A method and system for performance testing of electronic and electrical appliances based on pattern recognition

Abstract

This invention relates to the field of power equipment performance testing technology, and discloses a method and system for testing the performance of electronic power appliances based on pattern recognition. The method includes: collecting multi-dimensional raw performance data of an uninterruptible power supply (UPS) system, performing time alignment and quality assessment; collecting operating condition parameters, constructing a mapping model between operating conditions and performance parameters, and dynamically calculating performance benchmarks; identifying performance degradation patterns of UPS system functional modules and calculating confidence levels; analyzing the coupling relationships between functional modules and identifying the root cause functional modules of performance degradation; and performing graded early warning and fault location. This invention supports the transformation from traditional periodic inspection and maintenance to a predictive maintenance mode based on actual performance status, improving the targeting and efficiency of equipment maintenance.

Description

Technical Field

[0001] This invention relates to the field of power equipment performance testing technology, and more specifically, to a method and system for testing the performance of electronic power appliances based on pattern recognition. Background Technology

[0002] With the rapid development of the digital age, data centers, as the core infrastructure for information processing and storage, face increasingly stringent requirements for the reliability and stability of their power supply systems. Uninterruptible power supplies (UPS) systems, as critical equipment for ensuring power supply to data centers, need to maintain efficient and stable performance under 24 / 7 continuous operation. However, existing UPS system performance monitoring technologies mainly rely on periodic inspections and alarm mechanisms based on fixed thresholds, a traditional maintenance model with numerous limitations.

[0003] Traditional performance testing methods typically rely on single-parameter monitoring and static threshold judgments, which are insufficient to address the complexity of multi-module collaborative operation in UPS systems. A UPS system comprises multiple functional units, including rectifier modules, inverter modules, and battery modules, with complex coupling relationships between these modules. Performance degradation in one module often triggers a chain reaction in other modules. Existing technologies lack in-depth analysis of these inter-module relationships, making it difficult to accurately identify the root cause of performance degradation, leading to inaccurate fault location and low maintenance efficiency. Furthermore, traditional methods ignore the dynamic impact of operating conditions on equipment performance, relying on fixed performance benchmarks for judgment. When operating conditions such as load rate and ambient temperature change frequently, misjudgments and omissions are prone to occur, affecting the accuracy and reliability of the testing.

[0004] Furthermore, existing performance testing technologies lack the ability to accurately identify different performance degradation modes. UPS systems exhibit diverse performance degradation modes during long-term operation, including gradual decline, periodic fluctuations, and accelerated degradation. The corresponding failure mechanisms and maintenance strategies differ across these modes. Traditional methods cannot distinguish these differentiated degradation modes, resulting in a lack of targeted maintenance strategies. This can lead to over-maintenance, increasing costs, or untimely maintenance, causing serious failures and hindering the shift from reactive to predictive maintenance. Summary of the Invention

[0005] This invention provides a method and system for testing the performance of electronic power appliances based on pattern recognition, which solves the technical problems of low accuracy in UPS system performance testing, inability to identify differentiated degradation patterns, lack of inter-module correlation analysis, and insufficient predictive maintenance capabilities in related technologies.

[0006] This invention provides a method for performance testing of electronic electrical appliances based on pattern recognition, comprising the following steps: Collect multi-dimensional raw performance data of uninterruptible power supply systems, perform time alignment and quality assessment, and obtain multi-dimensional performance characteristic data; Collect operating condition parameters, combine them with multi-dimensional performance characteristic data, construct a mapping relationship model between operating conditions and performance parameters, dynamically calculate performance benchmarks, and obtain operating condition adaptive performance benchmark data. Based on the adaptive performance benchmark data, the performance degradation modes of the uninterruptible power supply system functional modules are identified and the confidence level is calculated to obtain the performance degradation mode identifier and mode confidence level of each functional module. Based on the performance degradation mode identifiers and mode confidence levels of each functional module, the coupling and correlation relationships between functional modules are analyzed, and the root cause functional modules of performance degradation are identified, resulting in hierarchical fault tracing results. Based on the performance degradation mode identification of each functional module and the hierarchical fault tracing results, graded early warning and fault location are carried out to obtain early warning decision results.

[0007] In a preferred embodiment, the acquisition of multi-dimensional raw performance data of the uninterruptible power supply system includes: Based on the sensor network deployed at various monitoring points of the data center uninterruptible power supply system, multi-dimensional raw performance data reflecting the operating status of the equipment is collected in real time. The functional modules of the data center uninterruptible power supply system include rectifier modules, inverter modules, and battery pack modules, and collect electrical and thermal parameters of the rectifier modules, inverter modules, and battery pack modules. Vibration characteristic parameters and acoustic characteristic parameters of operating noise of uninterruptible power supply systems are collected to form a multi-dimensional raw performance data set.

[0008] In a preferred embodiment, the time alignment includes: Establish a unified time reference system and use a network time protocol server as the time reference source; At each data acquisition node, a high-precision timestamp marking mechanism is used to mark the timestamp information of each data point collected. Analyze the sampling frequency characteristics of various types of sensors in the sensor network, and identify the sensor type with the highest sampling frequency as the target sampling frequency; Sensor data with sampling frequencies lower than the target frequency are upsampled. A cubic spline interpolation method is used to extract the original sampling points of thermal characteristic parameters within a continuous time window, construct a cubic spline interpolation function, and generate interpolation points between the original sampling points according to the target sampling frequency to obtain the upsampled thermal characteristic parameter data sequence.

[0009] In a preferred embodiment, the quality assessment includes: Establish a data quality assessment indicator system, including three dimensions: completeness indicators, consistency indicators, and reasonableness indicators; Check the integrity of the data points corresponding to each timestamp, count the number and distribution of missing values, and calculate the integrity index as the ratio of the number of valid data points to the total number of data points. Check the consistency of the change range of performance parameter values ​​between adjacent timestamps, calculate the difference between performance parameter values ​​at adjacent times, and determine that there is an abnormal change when the absolute value of the difference exceeds the product of the upper limit of the normal change rate of the performance parameter value and the sampling time interval. Check the reasonableness of performance parameter values, count the number of data points that exceed the reasonable range, and calculate the reasonableness index as the ratio of the number of reasonable data points to the total number of data points.

[0010] In a preferred embodiment, the mapping model between the construction conditions and performance parameters includes: Establish an equipment similarity assessment mechanism to retrieve similar equipment from historical databases that are similar to newly commissioned uninterruptible power supply equipment; Equipment similarity assessment considers three dimensions: equipment model, power level, and application scenario. The comprehensive similarity is calculated as a weighted average of model similarity, power level similarity, and application scenario similarity. Retrieve all historical devices from the historical database whose overall similarity is greater than a preset similarity threshold, and filter the source devices for transfer learning; For the historical operating data of each source device, the distribution characteristics of performance parameters under different operating conditions are analyzed. The load rate and ambient temperature are divided into five intervals. For each combination of load rate interval and ambient temperature interval, all performance parameter data points under different operating condition combinations are extracted, and the statistical distribution characteristics under different operating condition combinations are calculated.

[0011] In a preferred embodiment, the dynamic computing performance benchmark includes: Based on the initial knowledge base of performance parameter distribution characteristics, a mapping relationship model between operating conditions and normal operating range of performance parameters is established. The multiple linear regression method is adopted, with load rate, ambient temperature, ambient humidity and input grid voltage fluctuation rate as independent variables, and the normal operating center value and normal operating boundary range of performance parameters as dependent variables. As the operating time of newly commissioned uninterruptible power supply equipment increases, the actual operating data of the current equipment is gradually accumulated. An incremental learning method is used to integrate the actual operating data of the current equipment into the mapping relationship model. The trigger condition for incremental learning is set to the current device's accumulated running data volume reaching a preset data volume threshold. When incremental learning is triggered, the actual running data of the current device within the most recent data volume threshold time period is extracted, the deviation between the actual statistical distribution characteristics of the current device and the statistical distribution characteristics predicted by the initial model is calculated, and the coefficients of the mapping relationship model are adjusted based on the deviation.

[0012] In a preferred embodiment, identifying the performance degradation mode of the uninterruptible power supply system functional module includes: Based on multi-dimensional performance characteristic data and operating condition adaptive performance benchmark data, the deviation of each dimension of performance parameters from the benchmark value is calculated, and the time-series evolution matrix of performance parameters is constructed. A multi-scale sliding window mechanism is used for temporal feature extraction, with three sliding windows at different time scales. For each performance parameter, at the current time, extract the deviation time series of each performance parameter within a short time window, calculate the statistical characteristics of the deviation series within the short time window, including short-time mean, short-time variance, short-time skewness and short-time kurtosis, as well as time series characteristics, including short-time first derivative, short-time second derivative and short-time autocorrelation coefficient; For both medium-term and long-term windows, the corresponding statistical and time-series characteristics are calculated using the same method.

[0013] In a preferred embodiment, the calculation of confidence level includes: Based on the comprehensive performance degradation feature vector of various performance parameters, a performance degradation mode classifier is established, which divides the performance state into five categories: normal state, gradual degradation mode, periodic fluctuation mode, accelerated degradation mode and mixed degradation mode, each corresponding to a cluster center. Five cluster centers are initialized. Based on the initialized cluster centers, the cluster centers and membership matrix are updated using an iterative optimization method. Calculate the Euclidean distance between the comprehensive performance degradation feature vector of each performance parameter and each cluster center, and calculate the membership degree of the comprehensive performance degradation feature vector of each performance parameter to each category based on the Euclidean distance; Extract the membership degree of each performance parameter to each category and compare them. Select the category with the largest membership degree as the performance degradation mode identifier, and the corresponding membership degree value as the mode confidence.

[0014] In a preferred embodiment, the coupling relationships between the analysis function modules include: For any two functional modules, select the analysis objects for each functional module and extract the deviation time series; The continuous deviation time series is discretized into a symbol sequence, and the deviation values ​​are divided. The time delay parameter is set to one sampling period, and the history length parameter is set to three sampling periods. The propagation entropy value is calculated based on the joint probability distribution of the symbol sequence statistics. Compare the entropy values ​​of the two directions. When the entropy value of one direction is greater than that of the other direction, and the difference between the two is greater than a preset difference threshold, it is determined that there is a unidirectional causal relationship, and a directed causal relationship edge is established.

