New energy equipment asset performance multi-dimensional dynamic comparison and root cause analysis method and system

Abstract

The present application relates to the technical field of industrial data processing and equipment monitoring, and discloses a new energy equipment asset performance multi-dimensional dynamic comparison and root cause analysis method and system, which comprises the following steps: obtaining target equipment operation data and mapping the data into a normalized feature vector; searching for the nearest neighbor equipment according to the feature distance to construct a homologous set, calculating the mean to generate a virtual standard benchmark; calculating the statistical dispersion of the homologous set to generate a sensitive factor mask; constructing a state drift vector of the target equipment and the benchmark, calculating the direction similarity and generating a trajectory correction coefficient; using the sensitive factor mask and the trajectory correction coefficient to perform weighted calculation on the residual vector, determining the abnormality according to the weighted result, and outputting the root cause. By introducing the space weighting based on group dispersion and the evolution trend weighting based on the drift vector, the present application effectively reduces the environmental noise interference, and realizes the accurate identification and root cause positioning of early weak faults.

Description

Technical Field

[0001] This invention relates to the field of industrial data processing and equipment monitoring technology, and in particular to a method and system for multidimensional dynamic comparison and root cause analysis of the performance of new energy equipment assets. Background Technology

[0002] With the large-scale deployment of new energy infrastructure, the efficiency of asset management and operation and maintenance for distributed equipment has become a key focus of the industry. Existing equipment monitoring and fault diagnosis technologies mainly rely on preset rules based on physical mechanism models or static threshold alarm mechanisms based on single parameters.

[0003] In real-world operating environments, the performance parameters of new energy equipment are significantly affected by the coupling of external environmental factors (such as temperature, humidity, and grid fluctuations) and internal operating conditions (such as load rate and continuous operating time). Traditional static threshold setting methods struggle to accommodate different environmental stresses and load conditions. When equipment operates under low load or low temperature conditions, even if its core components experience abnormal fluctuations due to performance degradation, the monitoring system often fails to trigger an alarm if the fluctuation amplitude does not reach the high threshold set for compatibility with high-temperature full-load conditions, leading to missed fault detection. Conversely, tightening the threshold to improve sensitivity can easily trigger numerous false alarms under extreme environments or high-load conditions, increasing the troubleshooting burden on maintenance personnel.

[0004] Furthermore, existing monitoring technologies typically focus on absolute numerical characteristics at a single time point, lacking dynamic consistency analysis of equipment state evolution trends. In physical systems, the trajectory of equipment state changes over time should follow specific physical laws and environmental response logic. When a group of devices under the same operating conditions exhibits a consistent response trend, if an individual device shows reverse evolution or deviates from the group's direction of change, it often indicates the occurrence of early failures. However, diagnostic logic based solely on static numerical comparisons cannot perceive such differences in vector direction, making it difficult to identify latent failures before absolute parameter values ​​exceed limits. Simultaneously, the lack of a dynamic cross-sectional comparison mechanism based on massive amounts of homogeneous samples makes it difficult for diagnostic systems to effectively isolate common-mode interference from environmental factors, and to accurately locate the root cause of performance degradation in a multi-dimensional feature space. Summary of the Invention

[0005] The technical problem solved by this invention is that existing new energy equipment has low fault diagnosis accuracy under complex working conditions, static thresholds cannot adapt to dynamic environments, and early evolution trends are difficult to identify.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] The first aspect of this invention provides a method for multidimensional dynamic comparison and root cause analysis of the performance of new energy equipment assets.

[0008] This method first performs standardization processing on multidimensional data to obtain the target device's operating data, mapping it to a normalized feature vector. To eliminate differences in physical dimensions and construct a unified feature space, the normalized feature vector includes environmental feature sub-vectors, operating condition feature sub-vectors, and performance feature sub-vectors. Specifically, the environmental feature sub-vectors represent the external boundary conditions of the device's operation, the operating condition feature sub-vectors represent the device's current load pressure, and the performance feature sub-vectors represent the response state of the device's internal core components.

[0009] In the benchmark construction phase, this method employs a dynamic benchmark generation strategy based on a homogeneous group. The system searches the entire network set for multiple neighboring devices with the closest current feature distance to the target device to construct a homogeneous set, and calculates the feature mean of the homogeneous set to generate a virtual standard benchmark. When constructing the homogeneous set, to ensure the comparability of the comparison samples and avoid erroneous matching between faulty devices, the system defines a feature weight vector, assigning higher weight values ​​to environmental feature sub-vectors and operating condition feature sub-vectors, and lower weight values ​​to performance feature sub-vectors. Based on this weighting strategy, a weighted Euclidean distance is calculated, and devices with the smallest distance and satisfying similarity constraints are selected to form the homogeneous set.

[0010] To address the false positives and false negatives inherent in single-dimensional evaluation, this method introduces a dual-weighted correction mechanism based on both spatial and temporal dimensions.

[0011] In the spatial dimension, the statistical dispersion of the homogeneous set is calculated, and a sensitivity factor mask is generated based on this dispersion. Specifically, the root mean square deviation of all devices in the homogeneous set relative to a virtual standard benchmark in each feature dimension is calculated as the feature dispersion. An inverse mapping relationship is established between dispersion and sensitivity factor; that is, the smaller the feature dispersion, the higher the consistency of the parameter in the group, and the larger the assigned sensitivity factor value; conversely, the sensitivity is reduced. After normalizing the sensitivity factors of all feature dimensions, a sensitivity factor mask is generated to suppress interference from high background noise parameters in subsequent calculations.

[0012] In the time dimension, the system constructs the state drift vectors of the target device and the virtual benchmark, calculates the directional similarity between them, and generates a trajectory correction coefficient. The system calculates the difference between the eigenvectors of the target device and the virtual benchmark at the current time and the previous time, respectively, to obtain the state drift vector of the target device and the drift vector of the group benchmark. The consistency of the evolution direction is quantified by calculating the cosine similarity between the two vectors. When the evolution direction of the target device significantly deviates from the group benchmark (i.e., low similarity), the generated trajectory correction coefficient increases, thus amplifying the anomaly score; when in a steady state or with the same direction of change, the trajectory correction coefficient remains at the benchmark level.

