Distributed generator set production process state monitoring method and system

Abstract

The present application relates to the technical field of data processing, in particular to a kind of distributed generator production process state monitoring method and system, adaptive segmented collection is carried out to multimodal parameter in edge end, and operating characteristic vector containing statistics, trend and coupling feature is extracted;Gaussian mixture baseline model is constructed and dynamically updated, local health index is calculated, and early degradation evidence is generated when index is lower than threshold value or continuously decreases;Relationship graph is constructed based on electrical connection, geographical proximity and control domain in cloud, and system consistency score is calculated;Local health index and consistency score are fused, and alarm is output according to three priority levels of individual serious degradation, systematic abnormality and early weak degradation.The present application reduces communication load, improves early and accuracy of abnormal detection, and can effectively identify systematic operation disorder risk.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for monitoring the production process status of distributed generator sets. Background Technology

[0002] With the widespread integration of renewable energy, distributed generation systems (such as combinations of distributed wind turbines, photovoltaic arrays, and energy storage units) have become an important part of modern energy networks. Traditionally, the operational status monitoring of these distributed generators has mainly relied on two methods: one is a local monitoring system based on preset threshold alarms, which triggers an alarm when parameters such as temperature, vibration, and voltage exceed fixed ranges; the other is to centrally transmit all operational data to a cloud data center for offline analysis and status assessment.

[0003] However, these existing methods have significant shortcomings. First, distributed generator sets are typically deployed in geographically remote and environmentally complex areas with limited and unstable communication network bandwidth. Uploading massive amounts of high-frequency monitoring data to the cloud results in high data transmission latency and packet loss rates, making real-time status awareness difficult. Second, generator sets of different brands, models, and commissioning times have varying performance benchmarks. Using a uniform fixed threshold for anomaly detection easily leads to missed detections (insensitivity to early, subtle degradation) or false alarms (identifying normal operating fluctuations as faults). Furthermore, traditional methods focus on the independent analysis of individual units, failing to effectively utilize operational correlation information between units due to geographical proximity or electrical coupling. This results in the inability to identify comprehensive production status anomalies caused by decreased system coordination. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for monitoring the production process status of distributed generator sets, which reduces communication load, improves the early detection and accuracy of anomalies, and can effectively identify the risk of systemic operational misalignment.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a method for monitoring the production process status of a distributed generator set, comprising the following steps:

[0006] The multimodal parameter sequences of the distributed generator set are timestamped and aligned. Based on the comparison results of the change rate of each parameter sequence with the stationary threshold and active threshold of dynamic change, the continuous data stream is adaptively divided into stationary segments and active segments and encapsulated into basic monitoring data blocks.

[0007] For each basic monitoring data block, statistical feature sub-vectors, trend feature sub-vectors, and coupling feature sub-vectors are extracted in parallel, and the three are concatenated and normalized to generate a unified running feature vector;

[0008] At the edge, an adaptive state baseline model based on a Gaussian mixture model is maintained for each unit. This model is initially constructed using historical normal operation feature vectors and continuously evolved through a progressive incremental update mechanism based on sliding time windows and confidence filtering.

[0009] The negative log-likelihood probability is calculated based on the current running feature vector and the adaptive state baseline model. The local health index is obtained through normalization mapping. When the local health index is lower than the first threshold or decreases monotonically for three consecutive periods and the cumulative decrease exceeds the second threshold, the anomaly tracing process is initiated. The three feature dimensions with the highest contribution and their actual values ​​are extracted and packaged into an early degradation evidence data package.

[0010] A static relationship map is pre-constructed based on electrical connections, geographical proximity, and control domain correlation. For each unit, the weighted covariance of its local health index, weighted median deviation from the neighboring units in the region, and historical trends is calculated and fused to obtain a system consistency score.

[0011] Based on the local health index and system consistency score, the system judges three abnormal states in order of priority: severe individual degradation, systemic coordination abnormality, and early slight degradation. The corresponding alarm level and evidence information are output, and the normal status of the units that meet the normal conditions is recorded.

[0012] Among them, the stable threshold and the active threshold are dynamically changing reference values: the stable threshold is defined as the 90th percentile of the normal fluctuation rate of the unit's parameters under typical undisturbed operating conditions, and the active threshold is defined as the set percentage of the rate of change of the rated parameters per second or three times the standard deviation of normal operation. Both thresholds are recalculated weekly based on the historical operating characteristics of the unit and supplemented by adaptive correction based on event frequency.

[0013] The adaptive segmentation of the continuous data stream into stationary and active segments includes:

[0014] When the rate of change of any parameter exceeds the active threshold, the channel and all other channels that are physically coupled with it automatically switch from a low sampling frequency to a high sampling frequency and start a circular buffer window. The default range of the window is 5 seconds before the trigger point to 10 seconds after the trigger point. When multiple consecutive active events occur, the window is dynamically extended and merged into a complete active segment, cutting the continuous data stream into a stable segment and an active segment.

[0015] The extraction of the trend feature sub-vectors adopts an adaptive trend fitting based on segment type: a one-time overall linear fitting is performed on the stable segment data block to obtain the first-order trend slope and the second-order acceleration; for the active segment data block, a variable-width sliding time window is used to automatically identify inflection points and divide it into multiple monotonic segments, and the segment with the largest absolute value of the rate of change among all segments is taken as the main trend feature, while the average trend feature and the number of inflection points are calculated at the same time.

[0016] The extraction of the coupled feature sub-vectors includes: for each pair of predefined physical association parameters, calculating the average value and standard deviation of the sliding window correlation coefficient as an index of coupling degree and stability, and calculating the ratio of the change amplitude of the response parameter to the excitation parameter as the dynamic response ratio; the normalization is online rolling normalization, that is, performing zero-mean unit variance transformation based on the mean and standard deviation of each feature dimension of the most recent 100 normal data blocks.

[0017] The adaptive state baseline model adopts a Gaussian mixture model with a diagonal covariance matrix approximation, and the number of Gaussian components is automatically determined by the Bayesian information criterion. The incremental update only uses the running feature vectors that are selected as stationary segments and whose local health index is not lower than the health threshold, and maintains a sliding time window of length 200. Only one expectation maximization iteration is performed each time the data changes within the window. After the incremental update, an independent validation set is used to evaluate the model fit. If the fit deteriorates, the model parameters before the update are rolled back.

[0018] The normalization mapping of the local health index is based on the 5th and 95th percentiles of the negative log-likelihood probability under the historical normal state of the unit, which are linearly transformed to make the health index range between 0 and 1.

[0019] The first threshold is 0.3, and is temporarily adjusted to 0.2 within the first 100 hours of unit commissioning; the default cumulative reduction in the second threshold is 15%.

[0020] The anomaly tracing process first determines the principal Gaussian component to which the current running feature vector belongs. Then, it replaces the value of each dimension with the mean of the corresponding dimension of the principal Gaussian component and calculates the change in the negative log-likelihood probability to obtain the deviation contribution of each dimension. After sorting, the three dimensions with the highest contribution, along with their measured values ​​and deviation directions, are packaged into an early degradation evidence data package.

[0021] In the static relationship graph, the edge weights emphasize the adjacency strength coefficient, with the weight of the direct electrical connection criterion being 1.0, the weight of the geographical proximity criterion decreasing linearly to 0.2 with distance, and the weight of the control domain correlation criterion being 0.6. The health index deviation is calculated using the weighted median instead of the arithmetic mean, with the weight being the adjacency strength coefficient. The system consistency score is obtained by weighted fusion of the health index deviation component and the trend inconsistency component, with the health index deviation component having a weight of 0.6 and the trend inconsistency component having a weight of 0.4. The base value of the third threshold is 0.65, which is dynamically adjusted according to the node degree of the unit in the relationship graph. When the degree is 1, it is lowered to 0.55, and when the degree is greater than or equal to 5, it is raised to 0.70. Systemic operational misalignment risk is only confirmed when the value is below the adjusted threshold for two consecutive analysis periods.

[0022] The priority determination adopts a hit-and-stop strategy: the first priority is severe individual degradation, triggered when the local health index falls below the first threshold, outputting a red alarm and adding evidence of early degradation; the second priority is systemic coordination anomaly, triggered when the local health index is not lower than the first threshold, the system consistency score is lower than the third threshold, and the average local health index of the unit and all its neighbors is greater than 0.8, outputting a yellow alarm; the third priority is early minor degradation, triggered when the first two priorities are not met but the monotonically decreasing condition of the second threshold with a cumulative decrease is met, outputting a blue warning with evidence of early degradation and the estimated remaining health time; when none of the above abnormal conditions are met at the same time, it is determined to be in normal operation, the normal status record is recorded and a green icon is displayed.