[0015] In a preferred embodiment, a pattern recognition-based electronic electrical appliance performance testing system is used to perform the above-described pattern recognition-based electronic electrical appliance performance testing method, comprising: The data acquisition module is used to collect multi-dimensional raw performance data of the uninterruptible power supply system, perform time alignment and quality assessment, and obtain multi-dimensional performance characteristic data. The benchmark calculation module is used to collect operating condition parameters, combine them with multi-dimensional performance characteristic data, construct a mapping relationship model between operating conditions and performance parameters, dynamically calculate the performance benchmark, and obtain operating condition adaptive performance benchmark data. The pattern recognition module identifies the performance degradation patterns of the uninterruptible power supply system's functional modules and calculates the confidence level based on the adaptive performance benchmark data of the operating conditions, thereby obtaining the performance degradation pattern identifier and pattern confidence level of each functional module. The fault tracing module analyzes the coupling relationship between functional modules based on the performance degradation mode identifier and mode confidence of each functional module, and identifies the root cause functional module of performance degradation to obtain hierarchical fault tracing results. The early warning and decision-making module performs graded early warning and fault location based on the performance degradation mode identification and hierarchical fault tracing results of each functional module, and obtains early warning and decision-making results.

[0016] The beneficial effects of this invention are as follows: This invention effectively solves the problem of misjudgment that traditional fixed threshold judgment methods are prone to when operating conditions change by constructing a dynamic adjustment mechanism for performance benchmarks that adapts to operating conditions. It establishes a mapping relationship model between operating conditions and performance parameters based on historical data, and continuously optimizes the model parameters through incremental learning, so that the performance benchmark can be dynamically adjusted with real-time changes in operating conditions such as load rate and ambient temperature. The adaptive mechanism improves the accuracy of performance detection, avoids false alarms and false negatives caused by changes in operating conditions, and provides maintenance personnel with a more reliable basis for performance evaluation.

[0017] This invention achieves accurate identification and classification of differentiated performance degradation modes in UPS systems through multi-scale temporal feature extraction and fuzzy clustering analysis. It can accurately distinguish between different performance degradation modes such as gradual degradation, periodic fluctuations, and accelerated degradation. Furthermore, based on transfer entropy analysis, a multi-module coupled correlation network is established, enabling accurate location of the root cause functional modules of performance degradation. Through a hierarchical fault tracing mechanism, not only can the primary root cause functional modules be identified, but also the affected secondary root cause functional modules can be discovered, providing hierarchical fault location guidance for maintenance personnel. Combined with remaining time prediction and tiered early warning strategies, this invention can issue early warnings 3 to 7 days in advance and provide differentiated maintenance recommendations, effectively supporting the shift from traditional periodic inspection and maintenance to a predictive maintenance model based on actual performance status, improving the targeting and efficiency of equipment maintenance. Attached Figure Description

[0018] Figure 1 This is a flowchart of the main process of a pattern recognition-based performance testing method for electronic electrical appliances in this invention. Figure 2 This is a detailed flowchart of a method for testing the performance of electronic electrical appliances based on pattern recognition, as described in this invention. Figure 3 This is a block diagram of a pattern recognition-based performance testing system for electronic electrical appliances according to the present invention. Detailed Implementation

[0019] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0020] At least one embodiment of the present invention discloses a method for performance testing of electronic electrical appliances based on pattern recognition, such as... Figures 1 to 2 As shown, it includes the following steps: Step 1: Collect multi-dimensional raw performance data of the uninterruptible power supply system, perform time alignment and quality assessment, and obtain multi-dimensional performance characteristic data; Step 1.1: Collect raw performance data from multiple dimensions; Based on a sensor network deployed at various monitoring points within the UPS system, multi-dimensional raw performance data reflecting the equipment's operating status is collected in real time. For the rectifier module of the UPS system, electrical parameters such as input three-phase voltage amplitude, input three-phase current amplitude, input power factor, and input current harmonic distortion rate are collected, along with thermal characteristic parameters such as the surface temperature of the rectifier power switches, heat sink temperature, and module ambient temperature. For the inverter module, electrical parameters such as DC bus voltage, output three-phase voltage amplitude, output three-phase current amplitude, output voltage harmonic distortion rate, and output frequency deviation are collected, along with thermal characteristic parameters such as the surface temperature of the inverter power switches and the output transformer temperature. For the battery pack module, electrical parameters such as total battery pack voltage, charging and discharging current, individual cell voltage, and battery pack internal resistance are collected, along with thermal characteristic parameters such as battery pack surface temperature and battery cabinet internal temperature. For the entire UPS system, vibration characteristic parameters such as vibration amplitude and vibration frequency components of the equipment casing, as well as acoustic characteristic parameters such as the spectrum distribution of operating noise, are collected to form a multi-dimensional raw performance data set that includes electrical, thermal, vibration and acoustic dimensions.

[0021] Step 1.2: Establish a unified time reference and perform time synchronization processing on sensor data with different sampling frequencies; To address the issue of inconsistent sampling frequencies among different types of sensors during the acquisition of multi-dimensional raw performance data, a unified time reference system is established. A Network Time Protocol (NTP) server is used as the time reference source, providing a unified time reference for all data acquisition nodes. At each data acquisition node, a high-precision timestamp mechanism is employed to label each acquired data point with timestamp information accurate to the millisecond level. The sampling frequency characteristics of each type of sensor are analyzed, and the sensor type with the highest sampling frequency is identified as the target sampling frequency. For electrical parameter sensors, the sampling frequency is typically 100 times per second; for thermal characteristic parameter sensors, it is typically 1 time per second; for vibration parameter sensors, it is typically 50 times per second; and for acoustic parameter sensors, it is typically 10 times per second. After determining the target sampling frequency as 100 times per second, sensor data with sampling frequencies lower than the target frequency undergo upsampling processing.

[0022] For upsampling of thermal characteristic parameter data, a cubic spline interpolation method is employed. Original sampling points of thermal characteristic parameters within a continuous time window are extracted, and a cubic spline interpolation function is constructed. This function precisely matches the original data values ​​at the original sampling points and maintains continuity of the second derivative between adjacent sampling points. Based on the constructed cubic spline interpolation function, interpolation points are generated between the original sampling points according to the target sampling frequency, resulting in the upsampled thermal characteristic parameter data sequence. The same cubic spline interpolation method is used for upsampling acoustic parameter data. For vibration parameter data, since its sampling frequency is 50 times per second, close to half of the target frequency, a linear interpolation method is used to insert an interpolation point between adjacent sampling points, resulting in the upsampled vibration parameter data sequence.

[0023] For sensor data with a sampling frequency higher than the target frequency, although this situation does not exist in this embodiment, it may occur in other application scenarios. In such cases, a moving average filtering method is used for downsampling. The length of the sliding window is set to the ratio of the sampling frequency to the target frequency. The arithmetic mean of the data points within the sliding window is calculated as the downsampled data value. The sliding window moves according to the time interval of the target sampling frequency to obtain the downsampled data sequence. Through the above upsampling and downsampling processes, it is ensured that the performance parameter data of all dimensions are accurately aligned on the time axis. The data point corresponding to each timestamp contains the performance parameter values ​​of all dimensions, resulting in time-synchronized multi-dimensional performance data.

[0024] Step 1.3: Perform a quality assessment on the time-synchronized data and identify outlier data points; Based on the multi-dimensional performance data of time synchronization obtained in step 1.2, a comprehensive evaluation of data quality is conducted. A data quality evaluation index system is established, including three dimensions: integrity, consistency, and rationality. For the integrity index, it is checked whether the data point corresponding to each timestamp contains performance parameter values ​​for all dimensions, and the number and distribution of missing values ​​are counted. The integrity index value is calculated as the ratio of the number of valid data points to the total number of data points. When the integrity index value is lower than the set integrity threshold, the data for that time period is marked as having integrity issues. For the consistency index, it is checked whether the change in performance parameter values ​​between adjacent timestamps conforms to physical laws. The difference between performance parameter values ​​at adjacent times is calculated. When the absolute value of the difference exceeds the product of the upper limit of the normal change rate of the parameter and the sampling time interval, an abnormal mutation is determined. The number of abnormal mutation points is counted, and the consistency index value is calculated as the ratio of the number of normal change points to the total number of data points. For the rationality index, it is checked whether the performance parameter values ​​are within the physically permissible range.

[0025] For voltage parameters, the reasonable range is 0.7 to 1.3 times the rated voltage; for current parameters, the reasonable range is 0 to 1.5 times the rated current; for temperature parameters, the reasonable range is -20 degrees Celsius to +100 degrees Celsius; for power factor parameters, the reasonable range is 0 to 1; and for harmonic distortion rate parameters, the reasonable range is 0 to 1. Check whether the parameter value of each data point is within the corresponding reasonable range, count the number of data points exceeding the reasonable range, and calculate the reasonableness index as the ratio of the number of reasonable data points to the total number of data points. Combining the completeness index, consistency index, and reasonableness index, identify abnormal data points with missing values, abnormal mutations, or exceeding the reasonable range, and record the timestamp, parameter type, and anomaly type of the abnormal data points to obtain a set of abnormal data point identifiers.

[0026] Step 1.4: Adaptive interpolation repair method is used to process abnormal data points and reduce the weight of data that cannot be repaired; Based on the set of abnormal data point identifiers identified in step 1.3, abnormal data points are repaired. For abnormal data points with missing values, an adaptive interpolation repair method based on the data change trend of adjacent time points is adopted. Five normal data points before and after the abnormal data point are extracted, and the change trend of these normal data points is analyzed. The linear regression slope of the first five normal data points is calculated as a change trend indicator, and the linear regression slope of the last five normal data points is also calculated. When the difference between the first and last two slopes is less than the slope difference threshold, the data change trend is determined to be stable, and a linear interpolation method is used to perform linear interpolation based on the normal data values ​​of the time point before and after the abnormal data point to obtain the repaired data value. When the difference between the first and last two slopes is greater than the slope difference threshold, the data change trend is determined to have a turning point, and a cubic spline interpolation method is adopted. A cubic spline interpolation function is constructed based on the five normal data points before and after the abnormal data point, and the interpolation result is calculated at the timestamp position corresponding to the abnormal data point to obtain the repaired data value.