[0013] In the final diagnostic phase, a weighted calculation of the residual vector between the target device and the virtual standard benchmark is performed using a sensitivity factor mask and trajectory correction coefficients. First, the absolute deviation between the normalized feature vector and the virtual standard benchmark is calculated as the basic residual vector. Then, the element-wise Hadamard product operation is performed on the basic residual vector using the sensitivity factor mask. Finally, the overall gain of the calculation result is adjusted using scalar trajectory correction coefficients to generate a fused anomaly score vector.

[0014] The system determines abnormal states and outputs root cause parameters based on weighted results. A global anomaly index is obtained by calculating the Euclidean norm of the fused anomaly score vector, and a dynamic alarm threshold is generated using the Laida criterion based on the statistical distribution characteristics (such as mean and standard deviation) of the source set. When the global anomaly index exceeds this dynamic threshold, the equipment is deemed abnormal. Simultaneously, the element values ​​of each dimension in the fused anomaly score vector are sorted in descending order, and the features with the highest ranking are selected as the primary causes of the fault. Their contribution rates are calculated and output, thereby achieving precise location of the root cause of the fault.

[0015] The second aspect of this invention provides a system for multidimensional dynamic comparison and root cause analysis of the performance of new energy equipment assets.

[0016] The system includes a feature mapping module, a benchmark construction module, a weight generation module, and an attribution diagnosis module.

[0017] The feature mapping module is configured to acquire the target device's operating data and map it into a normalized feature vector, thereby achieving the standardization and vectorization of the original data.

[0018] The benchmark construction module is configured to search for the neighboring device with the closest feature distance to the normalized feature vector in the entire network device set to construct a homogeneous set, and calculate the feature mean of the homogeneous set to generate a virtual standard benchmark, thereby establishing a dynamic reference system under the current operating conditions.

[0019] The weight generation module is configured to perform multi-dimensional weight calculations. On one hand, it calculates the statistical dispersion of the homogeneous set and generates a sensitivity factor mask based on the statistical dispersion to quantify the signal-to-noise ratio of each parameter in the current environment. On the other hand, it constructs the state drift vectors of the target device and the virtual standard benchmark, calculates the directional similarity between the two, and generates trajectory correction coefficients to quantify the degree of anomalousness in the device's evolution trend.

[0020] The attribution diagnosis module is configured to perform weighted residual calculation and fault location. This module uses a sensitivity factor mask and trajectory correction coefficient to perform weighted calculation on the residual vector between the target device and the virtual standard baseline. Based on the weighted result, it determines the abnormal state and outputs the root cause parameters that lead to the performance decline.

[0021] In summary, the present invention has at least one of the following beneficial technical effects:

[0022] 1. This invention constructs a normalized feature vector that includes environmental, operating condition, and performance dimensions, and uses a weighted distance search for the homogeneous set that focuses on input conditions (environment and operating conditions). This effectively shields the interference of equipment performance abnormalities on the selection of the benchmark, ensuring the objectivity and accuracy of the virtual standard benchmark.

[0023] 2. This invention introduces a sensitivity factor mask based on statistical dispersion, enabling adaptive adjustment of the diagnostic logic. For parameters that fluctuate significantly due to environmental influences, the weight is automatically reduced, while the sensitivity of steady-state parameters is increased. This effectively reduces the false alarm rate caused by environmental noise and enhances the ability to detect minute deviations.

[0024] 3. This invention introduces a trajectory correction coefficient based on state drift vectors, expanding the diagnostic dimension from static numerical values ​​to dynamic evolution trends. By comparing the evolution direction of individual devices and group benchmarks in multidimensional space, early potential faults (such as counter-trend heating) can be identified even if the numerical values ​​have not exceeded limits but the trend of change violates physical laws, significantly improving the predictive maintenance capability of equipment.

[0025] 4. This invention employs a dual-weighted residual calculation based on sensitive factors and trajectory correction, combined with dynamic alarm thresholds and contribution rate ranking, to achieve quantitative assessment of faults and automatic root cause identification, providing maintenance personnel with intuitive and accurate maintenance guidance. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the architecture of a multi-dimensional dynamic comparison and root cause analysis system for the performance of new energy equipment assets according to an embodiment of the present invention;

[0027] Figure 2 This is a flowchart illustrating the multidimensional dynamic comparison and root cause analysis method for the performance of new energy equipment assets according to an embodiment of the present invention.

[0028] Figure 3 This is a schematic diagram of the multidimensional state data vectorization mapping and preprocessing process according to an embodiment of the present invention;

[0029] Figure 4 This is a schematic diagram of the process for constructing a dynamic homogeneous reference based on weighted Euclidean distance according to an embodiment of the present invention;

[0030] Figure 5 This is a schematic diagram of the dynamic mask generation process for sensitive factors based on statistical dispersion according to an embodiment of the present invention;

[0031] Figure 6 This is a schematic diagram of a time-series trajectory correction process based on state drift vector according to an embodiment of the present invention;

[0032] Figure 7 This is a schematic diagram of the weighted differential attribution and root cause locking process for multi-factor fusion according to an embodiment of the present invention;

[0033] Figure 8 This is a schematic diagram comparing the evolution trajectories of the target vector direction and the population direction in the initial stage of a fault, according to an embodiment of the present invention.

[0034] Figure 9 This is a schematic diagram comparing the abnormal index changes with the dynamic threshold triggering timing according to an embodiment of the present invention;

[0035] Figure 10 This is a schematic diagram of the root cause contribution rate of the fault according to an embodiment of the present invention.

[0036] Among them, 110 is the data preprocessing module; 120 is the feature mapping module; 130 is the benchmark construction module; 140 is the weight generation module; and 150 is the attribution diagnosis module. Detailed Implementation

[0037] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0038] See attached document Figure 1 This invention provides a multi-dimensional dynamic comparison and root cause analysis system for the performance of new energy equipment assets, which runs on a data processing center or cloud server platform.