[0023] Secondly, the present invention provides a distributed generator set production process status monitoring system for implementing the distributed generator set production process status monitoring method provided in the first aspect, comprising:

[0024] Several edge computing nodes, each edge computing node being associated with a group of distributed generator sets, wherein each edge computing node contains:

[0025] The data acquisition and preprocessing unit is used to timestamp the multimodal parameter sequences of the distributed generator set, and adaptively divide the continuous data stream into stable segments and active segments based on the comparison results of the change rate of each parameter sequence with the dynamic change stationary threshold and active threshold, and encapsulate them into basic monitoring data blocks.

[0026] The feature extraction unit is used to extract statistical feature sub-vectors, trend feature sub-vectors, and coupling feature sub-vectors in parallel for each basic monitoring data block, and then concatenate and normalize the three to generate a unified running feature vector;

[0027] The model management unit is used to maintain an adaptive state baseline model based on a Gaussian mixture model for each unit at the edge. The model is initially constructed using historical normal operation feature vectors and continuously evolves through a progressive incremental update mechanism based on sliding time windows and confidence filtering.

[0028] The evaluation unit is used to calculate the negative log-likelihood probability based on the current running feature vector and the adaptive state baseline model. After normalization mapping, the local health index is obtained. When the local health index is lower than the first threshold or monotonically decreasing for three consecutive periods and the cumulative decrease exceeds the second threshold, the anomaly tracing process is initiated. The three feature dimensions with the highest contribution and their actual values ​​are extracted and packaged into an early degradation evidence data package.

[0029] And a cloud analytics center, which communicates with each edge computing node via a network, and the cloud analytics center includes:

[0030] The collaborative analysis unit is used to pre-build a static relationship map based on electrical connections, geographical proximity and control domain correlation, and calculate the weighted covariance of the local health index of each unit with the weighted median deviation of neighboring units in the neighborhood and the historical trend, and integrate them to obtain a system consistency score.

[0031] The fusion decision unit is used to determine three abnormal states in order of priority: severe individual degradation, systemic coordination abnormality, and early minor degradation, based on the local health index and system consistency score. It outputs the corresponding alarm level and evidence information, and records the normal status of units that meet the normal conditions.

[0032] The edge computing node uploads local health index and early degradation evidence data packets to the cloud analysis center, which then distributes or displays the system consistency score and comprehensive status alarm results.

[0033] This invention discloses a method and system for monitoring the production process status of distributed generator sets. At the edge, multimodal parameters are adaptively segmented and collected to extract operational feature vectors containing statistical, trend, and coupling characteristics. A Gaussian mixture baseline model is constructed and dynamically updated to calculate a local health index. When the index falls below a threshold or continues to decline, early degradation evidence is generated. In the cloud, a relationship graph is constructed based on electrical connections, geographical proximity, and control domains to calculate a system consistency score. The local health index and consistency score are integrated, and alarms are output with three priority levels: severe individual degradation, systemic coordination anomalies, and early minor degradation. This invention reduces communication load, improves the early detection and accuracy of anomalies, and effectively identifies the risk of systemic operational misalignment. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0035] Figure 1 This is a schematic diagram of the steps of a distributed generator set production process status monitoring method according to the first embodiment of the present invention.

[0036] Figure 2 This is a flowchart illustrating a method for monitoring the production process status of a distributed generator set provided by the present invention.

[0037] Figure 3 This is a structural schematic diagram of a distributed generator set production process status monitoring system according to the second embodiment of the present invention. Detailed Implementation

[0038] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0039] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms "a," "the," and "the" as used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0040] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0041] The first embodiment of this application is as follows:

[0042] Please see Figures 1-3 This invention provides a method for monitoring the production process status of a distributed generator set, comprising the following steps:

[0043] S1. The multimodal parameter sequences of the distributed generator set are timestamped and aligned. Based on the comparison results of the change rate of each parameter sequence with the stationary threshold and active threshold of dynamic change, the continuous data stream is adaptively divided into stationary segments and active segments and encapsulated into basic monitoring data blocks.

[0044] Specifically, for the target distributed generator set, the data acquisition and preprocessing unit collects multimodal data streams reflecting the production process in parallel through the following three channels:

[0045] Acquisition of Electrical Parameter Sequences: Electrical parameters include output voltage, output current, frequency, and active power. These parameters are read directly from the digital communication interface of the generator set's grid-connected inverter controller or power quality monitoring module. Edge computing nodes receive periodically broadcast electrical measurement values ​​from the controller via a fieldbus using a polling or subscription method. To obtain transient details, the acquisition unit performs dual-channel sampling of active power and current: one channel receives the effective values ​​calculated internally by the controller at a lower rate, while the other channel samples the raw waveform directly from the analog outputs of the current transformers and voltage transformers via the analog-to-digital converter built into the edge node. The data from both sources will be fused based on the same time base in subsequent steps.

[0046] Mechanical parameter acquisition: Mechanical parameters include spindle vibration amplitude, gearbox temperature, and bearing acceleration. Temperature is sensed using platinum resistance temperature sensors attached to the gearbox housing and bearing housing; the output resistance value is converted to a digital value by the analog input module of the edge node. Vibration and acceleration parameters are acquired using piezoelectric accelerometers, either magnetically attached or bolted. The sensor's output charge signal is converted to a voltage signal by a built-in charge amplifier and then input to an analog-to-digital converter. For vibration parameters, the acquisition unit retains the original waveform characteristic values ​​for at least one rotation cycle.

[0047] Environmental parameter sequence acquisition: Environmental parameters include wind speed (for wind turbine generators), irradiance (for photovoltaic power generation units), and ambient temperature. Wind speed is acquired by an ultrasonic anemometer installed on the top of the generator nacelle or a standalone wind measurement tower and transmitted to the edge nodes in digital message form via an RS-485 interface. Irradiance is measured using a silicon photovoltaic cell total radiation meter, outputting an analog current signal of 0~20mA. Ambient temperature is measured using a semiconductor temperature sensor with a radiation shield. Each edge computing node is equipped with an independently operating real-time clock counter with a resolution of 1 millisecond. This clock is network-independent and driven by a crystal oscillator inside the node. When an edge node starts up, it first performs a coarse calibration with a reference clock in the cloud via a network time protocol, controlling the initial offset of the local clock within ±10 milliseconds. Subsequently, each time a data sample is received from any sensor, the acquisition unit immediately timestamps it with the local clock value at the time the data was read. The system performs an automatic alignment calibration upon initial installation. During calibration, an electrical pulse trigger signal generated by an edge node is simultaneously injected into all channels, inducing a transient in the current signal and a slow change in the ambient temperature sensor. By detecting the occurrence time of the response signal in each channel, the system calculates the inherent delay offset of each channel relative to the trigger signal. Subsequently, during real-time acquisition, the timestamp of each data sample is automatically subtracted from the inherent offset of that channel, thereby unifying the effective time reference of all sensor data to the same event start point. The stability threshold and activity threshold are adaptively calculated based on the historical operating characteristics and rated parameters of each generator set, and the specific determination process is as follows:

[0048] The stability threshold is defined as the upper limit of the rate of change of the parameter under normal random fluctuations under "typical undisturbed operating conditions". During the first week after the unit's initial commissioning, the system continuously records the 1-second sliding rate of change distribution of each parameter during periods of no faults and no scheduling actions, and takes the 90th percentile of this distribution as the initial stability threshold for that parameter. Thereafter, this threshold is recalculated weekly to slowly adapt to the natural aging of the unit's performance.

[0049] The active threshold is defined as the minimum rate of change that a parameter should reach when a critical change in the situation occurs. Its value is typically related to the rated response capability of the equipment: for electrical parameters (such as power), the active threshold is set to 15% of the rated power change per second; for mechanical vibration parameters, it is set to the rate of change corresponding to three times the standard deviation of the root mean square value of the vibration during normal operation; for environmental parameters, the active threshold for wind speed is set to a change exceeding 2 meters per second. Furthermore, the system employs a rule-based adaptive correction: if no high-sampling events have been triggered consecutively in the past 10 minutes, the active threshold is multiplied by 0.95 and slowly decreased to improve sensitivity to minor fluctuations; if high-sampling events are too frequent (more than 3 times per minute), the active threshold is multiplied by 1.05 and increased to avoid generating excessive redundant data. Each parameter sequence is configured with two sampling frequencies based on its physical rate of change: a low sampling frequency and a high sampling frequency.

