratio is taken as a theoretical joint probability density value of the feature tuple at the current moment; An abnormal score of the feature tuple at the current moment is determined through abnormal analysis, and a relative deviation of the instantaneous thermal resistance ratio of the to-be-measured point at the current moment from the theoretical instantaneous thermal resistance ratio is calculated, and the product of the abnormal score and the relative deviation is determined as a fault risk index of the current moment, and the power distribution cabinet is detected for faults according to the fault risk index.

2. The power distribution cabinet fault intelligent detection method according to claim 1, characterized in that, The theoretical joint probability density value of the feature tuple at the current moment and the actual joint probability density value of the feature tuple at the current moment are both obtained through two-dimensional joint probability density estimation based on space-time adaptation.

3. The power distribution cabinet fault intelligent detection method according to claim 1, characterized in that, The actual joint probability density value of the feature tuple at the current moment satisfies the following relationship: ； wherein, is the actual joint probability density value of the feature tuple at the current time instant is the actual joint probability density value of the feature tuple at the current time instant is the load current and the instantaneous thermal resistance ratio of the point under test at the current time instant, respectively, is the load current and the instantaneous thermal resistance ratio of the point under test at the current time instant, respectively, is the serial number and the total number of the feature tuples in the healthy sample set, respectively, is the serial number and the total number of the feature tuples in the healthy sample set, respectively, is the time weight of the th feature tuple in the healthy sample set, is the Gaussian kernel function, is the Gaussian kernel function, is the load current and the instantaneous thermal resistance ratio of the th feature tuple, respectively, is the adaptive bandwidth for acquiring the load current, is the adaptive bandwidth for acquiring the instantaneous thermal resistance ratio, is the adaptive bandwidth for , and is the adaptive bandwidth for .

4. The power distribution cabinet fault intelligent detection method according to claim 3, characterized in that, The time weight is determined based on the following method: The distance between the time of each feature tuple in the health sample set and the current moment is calculated, and the distance is taken as a negative index of the weight to construct a relationship formula in which the weight decays with the distance, so as to calculate the time weight of each feature tuple in the health sample set.

5. The power distribution cabinet fault intelligent detection method according to claim 3, characterized in that, The adaptive bandwidth of the load current and the adaptive bandwidth of the instantaneous thermal resistance ratio are respectively determined based on the following methods: For each load current, through The nearest neighbor matching algorithm finds the closest sample to the load current in the healthy sample set. Each load current value is calculated. The furthest distance between a load current value and that load current value is determined as the adaptive bandwidth of that load current. For each instantaneous thermal resistance ratio, a local range is set with the load current corresponding to the instantaneous thermal resistance ratio as the center, all feature tuples with the load current falling into the local range are screened out in the healthy sample set, and the standard deviation of the instantaneous thermal resistance ratio corresponding to all screened feature tuples is calculated, and the standard deviation is determined as the adaptive bandwidth of the instantaneous thermal resistance ratio.

6. The power distribution cabinet failure intelligent detection method according to claim 1, characterized in that, The theoretical instantaneous thermal resistance ratio corresponding to the load current of the to-be-tested point at the current moment is determined based on the following manner: A two-dimensional feature space is constituted with the load current as the horizontal axis and the instantaneous thermal resistance ratio as the vertical axis, each feature tuple of the healthy sample set corresponds to a feature point in the two-dimensional feature space, polynomial regression fitting is performed on all feature points to obtain a healthy center line, the instantaneous thermal resistance ratio corresponding to the load current of the to-be-tested point at the current moment on the healthy center line is obtained as the theoretical instantaneous thermal resistance ratio corresponding to the load current of the to-be-tested point at the current moment.

7. The method of claim 1, wherein, The abnormal score of the feature tuple at the current moment is determined based on the following manner: The number of feature tuples in the statistical health sample set whose load current is in the local range of the load current of the point under test at the current time instant is denoted as ; The abnormal score of the feature tuple at the current moment is calculated: wherein is the anomaly score of the feature tuple at the current time instant , and are the actual and theoretical joint probability density values of , is the natural exponential function, is the variance of the instantaneous thermal resistance ratio of all feature tuples whose load current is in the local range, is the max function, is a parameter to prevent the denominator from being zero.

8. The power distribution cabinet fault intelligent detection method according to claim 1, characterized in that, The fault detection of the power distribution cabinet according to the fault risk index is based on the following manner: all feature tuples in a preset time window before the current moment are obtained as the historical feature tuple set at the current moment; Calculate early warning indicators: In the formula, As the early warning indicator for the current moment, This represents the total number of feature tuples in the set of historical feature tuples. The first tuple in the set of historical feature tuples Fault risk index of each feature tuple This represents the fault risk index at the current moment. This is an indicator function; if the condition inside the parentheses is true, the function value is 1; if the condition inside the parentheses is false, the function value is 0. If the preset warning threshold is exceeded, the distribution cabinet is determined to have malfunctioned at that moment; if If the preset warning threshold is not exceeded, it is determined that the power distribution cabinet has not malfunctioned at the current moment.

9. A power distribution cabinet fault intelligent detection system, characterized in that, The power distribution cabinet fault intelligent detection system comprises a memory and a processor, the memory stores a computer program, and the processor executes the computer program to realize the steps of the power distribution cabinet fault intelligent detection method according to any one of claims 1-8.

Images