[0039] This system connects multiple new energy devices located in different geographical locations via a communication network. The new energy devices include DC charging piles. The communication network uses 4G, 5G, or industrial Ethernet communication protocols. The system is configured to acquire real-time telemetry data from the new energy devices and process the data to output fault diagnosis results.

[0040] like Figure 1 As shown, the system includes a data preprocessing module 110, a feature mapping module 120, a benchmark construction module 130, a weight generation module 140, and an attribution diagnosis module 150.

[0041] The data preprocessing module 110 is configured to receive raw operating data from new energy equipment. The raw operating data includes environmental parameters, operating condition parameters, and internal performance parameters. The data preprocessing module 110 performs data cleaning operations on the raw operating data, removing missing values ​​and invalid data with incorrect formats, and aligning the data according to a preset time window.

[0042] The feature mapping module 120 is configured to map the cleaned operational data into feature vectors in a high-dimensional feature space. Each feature vector consists of an environmental sub-vector, a working condition sub-vector, and a performance sub-vector. The feature mapping module 120 performs normalization calculations on the data in each dimension to eliminate the influence of different physical dimensions and generate a unified input vector.

[0043] The benchmark construction module 130 is configured to calculate the weighted Euclidean distance between the target device and other devices in the entire network based on feature vectors. The benchmark construction module 130 then selects devices with similar features to form a source set based on the weighted Euclidean distance. The benchmark construction module 130 calculates the mean feature value of all devices within the source set to synthesize a virtual standard benchmark for this operating condition.

[0044] The weight generation module 140 is configured to analyze the statistical characteristics and evolution patterns within the same source set. The weight generation module 140 calculates the standard deviation distribution of the same source set across each feature dimension and generates a sensitivity factor mask based on the standard deviation distribution. Simultaneously, the weight generation module 140 calculates the state drift vector of the device as its state changes over time and generates trajectory correction coefficients based on the angle between the target device's drift vector and the average drift vector of the same source set.

[0045] The attribution diagnosis module 150 is configured to perform the final fault location calculation. The attribution diagnosis module 150 calculates the residual vector between the target device feature vector and the virtual standard baseline. The attribution diagnosis module 150 uses a sensitivity factor mask and trajectory correction coefficients to weight the residual vector, generating a fault attribution score, and outputs a list of parameters that cause performance degradation based on the score ranking.

[0046] See attached document Figure 2 This invention provides a method for multidimensional dynamic comparison and root cause analysis of the performance of new energy equipment assets, comprising the following steps:

[0047] S100: Obtain equipment operation data within a preset time window and construct a normalized feature vector that includes environmental, operating condition and performance dimensions.

[0048] S200 searches for multiple neighboring devices in the entire network device set that are closest to the target device's current characteristics, constructs a homogeneous set, and calculates the centroid of the homogeneous set as a virtual standard benchmark.

[0049] S300 calculates the statistical dispersion of the homologous set on each feature dimension, and generates a sensitive factor mask that reflects the importance of each parameter under the current working condition based on the dispersion.

[0050] S400: Construct the state drift vector of the target device and the homogeneous set within a continuous time window, calculate the directional similarity between the two, and generate a trajectory correction coefficient that reflects the degree of anomalousness in the evolution trend.

[0051] The S500, combining a sensitivity factor mask and trajectory correction coefficients, performs weighted calculations on the difference results between the target device and the virtual standard benchmark, quantifies the abnormal contribution of each parameter, and identifies the root cause.

[0052] The following provides a detailed description of each step of the method of the present invention.

[0053] See attached document Figure 3 In step S100, the system performs multidimensional state data acquisition and vectorization mapping, transforming the physical world's device operating state into high-dimensional feature space data that can be processed by a computer. In this embodiment, the process specifically includes the following steps:

[0054] S110, construct the original state vector.

[0055] Assume the total number of devices in the entire network is For any device ( At that moment Define its original state vector as This vector consists of feature dimensions from three subspaces, namely, environmental feature subvectors. Operating condition feature sub-vectors and performance feature subvectors .

[0056] In this embodiment, the logic behind the division of the above three sub-vectors lies in decoupling external disturbances from internal states. Wherein:

[0057] Environmental feature sub-vectors The external boundary conditions used to characterize equipment operation are primarily influenced by the natural environment rather than equipment malfunctions. These conditions include ambient temperature, ambient humidity, and the voltage deviation rate of the local power grid. Ambient temperature and humidity data are obtained through built-in sensors or associated meteorological data interfaces, while the power grid voltage deviation rate is obtained through power grid-side monitoring instruments.

[0058] Operating condition feature vector The current load pressure of the device is used to characterize the device and is a basic variable for subsequent judgment of whether the device performance meets expectations. Specifically, it includes the current output power, the current requested by the vehicle battery management system (BMS), the initial state of charge (SOC), and the charging duration.

[0059] Performance feature subvectors It includes response variables that reflect the operating status of the core components inside the equipment and is the main dimension carrying the fault characteristics. Specifically, it includes the module air inlet temperature, the charging gun head temperature, the DC bus voltage ripple coefficient calculated and uploaded by the embedded controller at the device end at high frequency, and the power conversion efficiency calculated based on the input and output power.

[0060] S120 performs data cleaning and alignment.

[0061] The system receives the original state vector Then, the validity is first checked; for missing values, linear interpolation is used to fill them with data from the previous and next times; for outliers that exceed physical limits (e.g., temperature readings of negative absolute zero), they are directly removed and the data at that time is marked as invalid.

[0062] Simultaneously, the system maps data from different sampling frequencies onto a preset time axis. Specifically, it sets the sampling period. (e.g., 60 seconds) All asynchronously uploaded data will be aligned to the specified range using a resampling algorithm. , , At discrete time points, ensure that subsequent calculations are performed within the same time slice.