[0027] For anomalous data points exhibiting abnormal mutations, determine whether the mutation is a genuine change in device state. Check if other performance parameters also show mutations at the same time. If multiple parameters show mutations simultaneously, it is considered a genuine device state mutation, and the data point is retained without repair. If only a single parameter shows a mutation, it is considered a false mutation caused by sensor malfunction or interference, and median filtering is used for repair. Extract three data points before and after the anomalous data point, for a total of seven data points. Sort these seven data points by value, take the median as the repaired value, and replace the original anomalous data point value. For anomalous data points exceeding reasonable ranges, median filtering is also used for repair.

[0028] For data segments exhibiting anomalies across multiple consecutive time points, interpolation repair using adjacent normal data points is ineffective. To address this, a migration and repair method based on historical data from similar equipment is employed. Performance parameter data for other UPS devices of the same model and operating conditions within the same time period is retrieved from the historical database. A scaling factor is calculated between the normal data segments before and after the current UPS's anomaly segment and the corresponding time period data from the retrieved similar equipment. Based on this scaling factor, the data from the similar equipment during the anomaly period is scaled and adjusted to obtain a repaired data value sequence, which fills in the anomaly segment of the current equipment. For anomaly data points that cannot be effectively repaired using the above method, they are marked in the data record, and their weights are reduced during subsequent performance analysis. A weight decay coefficient of 0.3 is set, meaning that the weight of unrepairable anomaly data points in subsequent calculations is 0.3 times the weight of normal data points, resulting in multi-dimensional performance characteristic data.

[0029] In some embodiments, due to differences in sensor failure modes and data anomaly characteristics across different application scenarios, a single interpolation repair method may not be suitable for all scenarios. Therefore, an adaptive data repair method based on machine learning can be employed to improve the accuracy and adaptability of data repair. Specifically, a training sample set for the data repair model is established, containing manually labeled anomalous data points from historical operational data and their corresponding real data values. Features of the anomalous data points are extracted, including anomaly type, duration of anomaly, magnitude of anomaly, statistical characteristics of adjacent normal data points, and state characteristics of other dimensions and parameters at the same time. Based on the extracted features and corresponding real data values, the data repair model is trained. This model employs a random forest regression algorithm, capable of learning the mapping relationship between different anomaly features and the optimal repair value. In practical applications, when an anomalous data point is detected, its feature vector is extracted and input into the trained data repair model. The model outputs a predicted repair value, which is then used to replace the original anomalous data point, achieving adaptive data repair.

[0030] Step 2: Collect operating condition parameters, combine them with multi-dimensional performance characteristic data, construct a mapping relationship model between operating conditions and performance parameters, dynamically calculate the performance benchmark, and obtain operating condition adaptive performance benchmark data. Step 2.1: Collect the operating condition parameters of the UPS system; Based on the multi-dimensional performance characteristic data obtained in step 1, the operating condition parameters of the current UPS system are collected synchronously. These parameters include real-time load rate, ambient temperature, ambient humidity, and input grid voltage fluctuation rate. The real-time load rate is obtained by calculating the ratio of the current output power to the rated output power. Ambient temperature and humidity are collected by environmental sensors deployed around the UPS system. The input grid voltage fluctuation rate is obtained by calculating the ratio of the standard deviation to the average value of the input voltage over a continuous time window.

[0031] Step 2.2: Retrieve data on similar equipment from historical databases and extract the distribution characteristics of operating conditions and performance parameters as initial knowledge; A device similarity assessment mechanism is established to retrieve similar devices from the historical database that are comparable to newly commissioned UPS equipment. The device similarity assessment considers three dimensions: device model, power rating, and application scenario. For the device model dimension, the model similarity is 1 when the historical and current devices have identical models; 0.8 when they are different models but belong to the same manufacturer and series; and 0.5 when they are different models and from different manufacturers. For the power rating dimension, the ratio of the historical device's rated power to the current device's rated power is calculated. A power rating similarity of 1 is between 0.8 and 1.2; 0.7 is between 0.6 and 0.8 or 1.2 and 1.5; and 0.3 is between 0.6 and 1.5. For the application scenario dimension, the application scenario similarity is 1 when both the historical and current devices are used in data center scenarios; and 0.6 is when the historical device was used in an industrial scenario and the current device is used in a data center scenario.

[0032] The overall similarity score is calculated as a weighted average of model similarity, power rating similarity, and application scenario similarity, with weights of 0.5, 0.3, and 0.2, respectively. All historical devices with an overall similarity score greater than a similarity threshold of 0.7 are retrieved from the historical database. The retrieved historical devices are sorted from highest to lowest overall similarity score, and the top 5 devices with the highest overall similarity scores are selected as source devices for transfer learning. Historical operating data is extracted from these 5 source devices. This historical operating data includes at least 6 months of continuous operating records, with each record containing a timestamp, multi-dimensional performance parameter values, and corresponding operating condition parameter values.

[0033] For each source device, historical operating data was used to analyze the distribution characteristics of performance parameters under different operating conditions. The load rate was divided into five intervals: 0-20, 20-40, 40-60, 60-80, and 80-100. Ambient temperature was divided into five intervals: -20 to 0 degrees Celsius, 0 to 10 degrees Celsius, 10 to 20 degrees Celsius, 20 to 30 degrees Celsius, and 30 to 40 degrees Celsius. For each combination of load rate and ambient temperature intervals, all performance parameter data points were extracted for that operating condition combination. For each performance parameter, the statistical distribution characteristics under that operating condition combination were calculated, including the mean, standard deviation, 5th percentile, and 95th percentile. The mean represents the normal operating center value of the performance parameter under that operating condition, the standard deviation represents the degree of fluctuation of the performance parameter, and the 5th and 95th percentiles define the boundary range of the performance parameter's normal operation. The statistical distribution characteristics of five source devices under the same operating condition combination are fused using a weighted average method, with the weight being the comprehensive similarity between each source device and the current device. This yields initial knowledge of the performance parameter distribution characteristics under that operating condition combination. All operating condition combinations are then traversed to obtain an initial knowledge base of performance parameter distribution characteristics covering the entire operating condition space.

[0034] Step 2.3: Establish a mapping relationship model and dynamically adjust the model parameters through incremental learning; Based on the initial knowledge base of performance parameter distribution characteristics obtained in step 2.2, a mapping model between operating conditions and the normal operating range of performance parameters is established. This model employs a multiple linear regression method, using load rate, ambient temperature, ambient humidity, and input grid voltage fluctuation rate as independent variables, and the normal operating center value and normal operating boundary range of the performance parameters as dependent variables. For each performance parameter, based on the statistical distribution characteristics under different operating condition combinations in the initial knowledge base, a linear relationship between the normal operating center value and the operating condition parameter is fitted, expressed as the normal operating center value equals a constant term plus load rate multiplied by the load rate coefficient plus ambient temperature multiplied by the temperature coefficient plus ambient humidity multiplied by the humidity coefficient plus voltage fluctuation rate multiplied by the voltage fluctuation coefficient. Similarly, a linear relationship between the width of the normal operating boundary range and the operating condition parameter is fitted, with the boundary range width defined as the difference between the 95th percentile and the 5th percentile.

[0035] As the operating time of newly commissioned UPS equipment increases, actual operating data is gradually accumulated. An incremental learning method is employed to integrate this data into a mapping model, allowing the model to gradually adapt to the unique characteristics of each UPS unit. The trigger for incremental learning is set when the accumulated operating data reaches a threshold, initially set at one day's worth of data, gradually increasing to seven days' worth as operating time increases. When incremental learning is triggered, the actual operating data for the most recent threshold period is extracted. For each operating condition combination, the statistical distribution characteristics of the performance parameters of the actual operating data under that condition are calculated, including the actual mean, actual standard deviation, actual 5th percentile, and actual 95th percentile.

[0036] The deviation between the actual statistical distribution characteristics of the current device and the statistical distribution characteristics predicted by the initial model is calculated. The deviation includes center value deviation and boundary range deviation. The center value deviation is the difference between the actual mean and the predicted mean, and the boundary range deviation is the difference between the actual boundary range width and the predicted boundary range width. Based on the calculated deviations, the coefficients of the mapping relationship model are adjusted using the gradient descent method. For the mapping relationship of the normal operating center value, the gradient of the center value deviation with respect to each coefficient is calculated, and the coefficient values ​​are updated by multiplying the gradient by the learning rate, which is set to 0.01. For the mapping relationship of the normal operating boundary range, the gradient descent method is also used to update the coefficient values. Through multiple incremental learning iterations, the coefficients of the mapping relationship model are gradually adjusted from initial values ​​based on initial knowledge of similar devices to optimized values ​​adapted to the personalized characteristics of the current device, resulting in the updated mapping relationship model.

[0037] Step 2.4: Dynamically calculate the performance benchmark based on the updated mapping relationship model and real-time operating parameters; Based on the updated mapping relationship model obtained in step 2.3 and the currently collected operating condition parameters, the normal operating baseline value and allowable fluctuation range of each dimension of performance parameters under the current operating conditions are dynamically calculated.