[0050] Low sampling frequencies are used for recording steady-state conditions. For slowly changing parameters such as temperature and irradiance, the low sampling frequency is 0.1 Hz; for RMS voltage and current values ​​and active power, the low sampling frequency is 1 Hz; and for the root mean square value of vibration acceleration, the low sampling frequency is 5 Hz. High sampling frequencies are used to capture transient processes. When the rate of change of any parameter exceeds its activity threshold, that channel and other related channels physically coupled with it automatically switch to high sampling mode. The high sampling frequencies are set as follows: vibration and acceleration parameters are acquired at 1000 Hz for raw waveform acquisition; instantaneous electrical voltage and current are acquired at 4000 Hz (meeting the requirements for 50 Hz power frequency 80th harmonic analysis); and rapidly changing environmental physical quantities such as temperature and wind speed are acquired at 20 Hz. Different parameter channels can be independently or in conjunction with each other at different sampling frequencies. When a channel triggers high sampling due to activity, the edge node automatically raises all associated channels of the unit to the high sampling frequency simultaneously to maintain the synchronization integrity between multimodal data. When the rate of change of a parameter channel exceeds the active threshold and triggers high-sampling mode, the data acquisition and preprocessing unit does not start recording only from the moment of triggering. To preserve the complete process before and after the anomaly, the system starts a circular buffer window. The default time range of this window is 5 seconds before the trigger point to 10 seconds after the trigger point. The data for the 5 seconds before the trigger comes from a continuously running pre-buffer inside the edge node. This pre-buffer continuously records the data of all channels for the most recent 5 seconds at a high sampling frequency. The data for the 10 seconds after the trigger is recorded continuously at a high sampling frequency from the moment of triggering until the full 10 seconds are reached. If another channel triggers a new active event within 10 seconds after the trigger, the system will dynamically extend the window so that the new trigger point also has complete records for the 5 seconds before and 10 seconds after it, ultimately forming a continuous, variable-length active segment data stream.

[0051] The specific window range (5 seconds and 10 seconds) can be configured according to the type of generator set: for wind turbine generator sets with high inertia, the window can be extended to 10 seconds before triggering and 20 seconds after triggering; for photovoltaic inverters with fast response, the window can be shortened to 2 seconds before triggering and 3 seconds after triggering. Segmentation refers to cutting the original multimodal continuous data stream into several continuous "basic monitoring data blocks" with relatively consistent internal states according to the time axis. The segmentation process consists of the following steps:

[0052] Step A: Initialize an active segment. The system continuously monitors the real-time change rate of all channels. When the current change rate of any channel exceeds its active threshold for the first time, the system immediately creates a new "active segment" data block and extracts the data from the pre-buffered area for the 5 seconds prior to the trigger time as the header of the data block. At the same time, it begins writing data generated after the trigger time at a high sampling frequency.

[0053] Step B: Merge adjacent active events. If, before the current active segment ends (i.e., within the 10-second window after the recording is triggered), the rate of change of other channels in the same unit exceeds its active threshold, or the rate of change of the current channel experiences a second jump exceeding the active threshold, the system will reset the recording window of the current active segment, extending the window end point back 10 seconds from the current moment.

[0054] Step C: Termination of the active segment and formation of the stable segment. When the rate of change of all parameter channels is below its active threshold for 3 consecutive seconds, and a complete 10-second recording period has been completed since the last active event was triggered, the current active segment is closed. After the active segment is closed, the system automatically enters the "stable segment" recording mode. In the stable segment, all channels record data at a low sampling frequency. There are two conditions for the stable segment to end: First, if the rate of change of a certain channel is detected to exceed the active threshold again, the system will close the current stable segment and start a new active segment; second, if the duration of the stable segment exceeds a preset maximum length (e.g., 30 minutes), the system will forcibly close the current stable segment and start a new stable segment to prevent a single data block from becoming too large.

[0055] Step D: Generate basic monitoring data blocks. Each time an active or stable segment is closed, the system encapsulates all multimodal data (including timestamps, parameter type identifiers, and sampled value sequences) within that time period into an independent "basic monitoring data block". Each data block includes descriptive metadata: block type (active or stable segment), start and end timestamps, trigger channel (only for active segments), and the maximum rate of change for each parameter during the block's duration.

[0056] S2. Extract statistical feature sub-vectors, trend feature sub-vectors, and coupling feature sub-vectors in parallel for each basic monitoring data block, and then concatenate and normalize the three to generate a unified running feature vector.

[0057] Specifically, statistical feature sub-vectors are used to describe the overall distribution of each parameter sequence in the data block, independent of time order. For each parameter sequence contained in the basic monitoring data block, the feature extraction unit calculates the following five statistics: Mean: The arithmetic mean of all sampled values ​​of the parameter sequence over a period of time, reflecting its overall level. Standard Deviation: Measures the dispersion of sampled values ​​around the mean; a larger value indicates more severe fluctuations. Kurtosis: An indicator describing the thickness of the tail of the distribution. In calculation, the system first calculates the average of the fourth power of the deviation of each sampled value from the mean, then divides it by the fourth power of the standard deviation, and finally subtracts a baseline value to make the kurtosis of the normal distribution zero. A positive kurtosis indicates a thicker tail; a negative kurtosis indicates a thinner tail. Skewness: An indicator measuring the asymmetry of the distribution. In calculation, the system calculates the average of the cube of the deviation of the sampled values ​​from the mean, then divides it by the cube of the standard deviation. A positive skewness indicates a right-skewed distribution, and a negative skewness indicates a left-skewed distribution. Interquartile range: The difference between the 75th percentile and the 25th percentile after sorting the sampled values ​​from smallest to largest, reflecting the distribution range of the middle 50% of the data.

[0058] For stationary data blocks, the above statistics are calculated based on all low-sampling-frequency sampling points within the entire data block. For active data blocks, since the data blocks contain dense data with high sampling frequencies, the system performs a lightweight downsampling operation before calculation—compressing the high-sampling data by taking a median of every 50 points to reduce computational burden while preserving distribution characteristics. The five statistics for each parameter sequence are arranged sequentially to form the statistical sub-segment corresponding to that parameter.

[0059] Trend feature sub-vectors are used to capture the dynamic direction and velocity information of the evolution of each parameter within a data block over time. The feature extraction unit performs the following operations on each parameter sequence:

[0060] Step 1: Differentiating Trend Processing Strategies for Stable and Active Segments. The system reads the type label of the current data block. If the data block is a stable segment, a one-time linear trend fitting is performed on all sampling points within the entire block to obtain the overall trend. If the data block is an active segment, segmentation is performed: Since the changes within an active segment are drastic and may involve multiple stages, the system uses a widened sliding time window (gradually increasing the window width) to find the inflection point that minimizes the segmented fitting error, dividing the active segment into several monotonic segments, and then performing linear fitting on each segment separately. Finally, the segment with the largest absolute value of the rate of change among all segments is taken as the "primary trend feature" of this parameter, while the average of the rates of change of all segments is taken as the "average trend feature".

[0061] Step 2: Calculate the slope of the first-order trend. For the time interval to be fitted, use time as the independent variable and the parameter values ​​as the dependent variable, and fit a straight line using the least squares method. The slope of this line is the slope of the first-order trend. The sign of the slope indicates an increase or decrease, and the absolute value indicates the rate of change.

[0062] Step 3: Calculate the second-order acceleration. Based on the first-order slope, divide the time interval into two equal-length sub-intervals. Calculate the slope of the second sub-interval, subtract the slope of the first sub-interval from the slope of the second sub-interval, and then divide by the time interval to obtain the second-order acceleration. A positive acceleration indicates that the trend is accelerating and strengthening, while a negative acceleration indicates that the trend is slowing down or reversing.

[0063] Step 4: Generate trend feature sub-vectors. For each parameter sequence, the system outputs a quadruple: main trend slope, goodness of fit of the main trend, second-order acceleration, and the number of trend inflection points (0 for stationary segments, and potentially greater than 0 for active segments). The quadruples from all parameter sequences are concatenated in a fixed order to form the trend feature sub-vectors.