[0063] For the specific implementation of linear interpolation and resampling algorithms, those skilled in the art can refer to conventional data processing techniques, and will not elaborate further here.

[0064] S130, perform normalization mapping.

[0065] Due to the original state vector The physical dimensions and numerical ranges of the various feature dimensions differ significantly. For example, DC bus voltage is typically between 500V and 1000V, while the ripple factor is usually less than 1.0. If the raw data is used directly to calculate the Euclidean distance, large numerical features will dominate the distance calculation result, causing changes in small numerical features to be masked. Therefore, the system constructs a unified normalized feature vector. :

[0066] ;

[0067] in, This represents the total dimension of the features. For each feature dimension kk in the matrix, a Min-Max normalization method with a regularization term is used to map it to the interval [0,1]. The normalization calculation formula is as follows:

[0068] ;

[0069] in, For equipment At any moment No. The original observations of each feature dimension; For this feature dimension The preset statistical lower limit value is taken from the first percentile value of the parameter in the historical operation database of this type of equipment or the physical lower limit defined in the equipment technical specifications. For this feature dimension The preset statistical upper limit value is taken from the 99th percentile value of this parameter in the historical operation database of this type of equipment or the physical upper limit defined in the equipment technical specifications. For a preset small amount of regularization (e.g., 1×10), -6 ), used to prevent when The calculation error is caused by the denominator being zero. Furthermore, to ensure the numerical stability of the eigenvectors, for values ​​exceeding... For abnormally large or small values ​​within the range, the system truncates the normalized result to the interval [0,1].

[0070] This mapping transforms all features into dimensionless numerical values, making the parameters of different physical properties comparable for distance calculations within the same Euclidean space. The resulting feature vectors... This serves as the foundational input data for subsequent homology clustering and fault attribution.

[0071] See attached document Figure 4 In step S200, the system searches for multiple neighboring devices in the entire network device set that are closest to the target device's current characteristics, constructs a homogeneous set, and calculates the centroid of the homogeneous set as a virtual standard benchmark. The core principle of this step is to utilize the group distribution characteristics of massive devices to replace the traditional physical mechanism model. Since devices should theoretically exhibit similar performance responses (performance characteristics) under the same environmental stress (environmental characteristics) and workload (operating condition characteristics), the theoretical output standard that the device should possess at the current moment can be established by finding group samples with similar input conditions. In this embodiment, the process specifically includes the following sub-steps:

[0072] S210, Construct the feature weight vector. To ensure that the selected neighboring devices have a high degree of consistency with the target device in terms of operating conditions, the system defines a feature weight vector. ,in is the total dimension of the features. In this embodiment, the weight allocation strategy follows the principle of strong correlation between input conditions and weak correlation between output responses.

[0073] Specifically, feature dimensions are divided into two categories:

[0074] The first category is state-determined features, which include environmental feature subvectors. and working condition feature vectors These features are boundary conditions for equipment operation and are assigned higher weight values ​​(e.g.) ∈[0.8,1.0]);

[0075] The second category is state-response features, which include performance feature subvectors. These features are the device's feedback to boundary conditions and are assigned lower weight values ​​(e.g.) (∈[0,0.2]) or zero weight.

[0076] The technical effect of this strategy is that it forces the system to prioritize matching devices with similar external environments (such as temperature and voltage) and workloads (such as power and SOC) when searching in a multi-dimensional space, while ignoring the differences in the current performance of the devices. This avoids the problem of missing detection caused by faulty devices being neighbors, which may result in the target device being incorrectly matched with equally abnormal devices as a benchmark due to its own abnormal performance.

[0077] S220, calculate the weighted Euclidean distance.

[0078] Based on the aforementioned weight vector, the system calculates the target device. With any other device on the entire network ( At the current moment Weighted Euclidean distance This distance is used to quantify the similarity between devices in their operating states. The calculation formula is as follows:

[0079] ;

[0080] in, For target equipment At any moment No. Normalized eigenvalues ​​of dimension; For candidate devices At any moment No. Normalized eigenvalues ​​of dimension; For the first The weight coefficients corresponding to the dimensional features. By introducing weight coefficients... This formula enables non-uniform measurement of different feature dimensions, ensuring that the distance calculation results mainly reflect the differences in environment and working conditions.

[0081] S230, Filter sets of similar origins.

[0082] The system calculates the weighted Euclidean distance. All devices on the network are sorted in ascending order. In this embodiment, a dynamic K-nearest neighbor strategy is used to filter the set of devices from the same source. In practice, the nearest neighbor is selected. Each device is considered a source neighbor, among which For a preset sample size threshold (e.g.) =20).

[0083] To ensure the validity of the samples, the system also introduces a maximum distance constraint, i.e., only when... Less than the preset similarity threshold At that time, equipment Only then will they be included in the set of common origins.

[0084] In this embodiment, The value of is related to the sum of the elements of the feature weight vector, and the calculation formula is: ,in The allowable average normalization bias for a single dimension (e.g., taking...) =0.05 indicates that an average deviation of 5% is allowed.

[0085] If the number of devices that meet the conditions is insufficient If the number of devices that actually meet the conditions is zero, then the confidence level of the homologous samples at that moment is marked as insufficient. The system will automatically switch to the historical baseline of that device or skip the diagnosis at the current moment to avoid false alarms caused by forced comparison.

[0086] S240, Synthetic Virtual Standard Reference.

[0087] In determining the set of homologous origins Then, the system calculates the arithmetic mean of the feature vectors of all devices in the set, generating a virtual standard reference vector. This baseline vector represents the average performance level of a health device under a specific combination of environmental and operating conditions. The calculation formula is as follows:

[0088] ;

[0089] in, Indicates the number of devices in the same source set; For devices in the collection The normalized feature vectors generated. It is a vector with the same dimensions as the target device feature vector, which includes the group mean of three dimensions: environment, operating conditions, and performance.

[0090] This step utilizes group statistical laws to eliminate random noise from individual devices, constructing a digital mirror image of the ideal device at the current moment, which serves as a virtual standard benchmark. This will serve as the origin of the reference coordinates for calculating residuals and evaluating anomalies in subsequent steps.