[0038] By substituting the current real-time load rate, ambient temperature, ambient humidity, and input grid voltage fluctuation rate into the mapping model, the normal operating center value and normal operating boundary range width of each performance parameter are calculated. The normal operating baseline value is set as the normal operating center value, the upper limit of allowable fluctuation is set as the normal operating center value plus half of the normal operating boundary range width, and the lower limit of allowable fluctuation is set as the normal operating center value minus half of the normal operating boundary range width. For the rectifier input power factor parameter, when the load rate is 60%, the ambient temperature is 25 degrees Celsius, the ambient humidity is 50%, and the voltage fluctuation rate is 0.02, the calculated normal operating baseline value is 0.95, the upper limit of allowable fluctuation is 0.98, and the lower limit of allowable fluctuation is 0.92. For the inverter output voltage harmonic distortion rate parameter, under the same operating conditions, the calculated normal operating baseline value is 0.03, the upper limit of allowable fluctuation is 0.05, and the lower limit of allowable fluctuation is 0.01.

[0039] The above calculations are performed on all monitored performance parameters to obtain adaptive performance benchmark data, which includes the normal operating benchmark value, upper limit of allowable fluctuation, and lower limit of allowable fluctuation for each performance parameter. This performance benchmark data is dynamically updated according to real-time changes in operating conditions. When the load rate increases from 60% to 80%, the operating current of the rectifier and inverter increases, and the power loss increases, causing the normal operating benchmark values ​​of efficiency-related parameters to be adjusted accordingly, and the allowable fluctuation range to expand accordingly. When the ambient temperature rises from 25 degrees Celsius to 35 degrees Celsius, the heat dissipation conditions of the equipment deteriorate, the normal operating benchmark values ​​of temperature-related parameters will rise, and the upper limit of allowable fluctuation will also increase accordingly. By dynamically adjusting the adaptive performance benchmark, misjudgments caused by changes in operating conditions are avoided, the accuracy of performance testing is improved, and adaptive performance benchmark data is obtained.

[0040] In some embodiments, since linear mapping models have limitations in handling nonlinear relationships between operating parameters and performance parameters, a neural network-based nonlinear mapping model can be used to improve the accuracy of performance benchmark prediction. Specifically, a multi-layer feedforward neural network model is established. The input layer contains 4 neurons corresponding to 4 operating parameters, the hidden layer contains 2 layers (10 neurons in the first hidden layer and 5 neurons in the second hidden layer), and the output layer contains 2 neurons corresponding to the normal operating center value and the normal operating boundary range width, respectively. The modified linear unit ReLU is used as the activation function for the hidden layers, and a linear activation function is used for the output layer. The weight parameters of the neural network model are trained based on historical operating data of similar equipment and an initial knowledge base. During the current equipment operation, an incremental learning method is also used to fine-tune the neural network model using the actual operating data of the current equipment, allowing the model to gradually adapt to the personalized characteristics of the current equipment. In practical applications, real-time operating parameters are input into the trained neural network model, and the model outputs the predicted normal operating center value and the normal operating boundary range width. Performance benchmark data is calculated based on these output values.

[0041] Step 3: Based on the adaptive performance benchmark data, identify the performance degradation modes of the uninterruptible power supply system functional modules and calculate the confidence level to obtain the performance degradation mode identifier and mode confidence level of each functional module. Step 3.1: Construct the time-series evolution matrix of performance parameters; Based on the multi-dimensional performance feature data obtained in step 1 and the adaptive performance benchmark data obtained in step 2, the deviation of each performance parameter from the benchmark value is calculated. For each performance parameter, at each time point, the deviation is calculated as the difference between the actual measured value and the normal operating benchmark value, divided by half the width of the normal operating boundary range, to obtain a normalized deviation value. When the absolute value of the deviation is less than 1, it indicates that the performance parameter is within the normal operating range; when the absolute value of the deviation is greater than 1, it indicates that the performance parameter deviates from the normal operating range. A performance parameter time-series evolution matrix is ​​constructed. The rows of the performance parameter time-series evolution matrix correspond to different performance parameters, the columns correspond to consecutive timestamps, and the matrix elements are the deviation values ​​of the corresponding performance parameters at the corresponding times. The performance parameter time-series evolution matrix records the sequence of deviation changes of all performance parameters within a continuous time window, providing a data foundation for subsequent time-series feature extraction.

[0042] Step 3.2: Use a multi-scale sliding window mechanism to extract statistical and temporal features of performance parameters at different time scales; Based on the constructed performance parameter time-series evolution matrix, a multi-scale sliding window mechanism is employed for time-series feature extraction. Three sliding windows with different time scales are set: a short window of 1 hour, a medium window of 24 hours, and a long window of 7 days. The short window captures the instantaneous fluctuations and short-term anomalies of the performance parameters; the medium window captures daily periodic features and intraday variation patterns; and the long window extracts the long-term trend features and gradual changes of the performance parameters. For each performance parameter, at the current moment, the deviation time series of that parameter within the short window is extracted, containing all deviation data points from the past hour.

[0043] The statistical characteristics of the deviation sequence within a short-term window are calculated, including short-term mean, short-term variance, short-term skewness, and short-term kurtosis. The short-term mean reflects the average deviation of the parameter over the past hour; the short-term variance reflects the volatility of the deviation; the short-term skewness reflects the symmetry of the deviation distribution; and the short-term kurtosis reflects the sharpness of the deviation distribution. The temporal characteristics of the deviation sequence within a short-term window are also calculated, including the short-term first derivative, short-term second derivative, and short-term autocorrelation coefficient. The short-term first derivative is obtained by calculating the difference in deviation between adjacent time points and dividing by the time interval, reflecting the instantaneous rate of change of the deviation. The average of the first derivatives over all time points within the short-term window yields the short-term average rate of change. The short-term second derivative is obtained by calculating the difference in the first derivatives between adjacent time points and dividing by the time interval, reflecting the acceleration of the rate of change of the deviation. The average of the second derivatives over all time points within the short-term window yields the short-term average acceleration. The short-time autocorrelation coefficient is obtained by calculating the correlation coefficient between the deviation sequence and its sequence lagged by a certain time step, reflecting the periodic intensity of the deviation.

[0044] For the mid-term window, the deviation time series of this parameter over the past 24 hours is extracted. The same method is used to calculate mid-term statistical characteristics, including mid-term mean, mid-term variance, mid-term skewness, and mid-term kurtosis, as well as mid-term time-series characteristics, including mid-term average rate of change, mid-term average acceleration, and mid-term autocorrelation coefficient. For the long-term window, the deviation time series of this parameter over the past 7 days is extracted. Long-term statistical characteristics are calculated, including long-term mean, long-term variance, long-term skewness, and long-term kurtosis, as well as long-term time-series characteristics, including long-term average rate of change, long-term average acceleration, and long-term autocorrelation coefficient. The above feature extraction process is repeated for all monitored performance parameters to obtain a set of statistical and time-series characteristics for each performance parameter at the three time scales.

[0045] Step 3.3: The overall performance degradation feature vector is obtained by weighted fusion after calculating the feature contribution of each scale. Based on the multi-scale statistical features and time-series features extracted in step 3.2, feature fusion is performed to obtain a comprehensive performance degradation feature vector. For each performance parameter, the features extracted at three time scales include short-time mean, short-time variance, short-time skewness, short-time kurtosis, short-time average rate of change, short-time average acceleration, short-time autocorrelation coefficient, and corresponding medium-time and long-time features, totaling 21 feature dimensions. The contribution of each scale feature to the performance degradation judgment is analyzed; the contribution reflects the importance of the feature in distinguishing between normal and degraded states. The feature contribution is calculated using an information gain-based method, where information gain measures the degree to which the feature reduces the uncertainty of performance state classification.

[0046] Based on historical operational data, sample data labeled as normal and declining states are extracted, and the information gain value of each feature on these samples is calculated. The larger the information gain value, the higher the contribution of that feature to the performance status judgment.

[0047] For short-term features, since they are extracted based on a 1-hour time window, they can sensitively capture instantaneous fluctuations and short-term abnormal changes in performance parameters. Therefore, they make a high contribution to identifying sudden performance anomalies, with information gain values ​​typically between 0.15 and 0.25. However, due to the short time span, they make a low contribution to identifying gradual performance degradation that requires long-term observation, with information gain values ​​typically below 0.1.

[0048] For long-term features, since they are extracted based on a 7-day time window, they can effectively reflect the long-term trend of performance parameters and the gradual decline process. Therefore, they have the highest contribution to identifying gradual decline patterns, with information gain values ​​usually between 0.3 and 0.5. However, they are less sensitive to instantaneous fluctuations that require rapid response, with information gain values ​​usually below 0.08.

[0049] For the mid-time feature, since it is extracted based on a 24-hour time window, which corresponds exactly to the daily operating cycle of the equipment, it can effectively capture periodic fluctuation patterns related to load changes, daily ambient temperature, etc. Therefore, it makes a high contribution to the identification of periodic fluctuation patterns, and the information gain value is usually between 0.2 and 0.35.

[0050] Based on labeled normal and various decay state samples from historical operational data, the information gain value of each feature in distinguishing different states is calculated. The information gain calculation process involves statistically analyzing the state distribution of samples across different value ranges of the feature, calculating the information entropy of the original state distribution, and then calculating the weighted average information entropy of each subset after partitioning by the feature. The difference between these two values ​​is the information gain value of the feature. A larger information gain value indicates a higher contribution of the feature to state classification, and thus warrants a higher weighting coefficient during feature fusion.

[0051] Based on the calculated information gain values ​​of each feature, normalization is performed so that the sum of the contributions of all features is 1, thus obtaining the weight coefficient of each feature.

[0052] A weighted fusion method is used to combine the 21 feature dimensions according to their weight coefficients to obtain a comprehensive performance degradation feature vector. This feature vector still has 21 dimensions, but the values ​​of each dimension have been weighted, amplifying the values ​​of important features and reducing the values ​​of secondary features. For the rectifier input power factor parameter, the weight coefficient for its long-term average rate of change feature is 0.15, the weight coefficient for its short-term average rate of change feature is 0.05, and the weight coefficient for its long-term mean feature is 0.12. Each feature value of this parameter is multiplied by its corresponding weight coefficient to obtain the weighted feature values, which form the comprehensive performance degradation feature vector for this parameter. The above fusion process is repeated for all monitored performance parameters to obtain a set of comprehensive performance degradation feature vectors for each performance parameter.