[0064] Coupled feature sub-vectors are used to characterize the temporal synchronization and linkage between different parameter sequences. These coupling relationships reflect the constraints and transmission characteristics of the internal physical processes of the generator set. The feature extraction unit predefines a set of physically related parameter pairs, which are pre-configured according to the generator set type. For each predefined parameter pair, the system calculates the following two coupling features:

[0065] Sliding window correlation coefficient: On the time axis of the data block, the Pearson correlation coefficient (a statistic describing the degree of linear correlation between two variables, ranging from -1 to 1, with the absolute value closer to 1 indicating a stronger correlation) is calculated by sliding the window length (30-second window for stationary segments and 1-second window for active segments) between the two parameter sequences within that window. Then, the average of all window correlation coefficients is taken as the overall coupling degree; simultaneously, the standard deviation of all window correlation coefficients is taken as the coupling stability index.

[0066] Dynamic response ratio: For parameter pairs with a clear excitation-response relationship, the system calculates the ratio of the magnitude of change of the response parameter to that of the excitation parameter. Specifically, it first identifies the change events in the excitation parameter sequence, then finds the magnitude of change of the response parameter within the same time period. The response ratio is defined as (magnitude of change of response parameter) divided by (magnitude of change of excitation parameter). For multiple change events, the average value is taken.

[0067] The average value of the sliding correlation coefficient, the standard deviation of the correlation coefficient, and the dynamic response ratio of all predefined parameter pairs are arranged in order to form a coupled feature sub-vector.

[0068] After the feature extraction unit computes statistical feature vectors, trend feature vectors, and coupling feature vectors in parallel, it needs to fuse them into a single "running feature vector" that can be accepted by the subsequent baseline model. The fusion process includes three steps: concatenation, normalization, and metadata appending.

[0069] The three sub-vectors are concatenated end-to-end according to a sequence of first statistics, then trends, and finally coupling, forming a large, high-dimensional vector. This concatenation order ensures that the downstream model can perform calculations using the same feature indices in each processing iteration.

[0070] The feature extraction unit performs online normalization on the concatenated vector. The normalization parameters (mean and standard deviation) are derived from the statistical accumulation of historical normal data blocks for the unit: the system maintains a rolling buffer storing the feature values ​​of the most recent 100 data blocks confirmed as normal, and calculates the mean and standard deviation of each feature dimension in real time. Then, for each dimension of the currently running feature vector, a transformation is performed by subtracting the mean and dividing by the standard deviation, so that each feature dimension approximately follows a distribution with zero mean and unit variance over long-term operation. For newly commissioned units that have not yet accumulated sufficient historical data, the system uses the factory standard feature statistics of similar units as the initial normalization parameters, and gradually switches to self-learning parameters after running the first 200 data blocks.

[0071] The normalized feature vector is itself a numerical sequence. The feature extraction unit also adds three metadata tags to this vector: ① data block type (stationary segment / active segment); ② the absolute time interval corresponding to the data block (start time and end time); ③ the duration within the data block (in seconds). This metadata can be used to determine whether the feature vector is eligible to participate in model updates. Finally, a complete running feature vector is defined as: a composite data structure consisting of the above normalized numerical sequence and its attached metadata.

[0072] S3. At the edge, maintain an adaptive state baseline model based on a Gaussian mixture model for each unit. This model is initially constructed using historical normal operation feature vectors and continuously evolves through a progressive incremental update mechanism based on sliding time windows and confidence level filtering.

[0073] Specifically, the adaptive state baseline model employs an improved Gaussian mixture model that supports incremental updates as its core data structure. The model's specific structure comprises the following three components:

[0074] The model internally maintains multiple Gaussian components, each with three attributes: a mean vector, a covariance matrix, and a mixture weight. Initially, the number of Gaussian components is automatically selected using the Bayesian information criterion, typically between 3 and 8. Each Gaussian component essentially characterizes the "health fingerprint" distribution of the unit under a certain typical operating condition.

[0075] To support dynamic updates of the model, the model also needs to store some auxiliary statistics, including: the total number of "running feature vectors" involved in training, the current cumulative number of valid samples for each Gaussian component, and the timestamp of the last model update.

[0076] The metadata attached to the model includes: a unique unit identifier, model creation time, model version number, and a "baseline validity period" flag. When more than a certain period (e.g., 30 days) has passed since the last update and no new normal data blocks have been received, the model will be marked as "stale" and will be forced to undergo a compensatory update upon receiving new data again.

[0077] The covariance matrix is ​​approximated using a diagonalization method instead of the full matrix form. The diagonalization approximation assumes that the feature dimensions of each Gaussian component are independent of each other. Although it sacrifices some correlation modeling capabilities, it significantly reduces the computational and storage burden.

[0078] Each generating unit requires initial baseline model construction upon initial commissioning or after major maintenance. The initial construction phase uses operational feature vectors collected during the unit's "historical normal operating process" as training samples. The specific process is as follows:

[0079] The system automatically marks the previous complete production cycle after the unit's commissioning (e.g., 7 days or a continuous 300-hour operating interval) as the baseline period. During this period, maintenance personnel ensure the unit is in a healthy state free from faults, abnormal wear, and improper control. Steps one and two continue to run, extracting operational feature vectors from all basic monitoring data blocks generated within this baseline period. The system removes data blocks marked as "active segments," retaining only "stable segments." Furthermore, data blocks exhibiting significant measurement noise or communication packet loss are also removed. After filtering, the remaining stable segment feature vectors constitute the initial training sample set.

[0080] The system performs cluster analysis on the initial training sample set. Starting with one component, it incrementally fits a Gaussian mixture model, adding one component at a time and calculating the Bayesian information criterion. This criterion balances model goodness of fit with model complexity: as the number of components increases, goodness of fit improves, but model complexity also increases. When the decrease in the Bayesian information criterion after adding one component is less than a pre-set improvement threshold, further additions stop; this number of components represents the optimal value.

[0081] Using a defined number of Gaussian components, the Expectation-Maximization (EM) algorithm is run on the initial training sample set for parameter estimation. The process is as follows: First, the mean, covariance, and weights of each Gaussian component are randomly initialized. Then, two steps are iteratively executed: First, for each feature vector, its "responsibility" to each Gaussian component (i.e., the probability that the vector was generated by that component) is calculated; second, based on all vectors and their responsibility, the mean, covariance, and weights of each component are recalculated. This process is repeated iteratively until the parameters no longer change significantly. Finally, the initial mean vector, diagonal covariance vector, and mixed weights are obtained. At this point, the unit has its first usable adaptive state baseline model.

[0082] The model must be capable of dynamic updates, and the update process must be able to adapt to slow drift while preventing contamination by anomalous data. The specific process is as follows:

[0083] Step 1: Filter candidate feature vectors for updates. The model management unit continuously receives running feature vectors from Step 2. Only vectors that meet the following conditions are added to the update candidate pool: ① The data block type is "stable segment"; ② Within the time period corresponding to the vector, no local health index has fallen below the preset "health threshold" (the threshold is set to 0.6 by default); ③ There is at least a 10-block gap between the vector and the vector previously used for updates to avoid model bias due to excessive time correlation.

[0084] The second step involves maintaining a fixed-length sliding time window. The model management unit maintains a first-in, first-out queue called the "normal feature vector sliding window," with a fixed window length equal to the 200 most recently selected stationary feature vectors that are deemed "updatable." Whenever a new candidate vector passes the selection, it is pushed to the tail of the queue; if the queue is full, the oldest vector is removed from the head of the queue.

[0085] Step 3: Perform incremental updates. Each time the sliding window undergoes a substantial change, the model management unit initiates an incremental update. The model management unit first uses the number of Gaussian components, mean vector, covariance vector, and weights from the old model parameters as a starting point. Then, for all vectors within the sliding window, it performs an expectation-maximization iteration: first, it calculates the responsibility of each vector for each component based on the current parameters, and then recalculates the mean, covariance, and weights of each component based on these responsibility levels. After one iteration, the new model parameters become the current model. If the number of vectors within the sliding window does not reach the window length, only existing vectors are used for updating.

[0086] Step 4: Model Validation and Rollback Protection. The system retains the 50 most recent stationary feature vectors that were not selected into the sliding window (but passed health screening) as a validation set. After each incremental update, the system calculates the average negative log-likelihood probability of the updated model on the validation set. If this probability decreases relative to the previous update and the decrease exceeds a preset warning value, the update is deemed abnormal. The system discards the update result, restores the model parameters to the previous state, and sends a model update anomaly alarm to the cloud. Furthermore, if three consecutive updates are rolled back, the system triggers a higher-level alarm, prompting operations personnel to check for undetected changes in unit operating conditions or sensor drift.

[0087] Step 5: Model Snapshot and Recovery. Every 24 hours, the edge computing nodes save the current baseline model as a snapshot file, retaining the snapshots from the last 7 days. If the system unexpectedly restarts or detects corrupted model data, it can be restored to the most recent normal snapshot. Simultaneously, the cloud periodically pulls model summary information from each edge node for horizontal comparison of baseline differences between units of the same model, helping to identify common abnormal patterns.