[0091] See attached document Figure 5 In step S300, the system utilizes the group statistical characteristics within the same source set to drive the diagnostic logic of the individual device, generating dynamic sensitivity factor masks for different feature dimensions. The core principle of this step lies in introducing an adaptive gain control mechanism: treating the dispersion of the group samples in a specific dimension as the background noise level under that operating condition.

[0092] When background noise is high, the system automatically reduces the signal-to-noise ratio requirement, decreasing sensitivity to deviations; conversely, when background noise is extremely low, the system improves its ability to detect minute deviations. In this embodiment, the process specifically includes the following steps:

[0093] S310, calculate the feature dispersion of the homogeneous set.

[0094] The system is based on the homogeneous set constructed in step S200. This involves calculating the statistical dispersion of all devices within the set across each feature dimension to quantify the group volatility of that feature under the current environment and operating conditions. A feature dispersion vector is defined. , of which Discreteness of dimensional features The calculation formula is as follows:

[0095] ;

[0096] in, The number of samples in the homologous set; For devices in the collection In the Normalized eigenvalues ​​of dimension; The virtual standard reference vector calculated in step S240 is in the first... The numerical value of the dimension.

[0097] This formula calculates the root mean square deviation (RMSD) of homologous samples relative to the centroid. This value directly reflects the stability of the physical parameters: if A value approaching 0 indicates that the parameter has strong convergence under the current operating conditions and exhibits steady-state characteristics; if A large value indicates that the parameter is significantly affected by environmental disturbances and exhibits non-steady-state characteristics.

[0098] S320, construct the sensitivity factor mapping function.

[0099] To achieve a diagnostic logic that assigns high weights to features with high consistency and low weights to features with high volatility, the system establishes an inverse proportional mapping relationship between dispersion and sensitivity factors. In this embodiment, an inverse proportional function with a regularization term is used to calculate the original sensitivity factor. For feature dimensions... Its original sensitivity The calculation logic is as follows:

[0100] ;

[0101] in, The preset sensitivity benchmark coefficient is typically set to a value range of [0.5, 2.0] (preferably 1.0 in this embodiment) to adjust the order of magnitude of the overall weight. The preset smoothing regularization constant is typically set to a value range of

[10] . -4 10 -2 (In this embodiment, 1×10 is preferred) -3 ). Its function includes two aspects: first, to prevent the group from being completely uniform, i.e. The first is that the denominator being zero leads to computational overflow; the second is to set a physical upper limit for the sensitivity factor to prevent system instability caused by overfitting.

[0102] S330 generates a normalized sensitivity mask.

[0103] To facilitate the unified calculation of subsequent weighted residuals, the system normalizes the calculated original sensitivity factors to generate the final sensitivity factor mask vector. This embodiment employs a maximum value normalization strategy, and the calculation formula is as follows:

[0104] ;

[0105] in, For a set containing the original sensitivities of all dimensions, This is the maximum value in the set. After this processing, the mask vector... The range of values ​​for elements in the array is limited to the interval (0,1).

[0106] This sensitive factor mask Essentially, it's a dynamic filter that participates in subsequent calculations as a diagonal element of the weighting coefficient matrix. Its technical effect is that when certain parameters (such as battery temperature) exhibit significant differences across a group under specific operating conditions (such as high-temperature fast charging), the system automatically reduces the weight of that parameter in anomaly scoring, thereby filtering out common fluctuations caused by the environment. Conversely, for certain critical parameters (such as voltage ripple), if they typically remain highly stable within the group, the system will amplify this difference through highly sensitive weighting once the target device shows a slight deviation, thus achieving accurate detection of early, subtle faults. This dynamic mask generation mechanism based on statistical characteristics solves the technical challenge of traditional fixed-threshold methods failing to balance high sensitivity and low false alarm rates.

[0107] See attached document Figure 6In step S400, the system introduces dynamic evolution analysis in the time dimension. By calculating the consistency between the trajectory of equipment state changes over time and the group baseline evolution trajectory, a trajectory correction coefficient is generated. The core principle of this step lies in dynamic evolution consistency verification: In complex industrial systems, the absolute value at a single moment may be affected by various factors, but the trend of physical quantity changes (i.e., the direction of response to environmental stimuli) should follow physical laws. For example, under reduced load conditions, the equipment temperature should theoretically remain stable or decrease. If the temperature of a certain piece of equipment shows an upward trend at this time, even if its absolute temperature has not yet reached the alarm threshold, its evolution logic has violated physical laws. This step aims to identify latent faults by capturing this trend paradox. In this embodiment, the process specifically includes the following steps:

[0108] S410, construct the state drift vector.

[0109] The system retrieves the target device. At the present moment Normalized eigenvectors Compared with the previous sampling time Normalized eigenvectors Simultaneously retrieve the virtual standard reference vector at the corresponding time. and Based on this, the state drift vector of the target device is calculated respectively. Drift vector relative to the group baseline The calculation formula is as follows:

[0110] ;

[0111] ;

[0112] The aforementioned drift vector, in a physical sense, represents the differential change in the device state within a multidimensional feature space. Among these, Characterizes the target device at the sampling interval The actual state evolution direction and rate within, It characterizes the theoretical state evolution direction and rate of the health equipment group under the current environmental and operating condition changes.

[0113] S420 calculates the cosine similarity of the trajectories.

[0114] To quantify the consistency between the evolution direction of the target device and the evolution direction of the population theory, the system calculates the cosine similarity between the two drift vectors. This indicator, by decoupling the direction and magnitude of the vector, focuses on assessing the nature of the state change rather than its magnitude, thus effectively determining whether the device has experienced a reverse change. The calculation formula is as follows:

[0115] ;

[0116] in, Represents the dot product operation of vectors; The Euclidean norm of a vector; To prevent extremely small positive numbers with a denominator of zero (in this embodiment, the value is 1 × 10⁻⁶), -8 ). The value range is [-1, 1].