[0053] Step 3.4: Identify performance degradation patterns and calculate pattern confidence based on fuzzy clustering method; Based on the comprehensive performance degradation feature vector obtained in step 3.3, a performance degradation mode classifier is established. This classifier uses fuzzy C-means clustering to divide performance states into five categories: normal state, gradual degradation mode, periodic fluctuation mode, accelerated degradation mode, and mixed degradation mode, each corresponding to a cluster center. Fuzzy C-means clustering allows each data point to belong to multiple categories with different membership degrees, and compared to hard clustering methods, it can better handle situations with ambiguous mode boundaries. The five cluster centers are initialized. In the feature vector of the normal state cluster centers, the mean at each scale is close to 0, and the rate of change and acceleration at each scale are close to 0, indicating that the performance parameters are stable near the baseline value. In the feature vector of the gradual degradation mode cluster centers, the long-term mean deviates from 0, the long-term average rate of change is negative or positive, indicating a continuous downward or upward trend, and the long-term average acceleration is close to 0, indicating uniform change.

[0054] In the eigenvectors of cluster centers exhibiting periodic fluctuation patterns, large medium- and short-term autocorrelation coefficients indicate a certain degree of periodicity, while the near-zero medium- and long-term means suggest the absence of a long-term trend. In the eigenvectors of cluster centers exhibiting accelerated deterioration patterns, a non-zero long-term average acceleration indicates an accelerating rate of change, with a large absolute value of the long-term average rate of change. In the eigenvectors of cluster centers exhibiting both long-term mean deviation and long-term average rate of change, along with a large autocorrelation coefficient, indicating the simultaneous presence of both trend and periodicity.

[0055] Based on the initialized cluster centers, an iterative optimization method is used to update the cluster centers and membership matrix. For each performance parameter's comprehensive performance degradation feature vector, its Euclidean distance to each cluster center is calculated. The membership degree of this feature vector to each category is calculated based on the distance; the membership degree is inversely proportional to the distance, with closer distances resulting in higher membership degrees. A fuzzification method with a fuzziness exponent of 2 is used to calculate the membership degree, ensuring that the sum of the membership degrees for all categories is 1. Based on the updated membership matrix, each cluster center is recalculated; the cluster center is the weighted average of all feature vectors' membership degrees to that category. This iterative process is repeated until the change in cluster centers is less than the convergence threshold or the maximum number of iterations is reached, yielding the final cluster centers and membership matrix.

[0056] For each performance parameter, its membership degree to each category is extracted. The category with the highest membership degree is used as the performance degradation mode identifier for that parameter, and the corresponding membership degree value is used as the mode confidence level. When the rectifier input power factor parameter has a membership degree of 0.75 for the progressive degradation mode, 0.15 for the normal state, and less than 0.1 for all other modes, it is determined that the parameter is in a progressive degradation mode, with a mode confidence level of 0.75. Each performance parameter is grouped according to its corresponding UPS system functional module (including rectifier module, inverter module, and battery module). Relevant parameters for the rectifier module include input power factor, input current harmonic distortion rate, and rectifier temperature. Relevant parameters for the inverter module include output voltage harmonic distortion rate, output frequency deviation, and inverter temperature. Relevant parameters for the battery module include battery internal resistance, individual cell voltage difference, and battery temperature. For each functional module, the performance degradation mode identifier of all relevant parameters of the module is counted. When most parameters in the module are in the same degradation mode, the module as a whole is determined to be in that degradation mode, and the performance degradation mode identifier and corresponding mode confidence of each functional module are obtained.

[0057] In some embodiments, since the fuzzy C-means clustering method requires pre-setting the number of clusters and is sensitive to the initial cluster centers, a deep learning-based autoencoder can be used for performance degradation pattern recognition to improve the adaptability and accuracy of pattern recognition. Specifically, an autoencoder neural network model is established, consisting of an encoder and a decoder. The encoder compresses the 21-dimensional comprehensive performance degradation feature vector into a 5-dimensional latent feature representation, and the decoder reconstructs the 5-dimensional latent features into a 21-dimensional feature vector. The autoencoder is trained based on historical normal operation data to minimize the reconstruction error. In practical applications, the comprehensive feature vector of the current performance parameters is input into the trained autoencoder, and the reconstruction error is calculated. When the reconstruction error is less than a normal threshold, the system is considered to be in a normal state. When the reconstruction error is greater than the normal threshold, performance degradation is identified, and the feature distribution in the latent feature space is further analyzed. Density clustering is then used to automatically identify the degradation pattern category without pre-setting the number of clusters, thus improving the flexibility of pattern recognition.

[0058] Step 4: Based on the performance degradation mode identifiers and mode confidence levels of each functional module, analyze the coupling relationships between functional modules, identify the root cause functional modules of performance degradation, and obtain hierarchical fault tracing results. Step 4.1: Calculate the propagation entropy of the time-series performance parameters between modules and identify causal relationships; Based on the performance degradation mode identifiers and time-series evolution matrices of performance parameters of each functional module identified in step 3, the correlation between different functional modules is analyzed. For any two functional modules A and B, representative performance parameters of each module are selected as the analysis objects. For the rectifier module, the input power factor is selected as the representative parameter; for the inverter module, the output voltage harmonic distortion rate is selected as the representative parameter; and for the battery pack module, the battery pack internal resistance is selected as the representative parameter. The deviation time-series sequence of the representative parameters of module A is extracted and denoted as sequence X, and the deviation time-series sequence of the representative parameters of module B is extracted and denoted as sequence Y. The transfer entropy value from sequence X to sequence Y is calculated. The transfer entropy quantifies the additional contribution of the historical information of sequence X to the prediction of the future state of sequence Y, given the historical information of sequence Y.

[0059] The calculation of transfer entropy is based on conditional mutual information theory, discretizing the continuous deviation time series into a symbol sequence. Using an equal-frequency discretization method, the deviation values ​​are divided into five intervals, corresponding to very low, low, medium, high, and very high symbols. Sequences X and Y are discretized separately to obtain symbol sequences X and Y. The time delay parameter is set to 1 sampling period, and the history length parameter is set to 3 sampling periods. The state of sequence Y at time t is calculated and denoted as Y current; the historical state of sequence Y three sampling periods before time t is calculated and denoted as Y history; the historical state of sequence X three sampling periods before time t is calculated and denoted as X history.

[0060] Based on the symbol sequence statistics, we calculate the joint probability distribution of the current state of Y, the history of Y, and the history of X, as well as the joint probability distribution of the current state of Y and the history of Y. The transfer entropy from X to Y is equal to the conditional entropy of the current state of Y given both the history of Y and the history of X, minus the conditional entropy of the current state of Y given only the history of Y. This value reflects the additional information contribution of the history of X to predicting the current state of Y. When the transfer entropy value is greater than 0, it indicates that the historical information of X is helpful in predicting the future state of Y, that is, X has a causal influence on Y. Using the same method, we calculate the transfer entropy value from sequence Y to sequence X, denoted as the transfer entropy from Y to X.

[0061] Compare the transfer entropy from X to Y with the transfer entropy from Y to X. When the transfer entropy from X to Y is greater than the transfer entropy from Y to X, and the difference is greater than a threshold, it is determined that the performance change of module A is the cause of the performance change of module B. A directed causal relationship is established from module A to module B, with the weight of the edge being the transfer entropy from X to Y. When the transfer entropy from Y to X is greater than the transfer entropy from X to Y, it is determined that the performance change of module B is the cause of the performance change of module A. A directed causal relationship is established from module B to module A. When the two transfer entropy values ​​are similar, and the difference is less than a threshold, it is determined that there is no unidirectional causal relationship between the two modules; there may be bidirectional mutual influence or no causal relationship. Perform the above transfer entropy calculation and causal relationship determination on all functional modules in the UPS system to obtain a set of causal relationships between modules.

[0062] Step 4.2: Construct a multi-module coupled network and calculate the node out-degree weights and the root cause identification module; Based on the set of causal relationships between modules obtained in step 4.1, a multi-module coupled network is constructed. This network is represented by a directed graph, where nodes correspond to the functional modules of the UPS system, including the rectifier module, inverter module, and battery module. Directed edges in the graph represent causal relationships between modules, with the direction of the edges pointing from the cause module to the result module. The weight of each edge is its corresponding transit entropy value, reflecting the strength of the causal influence. For the rectifier module node, all directed edges emanating from that node are counted, and the sum of their weights is calculated to obtain the out-degree weight sum of the rectifier module. The out-degree weight sum represents the overall influence strength of this module on other modules; the larger the out-degree weight sum, the more widespread the impact of the module's performance changes on other modules in the system.

[0063] For the inverter module node, its out-degree weight sum is calculated. For the battery pack module node, its out-degree weight sum is also calculated. For the bypass switch module node, its out-degree weight sum is calculated. The out-degree weight sums of all module nodes are compared, and the node with the largest out-degree weight sum is identified as the primary root cause functional module. The primary root cause functional module is the main source of performance degradation, and its performance changes have the greatest impact on other modules in the system. In this embodiment, assuming the out-degree weight sum of the rectifier module is 0.85, the out-degree weight sum of the inverter module is 0.32, the out-degree weight sum of the battery pack module is 0.15, and the out-degree weight sum of the bypass switch module is 0.08, then the rectifier module is identified as the primary root cause functional module.

[0064] Further identification of secondary root-cause functional modules is conducted. Secondary root-cause functional modules are those directly influenced by primary root-cause functional modules and possess a relatively large sum of their out-degree weights. In the coupled network, all target modules directly pointed to by primary root-cause functional modules are searched, i.e., directed edges exist from primary root-cause functional modules to these target modules. For these target modules, their sum of out-degree weights is compared, and the target module with the largest sum of out-degree weights is identified as a secondary root-cause functional module. In this embodiment, the rectifier module directly points to the inverter module and the battery pack module. The inverter module's sum of out-degree weights (0.32) is greater than the battery pack module's (0.15), therefore the inverter module is identified as a secondary root-cause functional module. Although secondary root-cause functional modules are influenced by primary root-cause functional modules, their own performance degradation can further affect other modules, forming a cascading effect.