[0088] S4. Calculate the negative log-likelihood probability based on the current running feature vector and the adaptive state baseline model, obtain the local health index through normalization mapping, and start the anomaly tracing process when the local health index is lower than the first threshold or decreases monotonically for three consecutive periods and the cumulative decrease exceeds the second threshold. Extract the three feature dimensions with the highest contribution and their actual values ​​and package them into an early degradation evidence data package.

[0089] Specifically, the local health index aims to quantify the deviation between the current operational feature vector and the unit's normal state space. A higher value indicates a closer approximation to historical normal conditions, while a lower value indicates a greater likelihood of anomalies. The calculation process includes the following three sub-steps:

[0090] Sub-step 1: Calculate the negative log-likelihood probability. The evaluation unit first inputs the current running feature vector into the adaptive state baseline model (Gaussian mixture model). This model consists of multiple Gaussian components, each with its own mean vector, diagonal covariance vector, and mixture weights. The system sequentially calculates the probability density value of the current feature vector belonging to each Gaussian component. Due to the use of the diagonal covariance approximation, the probability density calculation of each component can be simplified as follows: first, calculate the deviation of each dimension of the feature vector from the mean of the corresponding dimension of the component; divide the deviation by the standard deviation of that dimension (the square root of the covariance) to obtain the standardized deviation; then sum the squares of the standardized deviations of each dimension; and then perform exponential transformation and normalization. Finally, multiply the probability density values ​​of all components by their respective mixture weights and sum them to obtain the total probability density value of the current feature vector under the model. To convert the probability density into a monotonic index that is convenient for anomaly detection, the system takes the negative natural logarithm of the total probability density value to obtain the "negative log-likelihood probability".

[0091] Sub-step 2: Based on the normalized mapping of the historical normal distribution, the evaluation unit maintains a rolling distribution statistics of the historical negative log-likelihood probability for the unit. Specifically, this includes calculating the 95th percentile (denoted as P95) and 5th percentile (denoted as P5) of the negative log-likelihood probability on the feature vectors of the past 200 stationary segments that were confirmed as normal (passing the health screening in step 3). P5 represents the minimum degree of anomaly obtained under normal conditions, and P95 represents the acceptable boundary degree of anomaly under normal conditions. Then, the evaluation unit performs a linear mapping on the current negative log-likelihood probability: if the current value is less than or equal to P5, the mapped health index is 1.0; if the current value is greater than or equal to P95, the mapped health index is 0.0; if it is between the two, it is mapped linearly decreasing from P5 to P95 to between 1.0 and 0.0. The specific mapping method is: Health Index = 1.0 - (Current Negative Log-Likelihood Probability - P5) / (P95 - P5).

[0092] Sub-step 3: Output the local health index. The value obtained after the above mapping is the local health index of the unit at the current moment. The evaluation unit temporarily stores it in the local database of the edge node, along with the corresponding timestamp and the identifier of the running feature vector.

[0093] To determine whether the local health index indicates an abnormal state, the assessment unit employs two complementary threshold conditions: a first threshold is used to identify severe deviations where the health index remains persistently low; the second threshold is used to detect early deterioration trends where the health index, although not yet below the first threshold, has shown an accelerated decline. The specific criteria for setting the two thresholds are as follows:

[0094] The first threshold represents the lower limit of "significant anomaly". Its default value is set to 0.3, meaning that when the local health index is below 0.3, the current operating feature vector has fallen to the tail of the normal state space distribution (at least one standard deviation away from the region where the negative log-likelihood probability is higher than P95). For critical equipment or high-value units, maintenance personnel can increase the first threshold to 0.4 to improve sensitivity; for auxiliary units with strong fault tolerance, it can be lowered to 0.2 to reduce false alarms. In addition, this threshold can also be dynamically adjusted according to the unit's operating time: in the first 100 hours of initial operation, the first threshold is temporarily set to 0.2 because the baseline model has not yet fully converged; after 100 hours, it is restored to 0.3. The second threshold is used to detect early weak degradation trends when the health index continues to decline but has not yet reached the first threshold. This threshold includes two independent conditions: the number of consecutive monitoring periods and the percentage of cumulative decline. Specifically, the local health index shows a monotonically decreasing trend over three consecutive monitoring periods, and the cumulative decrease from the first period to the third period exceeds the "second cumulative decrease threshold", which is set to 15% by default.

[0095] After each calculation of the local health index, the assessment unit immediately executes the following judgment logic:

[0096] Condition 1: Determine if the current local health index is below the first threshold (default 0.3). If it is, the activation condition is immediately met. Condition 2: Determine if the local health index of the current period and the previous two periods shows a monotonically decreasing trend, and whether the cumulative decrease of the current period relative to three periods ago exceeds the second threshold.

[0097] If either of the above two conditions is true, the assessment unit will immediately initiate the anomaly tracing process. If neither condition is met, the tracing process will not be initiated; instead, the local health index will be stored and reported to the cloud.

[0098] When the evaluation unit determines that an anomaly tracing process needs to be initiated, it needs to identify which feature dimensions in the currently running feature vector are the main factors causing the decline in the health index, and package this information into an "early degradation evidence" data package. The specific steps are as follows:

[0099] Step 1: Determine the principal Gaussian component of the currently running feature vector. Specifically, calculate the probability density value of the feature vector under each Gaussian component, multiply it by the mixture weight of that component, and select the component with the largest product as the principal Gaussian component. Step 2: Calculate the deviation contribution of each feature dimension. After determining the principal Gaussian component, the evaluation unit extracts the mean vector (denoted as μ) and diagonal covariance vector (denoted as σ², i.e., the variance of each dimension) of that component. For each feature dimension (denoted as x_i) in the currently running feature vector, the system calculates the "standardized deviation distance" of that dimension relative to the principal Gaussian component: (x_i - μ_i) divided by σ_i (standard deviation). Since σ_i has the same dimensions as x_i, the obtained standardized deviation distance is a dimensionless number; positive or negative indicates the direction of deviation, and the absolute value indicates the degree of deviation. To more accurately characterize the contribution, the evaluation unit employs an approximation method of "marginal likelihood change": The values ​​of all dimensions in the current running feature vector, except for the current dimension, remain unchanged. Then, the unit calculates how much the new negative log-likelihood probability would decrease if only the value of the current dimension were replaced with the mean μ_i of that principal Gaussian component. The greater the decrease, the greater the contribution of that dimension to the current anomaly level.

[0100] Step 3: Sort and filter the three feature dimensions with the highest contribution. The evaluation unit sorts the deviation contribution of all feature dimensions from largest to smallest. The contribution can be positive or negative. The system only takes the top three dimensions with the largest contribution. If more than two of the top three dimensions with the largest contribution are below a minimum effective threshold, it indicates that the deviation of the entire feature vector is scattered and has no dominant factor. At this time, the source tracing result indicates "no clear dominant feature, overall state drift". Step 4: Generate early degradation evidence data package. The evaluation unit packages the information of the three feature dimensions selected above. Each dimension records the following four items: dimension name, current actual value, deviation direction, and qualitative description of deviation contribution. In addition, the data package also includes the following metadata: current local health index, triggering conditions (whether it is triggered by the first threshold or the second threshold), current timestamp, and the block type and start and end time of the basic monitoring data block from which the running feature vector of this data package originates.

[0101] Step 5: Storage and transmission of evidence data packets. Edge nodes store the generated early degradation evidence data packets in local non-volatile memory, forming an evidence chain arranged chronologically. Simultaneously, this data packet is immediately marked as "pending upload" and prioritized for transmission to the cloud analysis center in the next communication window.

[0102] S5. Based on electrical connections, geographical proximity, and control domain correlation, a static relationship map is pre-constructed. For each unit, the weighted covariance of its local health index, the weighted median deviation of its neighboring units in the neighborhood, and the historical trend is calculated and fused to obtain a system consistency score.

[0103] Specifically, the static relationship graph is an undirected weighted graph used to describe the mutual influence relationships between distributed generator sets due to electrical connections, geographical proximity, or control coupling. This graph is pre-built in the cloud and incrementally updated periodically (e.g., monthly) based on changes in system configuration, but remains static during real-time monitoring to reduce online computational complexity. The specific construction process is as follows:

[0104] Step 1: Determine the nodes in the data map. The cloud analysis center retrieves a list of all generator sets that need to be included in the monitoring from the distributed generation system's configuration database. Each generator set corresponds to a node in the data map, and the node identifier is a unique number used by that generator set in the system.