[0117] when When this indicates that the changing trend of the target equipment is highly positively correlated with the group (e.g., synchronous heating); when When this occurs, it indicates that the target device exhibits a negatively correlated evolution that is diametrically opposed to that of the population (e.g., the population cools down while the device heats up); when When the two evolution directions are orthogonal, there is an uncorrelated perturbation.

[0118] S430 generates trajectory correction coefficients.

[0119] To avoid misjudgments of trends caused by random noise dominating the drift direction during steady-state operation, the system first executes steady-state filtering logic. The system calculates the target device's state drift vector. The modulus (i.e., the Euclidean norm) is determined, and this modulus is compared with a preset effective change threshold. (For example, set to 0.01) for comparison.

[0120] like If the device is currently in a steady state or a silent period, its minor directional differences are ignored, and the trajectory correction coefficient is directly set. =1.0;

[0121] like If the device is in a valid dynamic evolution state, the system determines that it is based on the calculated cosine similarity. Trajectory correction coefficients are generated using a nonlinear mapping function. .

[0122] This coefficient is used to dynamically adjust the gain of the static residuals in subsequent calculations. In this embodiment, to balance computational efficiency and the smoothness of the adjustment, piecewise linear penalty logic is preferred. The calculation formula is as follows:

[0123] ;

[0124] in, The preset trend penalty gain factor is usually set to a value range of [1.0, 5.0] (preferably 2.0 in this embodiment). This parameter determines the amplification factor of the final anomaly score when an anomaly is detected. The preset consistency judgment threshold is usually set to [0, 0.8] (preferably 0.5 in this embodiment).

[0125] The physical meaning of this formula is: Set an allowable trend deviation cone angle (determined by...). (Definition) When the evolution direction of the target device is within this cone angle (i.e.) When this happens, the trend is considered normal. Keep it at 1.0, without introducing additional penalty; when the evolution direction exceeds this cone angle (i.e. When similarity Further reduction, Linear increase, with weighted penalties for abnormal trends.

[0126] Through this step, the system extends from static numerical analysis to dynamic trend analysis. For localized temperature rise anomalies, such as those caused by poor contact, the absolute temperature may still be within a safe threshold in the early stages of the fault. However, because its rate of temperature change (in the direction of temperature rise) differs significantly from that of the group under the same low-load conditions (in the direction of constant or cooling), the system generates a large trajectory correction coefficient. This allows for the early identification of potential evolutionary anomalies before the values ​​exceed the limit.

[0127] See attached document Figure 7 In step S500, the system performs comprehensive calculations and outputs results, fusing the virtual standard benchmark, sensitive factor mask, and trajectory correction coefficient generated in the preceding steps in a multi-dimensional manner. The core logic of this step lies in constructing a multi-dimensional weighted evaluation system. First, the signal-to-noise ratio is adjusted for the residuals of different physical parameters using the spatial dimension sensitive factor mask; second, the trend gain is adjusted for the overall anomaly amplitude using the time dimension trajectory correction coefficient. Finally, the processed multi-dimensional feature vector is mapped to a single health index, achieving a quantitative assessment of the equipment status. In this embodiment, the process specifically includes the following steps:

[0128] S510, calculate the basic residual vector.

[0129] The system calculates the normalized feature vector of the target device at the current time. With the virtual standard reference vector generated in step S240 The absolute deviation between them. This deviation represents the Euclidean distance projection of the target device relative to the centroid of the common group before weighted processing. Define the basic residual vector. The calculation formula is as follows:

[0130] ;

[0131] in, This indicates that the absolute value operation is performed on each element of the vector; For feature dimension The same non-negative vector. At this point, the residual only reflects the difference at the numerical level, and does not yet include statistical dispersion information in the feature dimension or evolutionary trend information in the time dimension.

[0132] S520, calculate the fusion anomaly score vector.

[0133] The system utilizes the sensitive factor mask generated in step S330 and the trajectory correction coefficients generated in step S430 The basic residual vector is then weighted and corrected. This process employs a hierarchical weighting strategy: the inner layer uses Hadamard product to perform element-wise weighting of the feature dimensions to suppress the influence of high background noise parameters; the outer layer uses scalar multiplication to introduce a trend gain in the time dimension to amplify the influence of the reverse evolution trend. The final fusion anomaly score vector is defined. The calculation formula is as follows:

[0134] ;

[0135] in, This represents the Hadamard product operation, which is the element-wise multiplication of vectors. This is a sensitivity factor mask vector, where smaller element values ​​indicate higher tolerance for that dimension. This is a scalar trajectory correction coefficient with a value greater than or equal to 1.0. This formula ensures that the final score reflects both the degree of numerical deviation and consideration of evolutionary logic.

[0136] S530 calculates the global anomaly index and determines alarms.

[0137] To assess the overall health of the equipment, the system will integrate anomaly scoring vectors. Aggregates into a single scalar indicator, namely the global anomaly index. This embodiment preferably uses the Euclidean norm (L2 norm) for aggregation to highlight the contribution of significant anomalous features to the overall score. The calculation formula is as follows:

[0138] ;

[0139] in, For vectors The Middle The element value of the dimension. The system will With dynamic alarm threshold A comparison is made. To adapt to the group fluctuation levels under different operating conditions, this embodiment abandons the fixed threshold and adopts a dynamic threshold generation mechanism based on the Laida criterion (3σ criterion). Specifically, the system calculates the homogeneous set in real time. The arithmetic mean of the anomaly indices of all devices at the current moment. and standard deviation The alarm thresholds are set as follows:

[0140] ;

[0141] in, This is the confidence level coefficient, which is usually set to 3 (corresponding to a 99.7% confidence interval).

[0142] like If the condition is met, the target device is determined to be abnormal, triggering the root cause locking logic; otherwise, the device is determined to be operating normally.

[0143] S540, Fault Root Cause Ranking and Top-K Output.