[0065] For the identified primary root cause functional module, the rectifier module, combined with the performance degradation mode identifier obtained in step 3, the possible causes of failure within the module are analyzed. When the rectifier module is in a progressive degradation mode, and the representative parameter, the input power factor, shows a continuous downward trend, possible causes include aging of the power switching transistors leading to increased switching losses, deterioration of the heat dissipation system leading to increased operating temperature and thus affecting efficiency, and attenuation of the input filter capacitor capacity leading to a decrease in harmonic filtering effect. For the identified secondary root cause functional module, the inverter module, when it is in a periodic fluctuation mode, possible causes include drift in control circuit parameters leading to unstable output regulation, and changes in the performance of output filter components leading to periodic fluctuations in filtering characteristics. Through the above analysis, a hierarchical fault tracing result is obtained, including the identifiers of the primary and secondary root cause functional modules, and a list of possible fault causes for each root cause functional module.

[0066] Step 5: Based on the performance degradation mode identification and hierarchical fault tracing results of each functional module, perform hierarchical early warning and fault location to obtain early warning decision results; Step 5.1: Extract the current degradation status information of the key performance parameters of the root cause functional module; Based on the primary root cause functional module identifiers obtained in step 4 of the hierarchical fault tracing results, key performance parameters requiring focused prediction are identified. For the identified primary root cause functional module, the rectifier module, the input power factor is selected as the key performance parameter for degradation state analysis. Based on the performance parameter time-series evolution matrix constructed in step 3, the current deviation value of the rectifier input power factor parameter is extracted. This value reflects the degree of deviation of the performance parameter from the normal baseline value at the current moment.

[0067] The deviation time-series data of the performance parameter over the most recent 24 hours are extracted from its time-series evolution matrix. A linear regression method is used to fit the trend of deviation over time, and the slope of the regression line is calculated as the current decay rate. A positive decay rate indicates that the deviation of the performance parameter is increasing, while a negative value indicates that the deviation is decreasing. Further, the daily decay rate data for the parameter over the most recent 7 days are extracted, and the rate of change of the decay rate is calculated as the current decay acceleration. The decay acceleration reflects whether the decay process is accelerating or decelerating, providing an important parameter for subsequent evolution trajectory prediction. The extracted current deviation value, current decay rate, and current decay acceleration are used as the current decay state information for this key performance parameter.

[0068] Step 5.2: Establish a sub-model for predicting operating conditions and analyze the acceleration coefficient of influencing factors on the decay rate; Based on the historical operating data from the multi-dimensional performance characteristic data obtained in Step 1, the changing trend of operating conditions over a future period is analyzed. For the load rate parameter, time-series load rate data for the past 30 days is extracted to analyze its periodicity and trend changes. Data center load rates typically exhibit intra-weekly periodicity, with weekday load rates higher than weekend load rates, and show a slow upward long-term trend. A time-series decomposition method is used to decompose the load rate sequence into trend components, periodic components, and random components. Based on the linear extrapolation of the trend components, the load rate trend value for the next 7 days is predicted. Based on the periodicity of the periodic components, the periodic fluctuation of the load rate for the next 7 days is predicted. The predicted trend value and the periodic fluctuation are superimposed to obtain the predicted load rate values ​​for each time point in the next 7 days.

[0069] For the ambient temperature parameter, time-series data of the past 30 days were extracted to analyze its diurnal periodicity and seasonal trends. Ambient temperature exhibits a clear diurnal periodicity, with daytime temperatures higher than nighttime temperatures, and also shows a long-term variation trend influenced by seasonality. Using the same time-series decomposition method, the ambient temperature values ​​for each time point in the next 7 days are predicted. For the ambient humidity and input grid voltage fluctuation rate parameters, time-series decomposition and extrapolation methods are also used to predict their trends for the next 7 days. Combining the load factor prediction, ambient temperature prediction, ambient humidity prediction, and voltage fluctuation rate prediction yields a predicted operating condition sequence for the next 7 days.

[0070] This study analyzes the impact of different operating conditions and external environmental factors on the performance degradation rate, and establishes a relationship model between influencing factors and the degradation rate acceleration factor. Based on historical operating data from the multi-dimensional performance characteristic data obtained in step 1, statistical characteristics of the performance parameter degradation rate under different operating conditions are extracted. When the load rate increases from 60% to 80%, the operating current of the rectifier and inverter increases, power loss increases, temperature rise accelerates, and component aging speeds up, leading to an increased performance degradation rate. The average degradation rate at a load rate of 60% and the average degradation rate at a load rate of 80% are statistically analyzed, and the ratio of the two is calculated as the acceleration factor of the load rate effect. In this embodiment, when the load rate increases from 60% to 80%, the degradation rate acceleration factor is 1.3, indicating that the degradation rate increases by 30%.

[0071] When the ambient temperature rises from 25 degrees Celsius to 35 degrees Celsius, the equipment's heat dissipation deteriorates, the operating temperature increases, component aging accelerates, and the performance degradation rate increases. The average degradation rate at 25 degrees Celsius and the average degradation rate at 35 degrees Celsius are statistically analyzed, and an acceleration factor is calculated. In this embodiment, the degradation rate acceleration factor is 1.5 when the ambient temperature rises from 25 degrees Celsius to 35 degrees Celsius. The impact of ambient humidity and voltage fluctuation rate on the degradation rate is also analyzed, and the corresponding acceleration factors are calculated. An acceleration factor lookup table is established to record the degradation rate acceleration factors under different operating condition parameter values.

[0072] Step 5.3: Based on the current decline state and future operating conditions, predict and calculate the future evolution trajectory and remaining time of performance parameters; Based on the predicted operating condition sequence and acceleration coefficient lookup table obtained in step 5.2, and combined with the current degradation status information obtained in step 5.1, the future evolution trajectory of performance parameters is predicted. Based on the primary root cause functional module identifier in the hierarchical fault tracing results obtained in step 4, for the input power factor parameter of the rectifier module, its current deviation is -0.8, indicating that the current power factor is lower than the normal reference value; the current degradation rate is -0.05 per day, indicating that the deviation decreases by 0.05 per day; and the current degradation acceleration is -0.002 per day, indicating that the degradation rate is accelerating. A fault threshold is set at a deviation of -2.0, that is, when the power factor deviates from the normal reference value by twice the normal fluctuation range, it is determined to be a fault state requiring maintenance.

[0073] A daily recursive method is used to predict the deviation values ​​at each future time point. For the first day, the predicted load rate and predicted ambient temperature for the first day are extracted from the operating condition prediction sequence obtained in step 5.2. The corresponding decay rate acceleration coefficient is obtained by looking up the acceleration coefficient lookup table obtained in step 5.2. Assuming the predicted load rate for the first day is 70% and the predicted ambient temperature is 28 degrees Celsius, the corresponding comprehensive acceleration coefficient is 1.2. Multiplying the current decay rate by the comprehensive acceleration coefficient, the predicted decay rate for the first day is -0.05 per day multiplied by 1.2, which equals -0.06 per day. Adding the current deviation to the predicted decay rate for the first day multiplied by the time step of 1 day, the predicted deviation at the end of the first day is -0.8 plus -0.06, which equals -0.86. Simultaneously, the decay rate is updated by adding the current decay rate to the current decay acceleration multiplied by the time step, resulting in a predicted decay rate at the end of the first day of -0.06 per day plus -0.002 per day, which equals -0.062 per day.

[0074] For the next day, extract the predicted operating conditions for that day and find the corresponding acceleration factor, assuming it to be 1.15. Multiply the predicted degradation rate at the end of day 1 by the acceleration factor to obtain the predicted degradation rate for day 2 as -0.062 per day, multiplied by 1.15, which equals -0.071 per day. Add the predicted deviation at the end of day 1 to the predicted degradation rate for day 2 multiplied by the time step to obtain the predicted deviation at the end of day 2 as -0.86, plus -0.071, which equals -0.931. Repeat the above recursive process, calculating the predicted deviation and predicted degradation rate for each future time each day, until the predicted deviation reaches or exceeds the fault threshold of -2.0.

[0075] In this embodiment, through recursive calculation, the predicted deviation reaches -2.05 at the end of the 5th day, exceeding the fault threshold of -2.0. Therefore, the remaining time for the performance parameters to reach the fault threshold is calculated to be 5 days. Considering the uncertainties in the prediction process, including operating condition prediction errors, acceleration coefficient estimation errors, and random fluctuations in the degradation process, it is necessary to calculate the confidence interval of the remaining time. Using the Monte Carlo simulation method, random perturbations are introduced into the predicted operating condition values, acceleration coefficient, and degradation acceleration. The above recursive calculation process is repeated 1000 times to obtain 1000 remaining time samples. The distribution characteristics of these samples are statistically analyzed, and the 10th percentile is calculated as the lower confidence limit of the remaining time, and the 90th percentile is calculated as the upper confidence limit of the remaining time. In this embodiment, the lower confidence limit of the remaining time is 4 days, the upper confidence limit is 6 days, and the average is 5 days.

[0076] Step 5.4: Develop a tiered early warning strategy and generate differentiated maintenance recommendations based on the remaining time confidence interval; Based on the remaining time and its confidence interval calculated in step 5.3, a dynamic hierarchical early warning strategy is formulated. The early warning levels are divided based on the lower confidence limit of the remaining time. Using the lower confidence limit instead of the average value is a conservative estimate to avoid untimely maintenance due to overly optimistic predictions. When the lower confidence limit of the remaining time is greater than 7 days, the performance degradation is considered to be in the early stage, with a relatively small impact on system operation, and a low-level early warning is issued. A low-level early warning prompts maintenance personnel to pay attention to the performance trend of this module, but does not require immediate maintenance; it can be considered in the routine maintenance plan. When the lower confidence limit of the remaining time is between 3 and 7 days, the performance degradation is considered to have entered the mid-stage, requiring maintenance to be arranged in the near future, and an intermediate-level early warning is issued. An intermediate-level early warning requires maintenance personnel to schedule maintenance within the next 3 to 7 days to prevent further performance degradation from affecting power supply reliability.