[0105] Step 2: Define and calculate three criteria for edge connection. Whether an edge needs to be established between any two units depends on the weighted judgment of the following three independent criteria. An edge is added between the two nodes if any one of the criteria is satisfied. The three criteria are as follows:

[0106] Criterion 1: Direct Electrical Connection Criterion. If two generator sets are directly connected to the same feeder, there is strong electrical coupling between them. Power fluctuations or voltage disturbances in either generator set may affect the other. The system determines this relationship by consulting the primary electrical wiring diagram. For generator pairs that meet this criterion, the edge strength weight is preset to the maximum value of 1.0.

[0107] Criterion 2: Geographic Proximity Criterion. For units that are geographically close but lack direct electrical connections due to their dispersed layout, the Euclidean distance between the system computer unit installation locations is used. If the geographical distance between two units is less than a set threshold (default value is 500 meters), an edge is created. The edge strength weight decreases linearly with distance: 0.8 when the distance is zero, and 0.2 when the distance reaches 500 meters. No edge is created if the distance exceeds 500 meters.

[0108] Criterion 3: Control Domain Correlation Criterion. If two generating units are coordinated and scheduled by the same higher-level station controller, the system establishes this relationship by reading the control system's region partitioning configuration file. For generating unit pairs that satisfy this criterion, the edge strength weight is preset to 0.6.

[0109] Step 3: Edge merging and weight normalization. If a pair of edges simultaneously satisfies multiple criteria, the combined weight is the maximum value among the weights of each criterion to avoid overemphasizing multiple connections. Finally, each edge is assigned a weight value between 0.2 and 1.0, called the "adjacency strength coefficient".

[0110] Step 4: Storing the graph structure. After construction, the static relationship graph is stored in the cloud database in the form of an adjacency list. For each node, the table records the identifiers of all its neighboring nodes and their corresponding adjacency strength coefficients. In addition, the system also stores the degree of each node (i.e., the number of neighbors) and global statistics of the entire graph (such as average degree, maximum degree, etc.).

[0111] The cloud-based analytics center triggers a full traversal analysis of all units at a fixed interval (every 5 minutes by default). For each unit node, the collaborative analysis unit executes the following complete process to calculate its current "system consistency score." This score measures the degree of coordination between the unit and its neighboring units in terms of health indices and their trends.

[0112] Step A: Obtain the basic data required for analysis. For the target unit currently being analyzed (denoted as Unit A), the collaborative analysis unit reads the following data from the cloud database:

[0113] The local health index of unit A at the current moment (denoted as H_A). The time series of the local health index of unit A over the past hour (12 points in total, one value every 5 minutes, including the current moment and the previous 11 moments), denoted as sequence S_A. All neighboring units of unit A in the static relation graph (denoted as set N(A)), and the adjacency strength coefficient w_{A→neighbor} of each neighboring unit. For each neighboring unit (denoted as unit B), read its local health index H_B at the current moment, and its health index time series S_B over the past hour.

[0114] Step B: Calculate the neighborhood bias of the current health index. The specific calculation process is as follows:

[0115] First, calculate the weighted median of the current health index of all neighboring units. The weighted median is calculated as follows: sort all neighboring units by their respective health indices from smallest to largest, then sum the adjacency strength coefficients of each neighboring unit. When the sum first exceeds half of the total sum of all neighbor strength coefficients, the health index of that neighboring unit is the weighted median. After obtaining the weighted median, calculate the absolute deviation of the target unit's health index from this median, and then divide it by the weighted standard deviation of all neighboring health indices (i.e., summing the squared deviations of each neighboring unit's health index from the weighted median multiplied by its weight, then dividing by the total weight, and finally taking the square root), resulting in a dimensionless "deviation degree". The larger the deviation degree, the more inconsistent the health status of this unit is with its surrounding units. To prevent the deviation degree from exploding due to an excessively small denominator, the denominator is forcibly set to 0.05 when the weighted standard deviation is less than 0.05. Finally, the deviation degree is limited to between 0 and 3 (those exceeding 3 are counted as 3), and then mapped to a score between 0 and 1 through a linear transformation, called the "health index deviation component". Mapping rules: When the deviation is 0, the deviation component is 0; when the deviation is 3 or above, the deviation component is 1; intermediate linear mapping.

[0116] Step C: Calculate the neighborhood covariance of the health index change trend. The specific process is as follows:

[0117] First, for each unit (including the target unit and each of its neighbors), the health index time series (12 points) over the past hour is used, and a simplified first-order linear trend fitting method similar to step two is employed to calculate the overall slope of the series (i.e., the change in health index per unit time, with positive values ​​indicating an increase and negative values ​​indicating a decrease). This slope is denoted as the trend value T_A (for the target unit) and T_B (for each neighboring unit).

[0118] Then, the weighted covariance of the target unit's trend value and the trend values ​​of all its neighbors is calculated. The covariance is calculated using the following steps: first, the weighted average of all neighbor trend values ​​is calculated (weights are the adjacency strength coefficients); then, for each neighbor, the product of (T_B - weighted average) and (T_A - weighted average) is calculated and multiplied by the neighbor's weight; finally, all products are summed and divided by the total weight to obtain the covariance value Cov. A positive covariance value indicates that the target unit and its neighbors' trends change in the same direction, while a negative value indicates an inverse change. To convert the covariance into a score between 0 and 1, the system first divides the covariance by the weighted standard deviation of the neighbor trend values ​​and the standard deviation of the target unit's own trend value, obtaining a dimensionless quantity in the form of a correlation coefficient, ranging from -1 to 1. This dimensionless quantity is then mapped to a "trend inconsistency component" between 0 and 1 using the formula (1 - correlation coefficient) / 2. When the correlation coefficient is 1 (perfectly positive correlation), the trend inconsistency component is 0; when the correlation coefficient is -1 (perfectly negative correlation), the trend inconsistency component is 1; when the correlation coefficient is 0, the component is 0.5.

[0119] Step D: Merge into a system consistency score. The system consistency score is a value between 0 and 1. A higher score indicates better state coordination (i.e., greater consistency) between the unit and its neighbors, while a lower score indicates a risk of misalignment. The collaborative analysis unit weights and merges the above two components (health index deviation component and trend inconsistency component). The default fusion method is:

[0120] System consistency score = 1 - (0.6 × health index deviation component + 0.4 × trend inconsistency component).

[0121] The weighting coefficients 0.6 and 0.4 represent the weighting coefficients for the two components, reflecting that the deviation of the current health index has a greater impact on consistency than the trend deviation. After calculation, the system consistency score of the target unit, along with its identifier and timestamp, is stored in the cloud database.

[0122] The third threshold is used to determine whether the system consistency score is low enough to trigger a systemic anomaly alarm. The third threshold is set based on the following statistical and engineering experience principles: (1) Under good system coordination, the system consistency scores of each unit are usually distributed between 0.75 and 0.95. Most cases with scores below 0.7 correspond to observable coordination anomalies. Therefore, the default base value of the third threshold is set to 0.65. (2) The cloud-based collaborative analysis unit introduces a dynamic correction based on node degree: for units with a degree of 1, the third threshold is lowered to 0.55; for units with a degree greater than or equal to 5, the third threshold is raised to 0.70. For units with a degree between 2 and 4, linear interpolation is used to determine the threshold. (3) When the cloud determines that the system consistency score of a unit is lower than the corrected third threshold, it will not immediately output an alarm. Instead, a "two-times-confirmation" mechanism is adopted: if the system consistency score of the unit is lower than its corresponding third threshold in two consecutive analysis cycles (i.e., two time points 5 minutes apart), it is finally determined that there is a risk of systemic operational misalignment; if it is lower than the threshold in only one cycle, it is marked as "suspicious" but no alarm is triggered until the next cycle is verified again. (4) The cloud will also periodically (e.g., weekly) recalculate the 10th percentile of the system consistency score of each unit during the normal operation period (without any alarm marking) in the past week, multiply the value by 0.9 as the new suggested threshold, and perform an exponential weighted average with the current third threshold to achieve slow adaptive adjustment.

[0123] S6. Based on the local health index and system consistency score, judge the three abnormal states in order of priority: severe individual degradation, systemic coordination abnormality, and early weak degradation, output the corresponding alarm level and evidence information, and record the normal status of the units that meet the normal conditions.