[0144] When an anomaly alarm is triggered, the system merges the anomaly scoring vector. All elements are sorted in descending order. The larger the element value, the greater the contribution of that feature dimension to the overall anomaly after sensitivity weighting and trend correction, thus it is determined to be the main cause of the failure.

[0145] The system extracts the first sorted items. One feature (in this embodiment) Preferably, option 3) is output as the primary cause of the failure. To quantify the contribution of each root cause, the system further calculates the... Contribution rate of dimensional features :

[0146] ;

[0147] Ultimately, the system output includes the anomaly time, global anomaly index, and previous... A diagnostic report is generated based on the root cause characteristic name and its contribution rate. This step automates fault attribution, allowing operations personnel to analyze the contribution rate... Prioritize checking high-contribution physical components based on size to improve maintenance efficiency.

[0148] Example 1:

[0149] To more intuitively illustrate the fault identification capability of this invention under dynamic operating conditions, a specific case of "abnormal temperature rise caused by aging of the charging gun head contact" is selected below for explanation. This embodiment demonstrates how the system can identify potential problems in advance through multi-dimensional dynamic comparison before the numerical value triggers a traditional static threshold alarm.

[0150] The specific steps are as follows:

[0151] S100, Scene Background and State Vectorization:

[0152] DC charging pile #Dev_042 was selected as the monitoring target, and the monitoring period was from 14:00:00 (time t−1) to 14:01:00 (time t). During this time, vehicles in the charging station were generally transitioning from constant current fast charging to constant voltage trickle charging, and the output current showed a decreasing trend. At time t−1, the target equipment was under high load (current 200A), and the gun head temperature was 55℃. The normalized feature vector... Approximately [0.8, 0.55]. At time t, the output current drops to 100A (load halved), but due to the increased contact impedance of the nozzle, heat dissipation is slow, and the temperature only slightly drops to 54℃. At this point, the normalized vector... It becomes [0.4, 0.54].

[0153] S200~S400, Dynamic Benchmark Construction and Trend Analysis:

[0154] Based on environmental and operating condition characteristics, the system identifies 20 devices from the same source that have undergone the same load reduction process in real time. Calculations show that the average temperature of this group significantly decreased from 54℃ to 44℃, generating a virtual baseline vector. The drift vector is [0.4, 0.44]. In the evolutionary consistency analysis, the system calculates the drift vector of the target device as ΔX = [-0.4, -0.01] (temperature is almost static), while the population baseline drift vector is [0.4, 0.44]. =[−0.4,−0.10] (Temperature decreases synchronously).

[0155] Determined by steady-state filtering logic Then, the system calculates the cosine similarity of their evolution directions. ≈0.3, much smaller than the preset threshold. =0.5 indicates that the target device exhibits an abnormal evolutionary trend of lagging behind the group. According to the formula... (set up =2.0), the system automatically generated trajectory correction coefficients. =1.4.

[0156] S500, Attribution Diagnosis and Results Output:

[0157] In the final diagnostic stage, the normalized baseline residual between the target device temperature and the reference value is only 0.10 (absolute temperature difference of 10℃), which usually does not reach the alarm threshold in traditional algorithms. However, this invention utilizes a highly consistent sensitive factor mask ( ≈0.9) and the aforementioned trajectory correction coefficient ( =1.4) The residuals are double-weighted to calculate the final anomaly score. =1.4×(0.10×0.9)=0.126. This score successfully exceeded the dynamic alarm threshold, and the system immediately triggered an early warning, listing "charging gun head temperature" as the root cause with the highest contribution rate, thus achieving accurate early warning before the equipment overheated and burned out.

[0158] Example 2:

[0159] To verify the effectiveness and advancement of the method proposed in this invention, this embodiment conducted backtesting and comparative analysis based on a real historical dataset provided by a major domestic charging operator.

[0160] Experimental setup:

[0161] The experimental dataset covers 500 120kW DC charging piles from 50 public charging stations in a certain province, spanning from January to June 2024, with a total sample size exceeding 120 million. The dataset includes 158 real fault cases confirmed by maintenance work orders, covering various types such as fan failure, over-temperature, and voltage drift. The experiment selected the commonly used "static threshold method (Solution A)" and "isolated forest algorithm (Solution B)" as benchmarks and compared the detection results with the method of this invention (Solution C).

[0162] Results analysis:

[0163] As shown in Table 1 below, the present invention demonstrates significant advantages in all key indicators:

[0164] Table 1. Performance comparison of different algorithms on the full dataset:

[0165] Performance indicators Option A (Static Threshold) Option B (Isolated Forest) Solution C (This invention) Precision 0.985 0.723 0.942 Recall 0.456 0.889 0.968 F1-Score 0.623 0.797 0.955 Average early warning time 0 hours 12.5 hours 48.3 hours

[0166] The comparative results show that while the traditional scheme A has a low false alarm rate, it misses more than 50% of early faults (recall rate of only 45.6%). Scheme B, although improving the detection rate, suffers from a large number of false alarms due to its inability to distinguish environmental noise (precision rate of only 72.3%). In contrast, this invention (scheme C), by introducing a homogeneous dynamic benchmark and trajectory correction mechanism, not only improves the comprehensive performance index F1-Score to 0.955, but more importantly, thanks to its keen capture of evolutionary trends, this invention issues early warnings on average 48.3 hours before the occurrence of hard faults.

[0167] In conjunction with the above experiments, Figures 8-10 This further corroborates the experimental conclusions: Figure 8 The evolution trajectory comparison diagram shows the significant angle between the target vector direction and the group direction in the early stage of the fault; Figure 9 The timing diagram shows that the anomaly index calculated by the present invention exceeded the dynamic threshold in the early stage of the fault, while the traditional fixed threshold had not yet been triggered at this time. Figure 10The contribution rate histogram indicates the root cause of the failure, verifying the accuracy of the attribution logic.