[0077] When the lower confidence limit of the remaining time is less than 3 days, the performance degradation is determined to have entered the late stage, posing a high risk of failure, and a high-level warning is issued. The high-level warning requires maintenance personnel to immediately arrange emergency repairs, and if necessary, activate backup equipment or adjust load distribution to ensure system power supply safety. Based on the primary root cause functional module identifier in the hierarchical fault tracing results obtained in step 4, in this embodiment, the lower confidence limit of the remaining time for the rectifier module's input power factor parameter is 4 days, falling within the 3-7 day range; therefore, a medium-level warning is issued. The warning information includes the warning level, warning time, root cause functional module identifier, key performance parameter name, current deviation, predicted remaining time, and confidence interval.

[0078] Differentiated maintenance recommendations are generated for different warning levels and different performance degradation modes. Based on the primary root cause functional module identifiers in the hierarchical fault tracing results obtained in step 4 and the performance degradation mode identifiers of each functional module obtained in step 3, the maintenance recommendations for rectifier modules with intermediate warning levels and in a progressive degradation mode include the following: The recommended maintenance time window is within the next 3 to 5 days, prioritizing maintenance during periods of lower system load to minimize the impact on business operations.

[0079] Based on the list of possible fault causes for each root cause functional module in the hierarchical fault tracing results obtained in step 4, targeted checks are performed on the possible fault causes of the rectifier module, the primary root cause functional module. When the fault cause list indicates possible aging of the power switch transistor, the focus is on checking the on-state voltage drop and switching losses of the power switch transistor. When the fault cause list indicates possible performance degradation of the heat dissipation system, the focus is on checking the operating status of the cooling fan and the dust accumulation on the heat sink. When the fault cause list indicates possible capacitance decay of the input filter capacitor, the focus is on measuring the capacitance and equivalent series resistance of the filter capacitor. Modules and components requiring key checks include the rectifier power switch transistor, the heat dissipation system, and the input filter capacitor. For the power switch transistor, check whether its surface temperature rises abnormally, measure whether its on-state voltage drop increases, and replace the aging power switch transistor if necessary. For the heat dissipation system, check the operating status of the cooling fan, clean the dust accumulation on the heat sink surface, and ensure unobstructed heat dissipation channels. For the input filter capacitor, measure its capacitance and equivalent series resistance; when the capacitance decay exceeds 20 ohms or the equivalent series resistance increases by more than 50 ohms, replace the filter capacitor.

[0080] Simultaneously, considering the secondary root cause functional module identifiers in the hierarchical fault tracing results obtained in step 4, auxiliary maintenance recommendations are formulated for the identified secondary root cause functional modules, specifically the inverter modules. Although the inverter modules are affected by the primary root cause functional modules, their own performance degradation also needs attention. It is recommended that while maintaining the primary root cause functional modules, preventative checks be performed on the secondary root cause functional modules, including checking for control circuit parameter drift and the normal performance of output filter components, to avoid cascading faults.

[0081] The estimated maintenance time is 2 to 3 hours, including 1 hour of shutdown for inspection, 1 hour for component replacement, and 1 hour for recovery testing. Spare parts required during maintenance include 2 spares of power switching transistor type A and 3 spares of filter capacitor type B. After maintenance, performance verification testing is required to measure whether the rectifier input power factor has returned to the normal range, and 24 hours of monitoring operation is needed to confirm stable performance. For modules with advanced warnings and in accelerated degradation mode, maintenance is more urgent, requiring immediate repair and, if necessary, activation of the emergency plan to switch the load to the backup UPS system to avoid power outages due to faults. For modules with low-level warnings and in periodic fluctuation mode, maintenance recommendations focus on optimizing operating conditions, adjusting load distribution, and improving environmental conditions to slow down performance degradation and extend equipment lifespan.

[0082] Through the aforementioned dynamic hierarchical early warning strategy and differentiated maintenance recommendations, early warning decision results are obtained. These results also include complete information based on the hierarchical fault tracing results obtained in step 4, including the identifiers of primary and secondary root cause functional modules, as well as a list of possible fault causes for each root cause functional module. This information provides hierarchical guidance for fault location by operations and maintenance personnel, clarifying not only the primary root cause functional modules that need priority handling but also indicating potentially affected secondary root cause functional modules and specific directions of fault causes, thus achieving a complete closed loop from performance degradation detection to fault location and maintenance guidance. This result not only informs operations and maintenance personnel when maintenance is needed but also clearly indicates which components need to be inspected and replaced, how to perform maintenance operations, and how to verify the effects after maintenance, providing actionable guidance and supporting the shift from traditional periodic inspection and maintenance models to predictive maintenance models based on actual performance status.

[0083] A pattern recognition-based performance testing system for electronic electrical appliances, such as Figure 3 As shown, a method for performing the above-described pattern recognition-based performance testing of electronic electrical appliances includes: The data acquisition module is used to collect multi-dimensional raw performance data of the uninterruptible power supply system, perform time alignment and quality assessment, and obtain multi-dimensional performance characteristic data. The benchmark calculation module is used to collect operating condition parameters, combine them with multi-dimensional performance characteristic data, construct a mapping relationship model between operating conditions and performance parameters, dynamically calculate the performance benchmark, and obtain operating condition adaptive performance benchmark data. The pattern recognition module identifies the performance degradation patterns of the uninterruptible power supply system's functional modules and calculates the confidence level based on the adaptive performance benchmark data of the operating conditions, thereby obtaining the performance degradation pattern identifier and pattern confidence level of each functional module. The fault tracing module analyzes the coupling relationship between functional modules based on the performance degradation mode identifier and mode confidence of each functional module, and identifies the root cause functional module of performance degradation to obtain hierarchical fault tracing results. The early warning and decision-making module performs graded early warning and fault location based on the performance degradation mode identification and hierarchical fault tracing results of each functional module, and obtains early warning and decision-making results.

[0084] In one embodiment of the present invention, a specific example is provided: A large internet company's data center is equipped with 50 three-phase UPS units, each with a rated power of 500 kVA, to provide uninterrupted power to the data center's server clusters, storage arrays, and network equipment. The data center employs an N+1 redundancy configuration, with 45 UPS units online and 5 in hot backup mode during normal operation. The data center faces complex operating conditions, including 24 / 7 continuous operation, frequent dynamic changes in business load, and seasonal temperature fluctuations in the data center environment.

[0085] In a practical application, the method of this invention was used to perform performance testing on a device numbered UPS-15. During the data acquisition phase, multi-dimensional performance data was collected in real time through a sensor network deployed at various monitoring points of UPS-15. Table 1 shows some of the raw performance parameter data collected at a certain moment.

[0086] Table 1: Example of raw data for UPS-15 equipment performance parameters;

[0087] In the time synchronization preprocessing stage, a unified time reference was established to address the differences in sampling frequencies among different sensors. Cubic spline interpolation was used to upsample the thermal characteristic parameters, ensuring precise alignment of all dimensions of data on the time axis. In the data quality assessment stage, an abnormal abrupt change in the harmonic distortion rate of the rectifier input current was detected at a certain moment. Median filtering was used to repair this, resulting in reliable multi-dimensional performance characteristic data.

[0088] During the establishment of the adaptive performance benchmark, the operating condition parameters of the UPS-15 were simultaneously collected. The current load rate was 68%, the ambient temperature was 27 degrees Celsius, the ambient humidity was 52%, and the input grid voltage fluctuation rate was 0.018. Since the UPS-15 has been running for 3 years, sufficient historical data has been accumulated, and the mapping model has been fully adapted to the characteristics of the equipment through incremental learning. Based on the updated mapping model and the current operating condition parameters, the normal operating benchmark value of the rectifier input power factor was calculated to be 0.945, the upper limit of allowable fluctuation was 0.975, and the lower limit of allowable fluctuation was 0.915.

[0089] In the performance degradation pattern identification phase, the deviation of the rectifier input power factor was calculated as 0.91 minus 0.945, divided by 0.03, which equals -1.17, indicating that the current power factor is lower than the normal baseline value. A time-series evolution matrix of this parameter over the past 7 days was constructed, and features were extracted using a multi-scale sliding window. The average rate of change in the short-term window (1 hour) was -0.002 per hour, the average rate of change in the medium-term window (24 hours) was -0.0018 per hour, and the average rate of change in the long-term window (7 days) was -0.0021 per hour. The long-term average acceleration was -0.00003 per hour, indicating a slow acceleration in the degradation rate. Autocorrelation analysis showed no significant periodicity. Based on the comprehensive performance degradation feature vector, a fuzzy clustering classifier determined that the parameter was in a gradual degradation mode, with a mode confidence level of 0.78.

[0090] In the multi-module coupling correlation analysis phase, the transfer entropy between the rectifier module and the inverter module is calculated. The transfer entropy from the rectifier to the inverter is 0.42, and the transfer entropy from the inverter to the rectifier is 0.15. The difference between the two is 0.27, which is greater than the difference threshold of 0.2. Therefore, it is determined that the performance change of the rectifier module is the cause of the performance change of the inverter module. A coupling correlation network is constructed, and the sum of the out-degree weights of each module is calculated. The sum of the out-degree weights is 0.78 for the rectifier module, 0.31 for the inverter module, and 0.12 for the battery pack module. The rectifier module is identified as a first-level root cause functional module, and the inverter module is identified as a second-level root cause functional module.

[0091] During the degradation rate prediction phase, the current deviation of the rectifier input power factor is extracted to be -1.17, the current degradation rate is -0.05 per day, and the current degradation acceleration is -0.0007 per day. Based on the operating condition prediction sub-model, the load rate is predicted to fluctuate between 65% and 72% over the next 7 days, and the ambient temperature is predicted to fluctuate between 26% and 29 degrees Celsius. Referring to the acceleration factor table, the overall acceleration factor for the next 7 days is between 1.1 and 1.25. Using a daily recursive method, the remaining time for the performance parameters to reach the fault threshold of -2.0 is predicted to be 5 days, with a lower confidence limit of 4 days and an upper confidence limit of 6 days.