[0124] Specifically, a generator unit is considered to be in normal operation if and only if it simultaneously meets the following three conditions: Condition 1: The unit's current local health index is not lower than the first threshold. Condition 2: The unit has not experienced a monotonically decreasing local health index for three consecutive monitoring cycles with a cumulative decrease exceeding the default value of 15%. Condition 3: The unit's system consistency score is not lower than the third threshold (a threshold dynamically adjusted by node degree, with a default base value of 0.65, adjusted between 0.55 and 0.70 based on the number of neighbors).

[0125] When a unit simultaneously meets all three of the above conditions, the fusion decision unit marks its status as "normal operation." For units in normal operation, no alarm information is issued, but the system will still package its local health index, system consistency score, and timestamp into a "normal status record" and store it in the cloud historical database for subsequent trend analysis, report generation, and threshold adaptive learning. At the same time, the visualization interface displays the unit in green or the default color, indicating that the status is normal.

[0126] When a unit fails to meet all the conditions for normal operation, i.e., at least one of the three abnormal conditions is triggered, the fusion decision unit judges according to a pre-set priority order. The priorities from highest to lowest are: first priority (severe individual degradation), second priority (systemic coordination anomaly), and third priority (early minor degradation). Once a high-priority judgment condition is met, the corresponding alarm is immediately output, and lower-priority conditions are no longer checked.

[0127] First priority: Severe individual degradation (red alert level).

[0128] Triggering condition: The unit's current local health index is below the critical alarm threshold (i.e., the first threshold, default 0.3). This condition is independent of any other factors and will trigger this level of alarm even if the system consistency score is normal or abnormal.

[0129] Alarm Content: The system immediately generates a red-level alarm with the title "Severe Degradation Alarm for Unit". The alarm text must include at least the following information: unit identification number, current local health index, and trigger timestamp. Simultaneously, the fusion decision unit extracts key information from the "early degradation evidence" data packet uploaded from edge nodes, appending the three highest-contributing feature dimensions, their actual values, and deviation directions to the alarm. Furthermore, the system automatically recommends possible fault types (such as "early bearing wear" or "pitch mechanism response lag"), derived by a lightweight rule engine in the cloud based on the combination of evidence features.

[0130] Response Action: A red alarm will flash on the visual interface, and the operations manager will be immediately notified via SMS, app push notification, or email. The system recommends arranging on-site inspection or remote shutdown maintenance as soon as possible (e.g., within 2 hours).

[0131] Second priority: Systemic coordination anomaly (yellow alert).

[0132] Triggering conditions: The unit's local health index is not lower than the critical alarm threshold (i.e., it does not meet the first priority), but the system consistency score is lower than the third threshold (the value after node degree correction), and an auxiliary verification condition is met: the average local health index of this unit and all its neighboring units is higher than 0.8 (i.e., the neighboring group itself is in normal condition). The purpose of the second condition is to exclude the situation of "low consistency score caused by large-scale failure of neighboring units" - if the neighboring units themselves have already experienced severe degradation, then the low consistency score may be caused by the neighboring units rather than system coordination issues. In this case, individual alarms of neighboring units should be handled first, and this unit should not trigger a yellow alarm alone.

[0133] Alarm Content: The system generates a yellow-level alarm with the title "Unit Systemic Inconsistency Risk Alarm". The alarm text includes: unit ID, its system consistency score, current local health index (still within the normal range), and a possible analysis of the reasons for the low consistency score. The cause analysis is based on the two components calculated in step five: if the health index deviation component contributes more, it indicates "the current health index deviates too much from the median of the neighboring group"; if the trend inconsistency component contributes more, it indicates "the health index trend deviates significantly from the neighboring group". In addition, the system will list the three units with the largest health index deviations among the unit's neighbors for maintenance personnel to trace.

[0134] Response Action: The yellow alarm is highlighted in yellow on the visualization interface, and a daily summary report is sent. It is recommended that maintenance personnel check the parameter configuration of the site controller, communication synchronization status, and whether there are any localized meteorological anomalies (e.g., gusts or shadows affecting only this unit). This alarm typically does not require immediate shutdown, but should be analyzed within 24 hours.

[0135] Third priority: Early slight degradation (blue level warning).

[0136] Triggering conditions: The unit does not meet the first and second priority levels (i.e., local health index ≥ critical alarm threshold and system consistency score ≥ third threshold), but meets the triggering condition for the second threshold—the local health index has monotonically decreased for three consecutive monitoring cycles and the cumulative decrease exceeds the default value of 15%. This indicates that although the unit's current health index is acceptable, it is experiencing a continuous and significant performance decline trend.

[0137] Alarm content: The system generates a blue-level alert with the title "Early Performance Degradation Alarm". The alert text includes: the unit identification number, the local health index sequence for the most recent three periods, and the three main deviation feature dimensions from the "Early Degradation Evidence" data packet generated in step four. The alert also provides a rough estimate of the "estimated remaining health time": calculated by linearly extrapolating the rate of decline over the past hour to determine the time required for the health index to drop to the critical alarm threshold (0.3).

[0138] Response Action: Blue alerts are indicated by a light blue icon on the visual interface and recorded in the alert log. The system will not proactively send SMS messages, but will summarize the alerts in the daily report. It is recommended that maintenance personnel conduct a dedicated inspection or online test of the unit within the next 3-7 days, focusing on the components identified by the evidence.

[0139] For units classified as "normally operating," the fusion decision unit performs the following operations: Generates a normal status record: Each record includes the unit identifier, current local health index, system consistency score, timestamp, and a brief qualitative description. This record is appended to the cloud database for long-term trend analysis and report generation. Updates the unit status panel: In the site panorama of the visualization interface, the unit is displayed in green or the default color. When the mouse hovers over it, an information card pops up displaying its latest health index and consistency score. Does not trigger any alarms or alerts: Normal operation status does not generate outgoing messages to keep maintenance personnel focused on abnormal units. Periodic report summaries: Even without alarms, the system still generates normal operation summary statistics hourly, daily, and weekly.

[0140] The second embodiment of this application is as follows:

[0141] Please see Figure 3 This invention provides a distributed generator set production process status monitoring system, applied to a distributed generator set production process status monitoring method as provided in the first embodiment, comprising:

[0142] Several edge computing nodes, each edge computing node being associated with a group of distributed generator sets, wherein each edge computing node contains:

[0143] The data acquisition and preprocessing unit is used to timestamp-align the multimodal parameter sequences of the distributed generator sets, and adaptively segment the continuous data stream into stationary and active segments based on the comparison results of the change rate of each parameter sequence with the stationary and active thresholds of dynamic changes, encapsulating them into basic monitoring data blocks. The feature extraction unit is used to extract statistical feature sub-vectors, trend feature sub-vectors, and coupling feature sub-vectors in parallel from each basic monitoring data block, and then concatenate and normalize these three to generate a unified operational feature vector. The model management unit is used to maintain a Gaussian mixture-based model for each unit at the edge. The adaptive state baseline model of the model is initially constructed using historical normal operation feature vectors and continuously evolved through a progressive incremental update mechanism based on sliding time windows and confidence level screening. The evaluation unit is used to calculate the negative log-likelihood probability based on the current operation feature vector and the adaptive state baseline model, and obtain the local health index after normalization mapping. When the local health index is lower than the first threshold or monotonically decreasing for three consecutive periods and the cumulative decrease exceeds the second threshold, the anomaly tracing process is initiated, and the three feature dimensions with the highest contribution and their actual values ​​are extracted and packaged into an early degradation evidence data package.

[0144] And a cloud analytics center, which communicates with each edge computing node via a network, and the cloud analytics center includes:

[0145] The collaborative analysis unit is used to pre-construct a static relationship map based on electrical connections, geographical proximity, and control domain correlation. For each unit, it calculates the weighted covariance of its local health index with the weighted median deviation and historical trend of neighboring units, and then integrates these to obtain a system consistency score. The fusion decision unit is used to determine three abnormal states—severe individual degradation, systemic coordination anomaly, and early minor degradation—in order of priority based on the local health index and system consistency score. It outputs the corresponding alarm level and evidence information, and records the normal status of units that meet the normal conditions. The edge computing node uploads the local health index and early degradation evidence data package to the cloud analysis center, which then distributes or displays the system consistency score and comprehensive status alarm results.