[0168] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for multidimensional dynamic comparison and root cause analysis of the performance of new energy equipment assets, characterized in that, Includes the following steps: Obtain the target device's operating data and map it into a normalized feature vector; Search for neighboring devices that are closest to the feature distance of the normalized feature vector to construct a homogeneous set, and calculate the feature mean of the homogeneous set to generate a virtual standard benchmark. Calculate the statistical dispersion of the homogeneous set, and generate a sensitivity factor mask based on the statistical dispersion; Construct the state drift vectors of the target device and the virtual standard reference, calculate the directional similarity between the two, and generate trajectory correction coefficients; Using the sensitive factor mask and the trajectory correction coefficient, the residual vector between the target device and the virtual standard reference is weighted and calculated. Based on the weighting result, the abnormal state is determined and the root cause parameter is output.

2. The method for multidimensional dynamic comparison and root cause analysis of new energy equipment asset performance according to claim 1, characterized in that, The normalized feature vector includes environmental feature sub-vectors, operating condition feature sub-vectors, and performance feature vectors; The environmental feature sub-vectors are used to characterize the external boundary conditions of the equipment operation; The operating condition feature sub-vector is used to characterize the current load pressure of the equipment; The performance feature sub-vectors are used to characterize the response state of the core components inside the device.

3. The method for multidimensional dynamic comparison and root cause analysis of new energy equipment asset performance according to claim 2, characterized in that, The step of searching for the nearest neighboring device with the feature distance of the normalized feature vector to construct a homology set specifically includes: Define a feature weight vector, wherein the environmental feature sub-vector and the operating condition feature sub-vector are assigned a first weight value, and the performance feature sub-vector is assigned a second weight value, wherein the first weight value is greater than the second weight value; Based on the feature weight vector, calculate the weighted Euclidean distance between the target device and other devices in the entire network; A preset number of devices with the smallest weighted Euclidean distance that is less than a preset similarity threshold are selected to form the homogeneous set.

4. The method for multidimensional dynamic comparison and root cause analysis of new energy equipment asset performance according to claim 1, characterized in that, The step of generating the sensitivity factor mask based on the statistical dispersion specifically includes: Calculate the root mean square deviation of all devices in the homogeneous set relative to the virtual standard benchmark in each feature dimension, as the feature dispersion; Establish an inverse mapping relationship between the feature dispersion and the sensitivity factor; the smaller the feature dispersion, the larger the corresponding sensitivity factor value. The sensitivity factors of all feature dimensions are normalized to generate the sensitivity factor mask.

5. The method for multidimensional dynamic comparison and root cause analysis of new energy equipment asset performance according to claim 1, characterized in that, The step of constructing the state drift vectors of the target device and the virtual standard reference, and calculating the directional similarity between them, specifically includes: Calculate the feature vector differences between the target device and the virtual standard reference at the current time and the previous time, respectively, to obtain the target device state drift vector and the group reference drift vector; Calculate the cosine similarity between the target device state drift vector and the group baseline drift vector to quantify the consistency of their evolution directions.

6. The method for multidimensional dynamic comparison and root cause analysis of new energy equipment asset performance according to claim 5, characterized in that, The step of generating trajectory correction coefficients specifically includes: Calculate the magnitude of the target device state drift vector. If the magnitude is less than a preset effective change threshold, set the trajectory correction coefficient to a preset reference value. If the magnitude is greater than or equal to the preset effective change threshold, the trajectory correction coefficient is calculated based on the cosine similarity; the smaller the cosine similarity, the larger the generated trajectory correction coefficient, and the trajectory correction coefficient is greater than or equal to the preset benchmark value.

7. The method for multidimensional dynamic comparison and root cause analysis of new energy equipment asset performance according to claim 1, characterized in that, The step of using the sensitivity factor mask and the trajectory correction coefficient to perform weighted calculation of the residual vector between the target device and the virtual standard reference specifically includes: Calculate the absolute deviation between the normalized feature vector and the virtual standard benchmark, and use it as the basic residual vector; The Hadamard product operation is performed on the basic residual vector using the sensitivity factor mask to achieve element-wise weighting of the feature dimensions; The trajectory correction coefficients are used to perform scalar multiplication on the vector after the Hadamard product to generate a fusion anomaly score vector as the weighted result.

8. The method for multidimensional dynamic comparison and root cause analysis of new energy equipment asset performance according to claim 7, characterized in that, The step of determining the abnormal state based on the weighted result specifically includes: Calculate the Euclidean norm of the fusion anomaly scoring vector to obtain the global anomaly index of the target device; Calculate the mean and standard deviation of the global anomaly index of all devices within the same source set, and generate a dynamic alarm threshold based on the Raida criterion; If the global anomaly index of the target device is greater than the dynamic alarm threshold, then the target device is determined to be abnormal.

9. The method for multidimensional dynamic comparison and root cause analysis of new energy equipment asset performance according to claim 7, characterized in that, The step of outputting the root cause parameters specifically includes: The element values ​​of each dimension in the fusion anomaly scoring vector are sorted in descending order; Select the features with the highest sorting count as the main causes of failure; Calculate the proportion of the element value corresponding to each major fault cause to the sum of all element values ​​in the fused anomaly score vector, and output it as the contribution rate.

10. A system for multidimensional dynamic comparison and root cause analysis of new energy equipment asset performance, used to implement the method for multidimensional dynamic comparison and root cause analysis of new energy equipment asset performance as described in any one of claims 1-9, characterized in that, include: The feature mapping module is configured to acquire the target device's operating data and map it into a normalized feature vector; The benchmark construction module is configured to search for neighboring devices with the closest feature distance to the normalized feature vector to construct a homologous set, and to calculate the feature mean of the homologous set to generate a virtual standard benchmark. The weight generation module is configured to calculate the statistical dispersion of the homogeneous set and generate a sensitivity factor mask based on the statistical dispersion. In addition, the state drift vectors of the target device and the virtual standard reference are constructed, the directional similarity between the two is calculated, and trajectory correction coefficients are generated; The attribution diagnosis module is configured to use the sensitive factor mask and the trajectory correction coefficient to perform weighted calculation on the residual vector between the target device and the virtual standard benchmark, determine the abnormal state based on the weighted result, and output the root cause parameters.

Images