[0092] Based on the remaining time confidence limit of 4 days, falling within the 3-7 day range, a medium-level warning was issued. The generated maintenance recommendations included a suggested maintenance window of 3-5 days, prioritizing the period from 2 AM to 5 AM when load rates are low. Key areas to check included the rectifier power switches, cooling system, and input filter capacitors. The estimated maintenance time was 2-3 hours, and spare power switches and filter capacitors should be readily available. Following the warning recommendations, maintenance was scheduled for the early morning of the 4th day. Inspection revealed abnormally high surface temperatures and increased on-state voltage drops in the three power switches of the rectifier module. After replacing the aging switches, the rectifier input power factor recovered to 0.94, effectively curbing performance degradation and preventing downtime due to continued performance degradation.

[0093] Through the method of this invention, the UPS-15 equipment has achieved a transformation from passive maintenance to predictive maintenance. It issues an early warning 4 days before the performance degradation reaches the failure threshold, providing maintenance personnel with sufficient preparation time for maintenance, avoiding unplanned downtime, ensuring the reliability of power supply to the data center, and improving equipment operation and maintenance efficiency and system security.

[0094] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for electronic power appliance performance detection based on pattern recognition, characterized in that, The method comprises the following steps: Collecting multi-dimensional original performance data of the uninterruptible power supply system, performing time alignment and quality evaluation, and obtaining multi-dimensional performance characteristic data; Collecting working condition condition parameters, combining the multi-dimensional performance characteristic data, constructing a mapping relationship model of working conditions and performance parameters, dynamically calculating performance benchmarks, and obtaining working condition adaptive performance benchmark data; Based on the working condition adaptive performance benchmark data, the performance degradation mode of the functional module of the uninterruptible power supply system is identified and the confidence is calculated, and the performance degradation mode identification and mode confidence of each functional module are obtained; Based on the performance degradation mode identification and mode confidence of each functional module, the coupling relationship between the functional modules is analyzed, and the root functional module of performance degradation is identified, and a hierarchical fault tracing result is obtained; Based on the performance degradation mode identification and hierarchical fault tracing result of each functional module, a hierarchical warning and fault positioning is performed, and a warning decision result is obtained.

2. The method for detecting the performance of an electronic power appliance based on pattern recognition according to claim 1, characterized in that, The collection of multi-dimensional original performance data of the uninterruptible power supply system comprises: Based on the sensor network deployed at each monitoring point of the uninterruptible power supply system in the data center, multi-dimensional original performance data reflecting the running state of the equipment is collected in real time; The functional modules of the uninterruptible power supply system in the data center include rectifier modules, inverter modules and battery modules, and the electrical parameters and thermal characteristic parameters of the rectifier modules, inverter modules and battery modules are collected; The vibration characteristic parameters and the acoustic characteristic parameters of the running noise of the uninterruptible power supply system are collected to form a multi-dimensional original performance data set.

3. The method for detecting the performance of an electronic power appliance based on pattern recognition according to claim 1, characterized in that, The time alignment comprises: Establishing a unified time reference system, and using a network time protocol server as a time reference source; At each data collection node, a high-precision time stamp marking mechanism is used to mark time stamp information for each data point collected; Analyze the sampling frequency characteristics of each type of sensor in the sensor network, and identify the sensor type with the highest sampling frequency as the target sampling frequency; For the sensor data with a sampling frequency lower than the target frequency, perform upsampling processing, use a cubic spline interpolation method to extract the original sampling points of the thermal characteristic parameters within a continuous time window, construct a cubic spline interpolation function, and generate interpolation points between the original sampling points according to the target sampling frequency to obtain the upsampled thermal characteristic parameter data sequence.

4. The method for detecting the performance of an electronic power appliance based on pattern recognition according to claim 1, characterized in that, The quality evaluation comprises: Establishing a data quality evaluation index system, including three dimensions of integrity index, consistency index and rationality index; Check the integrity of each time stamp corresponding data point, count the number and distribution position of missing values, and calculate the integrity index value as the ratio of the number of valid data points to the total number of data points; Check the consistency of the change amplitude of the performance parameter value between adjacent time stamps, calculate the difference value of the performance parameter value at adjacent time, and determine that there is an abnormal mutation when the absolute value of the difference value exceeds the product of the upper limit of the normal change rate of the performance parameter value and the sampling time interval; Check the rationality of the performance parameter value, count the number of data points exceeding the reasonable range, and calculate the rationality index value as the ratio of the number of reasonable data points to the total number of data points.

5. The method for detecting the performance of an electronic power appliance based on pattern recognition according to claim 1, characterized in that, The construction of the mapping relationship model of working conditions and performance parameters comprises: A device similarity evaluation mechanism is established to retrieve similar devices from a historical database; The device similarity evaluation considers three dimensions of device model, power level and application scenario, and calculates a comprehensive similarity as a weighted average of model similarity, power level similarity and application scenario similarity; All historical devices with a comprehensive similarity greater than a preset similarity threshold are retrieved from the historical database to screen source devices for transfer learning; For the historical operation data of each source device, the distribution characteristics of performance parameters under different working conditions are analyzed, the load rate and environmental temperature are divided into five intervals respectively, and for each combination of load rate interval and environmental temperature interval, all performance parameter data points under different working condition combinations are extracted, and the statistical distribution characteristics under different working condition combinations are calculated.

6. The method for detecting the performance of an electronic power appliance based on pattern recognition according to claim 5, characterized in that, The dynamic calculation performance benchmark includes: Based on the performance parameter distribution characteristic initial knowledge base, a mapping relationship model between working condition and performance parameter normal operation range is established, and a multiple linear regression method is used to take the load rate, environmental temperature, environmental humidity and input grid voltage fluctuation rate as independent variables, and the normal operation center value and normal operation boundary range of the performance parameter as dependent variables; As the running time of the current newly commissioned uninterruptible power supply device increases, the actual operation data of the current device is gradually accumulated, and an incremental learning method is used to integrate the actual operation data of the current device into the mapping relationship model; The trigger condition for incremental learning is set as the cumulative running data volume of the current device reaching a preset data volume threshold, when incremental learning is triggered, the actual running data of the current device in the last data volume threshold period is extracted, the deviation between the actual statistical distribution characteristics of the current device and the statistical distribution characteristics predicted by the initial model is calculated, and the coefficients of the mapping relationship model are adjusted based on the deviation.

7. The method for detecting the performance of an electronic power appliance based on pattern recognition according to claim 1, characterized in that, The performance degradation mode of the uninterruptible power supply system functional module includes: Based on multi-dimensional performance characteristic data and working condition adaptive performance benchmark data, the deviation of each dimension performance parameter relative to the benchmark value is calculated, and a performance parameter time evolution matrix is constructed; A multi-scale sliding window mechanism is used for time sequence feature extraction, and three sliding windows of different time scales are set; For each performance parameter, at the current time, the deviation time series of each performance parameter in the short-time window is extracted, the statistical characteristics of the deviation sequence in the short-time window are calculated, including short-time mean, short-time variance, short-time skewness and short-time kurtosis, and time sequence characteristics, including short-time first derivative, short-time second derivative and short-time autocorrelation coefficient; For the medium-time window and the long-time window, the same method is used to calculate the corresponding statistical characteristics and time sequence characteristics.

8. The method for detecting the performance of an electronic power appliance based on pattern recognition according to claim 1, characterized in that, The confidence level calculation includes: Based on the comprehensive performance degradation feature vector of each performance parameter, a performance degradation mode classifier is established, and the performance state is divided into five categories, namely normal state, gradual degradation mode, periodic fluctuation mode, accelerated degradation mode and mixed degradation mode, corresponding to five cluster centers; The five cluster centers are initialized, and based on the initialized cluster centers, an iterative optimization method is used to update the cluster centers and membership matrix; Calculate the Euclidean distance between the comprehensive performance degradation feature vector of each performance parameter and the cluster center, and calculate the membership of each performance parameter to each category based on the Euclidean distance; Extract and compare the membership of each performance parameter to each category, select the category with the maximum membership as the performance degradation mode identifier, and the corresponding membership value as the mode confidence.

9. The method for detecting the performance of an electronic power appliance based on pattern recognition according to claim 1, characterized in that, The coupling relationship between the analysis function modules includes: For any two function modules, select the analysis object of each function module and extract the deviation degree time sequence; Discretize the continuous deviation degree time sequence into a symbol sequence, and divide the deviation degree value; Set the time delay parameter to one sampling period and the history length parameter to three sampling periods, calculate the transfer entropy value based on the symbol sequence statistical joint probability distribution; Compare the size relationship of the transfer entropy values in two directions, when the transfer entropy value in one direction is greater than that in the other direction, and the difference is greater than the preset difference threshold, it is determined that there is a one-way causal relationship, and a directed causal association edge is established.

10. A pattern recognition based electronic power appliance performance detection system, characterized in that, A mode recognition-based electronic power appliance performance detection method for performing any one of claims 1-9, comprising: A data acquisition module for acquiring multi-dimensional original performance data of the uninterruptible power supply system, performing time alignment and quality evaluation, and obtaining multi-dimensional performance feature data; A reference calculation module for acquiring operating condition parameters, combining the multi-dimensional performance feature data, constructing a mapping relationship model between operating conditions and performance parameters, dynamically calculating performance references, and obtaining operating condition adaptive performance reference data; A pattern recognition module for identifying performance degradation patterns of function modules of the uninterruptible power supply system and calculating confidence based on the operating condition adaptive performance reference data, and obtaining performance degradation pattern identifiers and mode confidence of each function module; A fault tracing module for analyzing the coupling relationship between function modules based on the performance degradation pattern identifiers and mode confidence of each function module, and identifying the root function module of performance degradation, and obtaining hierarchical fault tracing results; An early warning decision module for hierarchical early warning and fault positioning based on the performance degradation pattern identifiers and hierarchical fault tracing results of each function module, and obtaining early warning decision results.

Images