[0146] Regarding the system in the above embodiments, the specific ways in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0147] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0148] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for monitoring the production process status of a distributed generator set, characterized in that, Includes the following steps: The multimodal parameter sequences of the distributed generator set are timestamped and aligned. Based on the comparison results of the change rate of each parameter sequence with the stationary threshold and active threshold of dynamic change, the continuous data stream is adaptively divided into stationary segments and active segments and encapsulated into basic monitoring data blocks. For each basic monitoring data block, statistical feature sub-vectors, trend feature sub-vectors, and coupling feature sub-vectors are extracted in parallel, and the three are concatenated and normalized to generate a unified running feature vector; At the edge, an adaptive state baseline model based on a Gaussian mixture model is maintained for each unit. This model is initially constructed using historical normal operation feature vectors and continuously evolved through a progressive incremental update mechanism based on sliding time windows and confidence filtering. The negative log-likelihood probability is calculated based on the current running feature vector and the adaptive state baseline model. The local health index is obtained through normalization mapping. When the local health index is lower than the first threshold or decreases monotonically for three consecutive periods and the cumulative decrease exceeds the second threshold, the anomaly tracing process is initiated. The three feature dimensions with the highest contribution and their actual values ​​are extracted and packaged into an early degradation evidence data package. A static relationship graph is pre-constructed based on electrical connections, geographical proximity, and control domain relevance. For each unit, the weighted covariance of its local health index, weighted median deviation from the neighboring units in its domain, and historical trends are calculated and fused to obtain a system consistency score. The edge weights of the static relationship graph emphasize adjacency strength coefficients, with the direct electrical connection criterion having a weight of 1.0, the geographical proximity criterion having a weight that linearly decreases to 0.2 with distance, and the control domain relevance criterion having a weight of 0.

6. The health index deviation is calculated using the weighted median rather than the arithmetic mean, with the weight being the adjacency strength coefficient. The system consistency score is obtained by weighted fusion of the health index deviation component and the trend inconsistency component, with the health index deviation component having a weight of 0.6 and the trend inconsistency component having a weight of 0.

4. The base value of the third threshold is 0.65, and it is dynamically adjusted based on the node degree of the unit in the relationship graph: it is lowered to 0.55 when the degree is 1, and raised to 0.70 when the degree is greater than or equal to 5. A systemic operational misalignment risk is only confirmed when the value is below the adjusted threshold for two consecutive analysis periods. Based on the local health index and system consistency score, the system judges three abnormal states in order of priority: severe individual degradation, systemic coordination abnormality, and early slight degradation. The corresponding alarm level and evidence information are output, and the normal status of the units that meet the normal conditions is recorded.

2. The method for monitoring the production process status of a distributed generator set as described in claim 1, characterized in that, The stable threshold and the active threshold are dynamically changing reference values: the stable threshold is defined as the 90th percentile of the normal fluctuation rate of the unit's parameters under typical undisturbed operating conditions, and the active threshold is defined as the set percentage of the rate of change of the rated parameters per second or three times the standard deviation of normal operation. Both thresholds are recalculated weekly based on the historical operating characteristics of the unit and are supplemented by adaptive correction based on event frequency.

3. The method for monitoring the production process status of a distributed generator set as described in claim 1, characterized in that, Adaptively segmenting continuous data streams into stable and active segments, including: When the rate of change of any parameter sequence exceeds the active threshold, the sampling channels corresponding to that parameter sequence and all other parameter sequences that are physically coupled with it will automatically switch from a low sampling frequency to a high sampling frequency, and start a circular buffer window. The default range of the window is 5 seconds before the trigger point to 10 seconds after the trigger point. When multiple consecutive active events occur, the window is dynamically extended and merged into a complete active segment, cutting the continuous data stream into a stable segment and an active segment.

4. The method for monitoring the production process status of a distributed generator set as described in claim 1, characterized in that, The extraction of the trend feature sub-vectors adopts an adaptive trend fitting based on segment type: a one-time overall linear fitting is performed on the stable segment data block to obtain the first-order trend slope and the second-order acceleration; for the active segment data block, a variable-width sliding time window is used to automatically identify inflection points and divide them into multiple monotonic segments, and the segment with the largest absolute value of the rate of change among all segments is taken as the main trend feature, while the average trend feature and the number of inflection points are calculated at the same time.

5. The method for monitoring the production process status of a distributed generator set as described in claim 1, characterized in that, The extraction of the coupled feature sub-vectors includes: for each pair of predefined physical association parameters, calculating the average value and standard deviation of the sliding window correlation coefficient as an index of coupling degree and stability, and calculating the ratio of the change amplitude of the response parameter to the excitation parameter as the dynamic response ratio; the normalization is online rolling normalization, that is, performing zero-mean unit variance transformation based on the mean and standard deviation of each feature dimension of the most recent 100 normal data blocks.

6. The method for monitoring the production process status of a distributed generator set as described in claim 1, characterized in that, The adaptive state baseline model adopts a Gaussian mixture model with a diagonal covariance matrix approximation, and the number of Gaussian components is automatically determined by the Bayesian information criterion. The incremental update only uses the running feature vectors that are selected as stationary segments and whose local health index is not lower than the health threshold, and maintains a sliding time window of length 200. Only one expectation maximization iteration is performed each time the data changes within the window. After the incremental update, an independent validation set is used to evaluate the model fit. If the fit deteriorates, the model parameters before the update are rolled back.

7. The method for monitoring the production process status of a distributed generator set as described in claim 1, characterized in that, The normalized mapping of the local health index is based on a linear transformation of the 5th and 95th percentiles of the negative log-likelihood probability under the historical normal state of the unit, so that the health index is between 0 and 1. The first threshold is 0.3, and is temporarily adjusted to 0.2 within the first 100 hours of unit commissioning; the default cumulative reduction in the second threshold is 15%. The anomaly tracing process first determines the principal Gaussian component to which the current running feature vector belongs. Then, it replaces the value of each dimension with the mean of the corresponding dimension of the principal Gaussian component and calculates the change in the negative log-likelihood probability to obtain the deviation contribution of each dimension. After sorting, the three dimensions with the highest contribution, along with their measured values ​​and deviation directions, are packaged into an early degradation evidence data package.

8. The method for monitoring the production process status of a distributed generator set as described in claim 1, characterized in that, Priority judgment adopts a hit-and-stop strategy: the first priority is severe individual degradation, triggered when the local health index falls below the first threshold, outputting a red alarm with early degradation evidence; the second priority is systemic coordination anomaly, triggered when the local health index is not lower than the first threshold, the system consistency score is lower than the third threshold, and the average local health index of the unit and all its neighbors is greater than 0.8, outputting a yellow alarm; the third priority is early minor degradation, triggered when the first two priorities are not met but the monotonically decreasing condition of the second threshold with a cumulative decrease is met, outputting a blue warning with early degradation evidence and estimated remaining health time; when none of the above priority conditions are met simultaneously, it is judged as normal operation, the normal status record is recorded and a green icon is displayed.

9. A distributed generator set production process status monitoring system, used to implement the distributed generator set production process status monitoring method as described in claim 1, characterized in that, include: Several edge computing nodes, each edge computing node being associated with a group of distributed generator sets, wherein each edge computing node contains: The data acquisition and preprocessing unit is used to timestamp the multimodal parameter sequences of the distributed generator set, and adaptively divide the continuous data stream into stable segments and active segments based on the comparison results of the change rate of each parameter sequence with the dynamic change stationary threshold and active threshold, and encapsulate them into basic monitoring data blocks. The feature extraction unit is used to extract statistical feature sub-vectors, trend feature sub-vectors, and coupling feature sub-vectors in parallel for each basic monitoring data block, and then concatenate and normalize the three to generate a unified running feature vector; The model management unit is used to maintain an adaptive state baseline model based on a Gaussian mixture model for each unit at the edge. The model is initially constructed using historical normal operation feature vectors and continuously evolves through a progressive incremental update mechanism based on sliding time windows and confidence filtering. The evaluation unit is used to calculate the negative log-likelihood probability based on the current running feature vector and the adaptive state baseline model. After normalization mapping, the local health index is obtained. When the local health index is lower than the first threshold or monotonically decreasing for three consecutive periods and the cumulative decrease exceeds the second threshold, the anomaly tracing process is initiated. The three feature dimensions with the highest contribution and their actual values ​​are extracted and packaged into an early degradation evidence data package. And a cloud analytics center, which communicates with each edge computing node via a network, and the cloud analytics center includes: The collaborative analysis unit is used to pre-build a static relationship map based on electrical connections, geographical proximity and control domain correlation, and calculate the weighted covariance of the local health index of each unit with the weighted median deviation of neighboring units in the neighborhood and the historical trend, and integrate them to obtain a system consistency score. The fusion decision unit is used to determine three abnormal states in order of priority: severe individual degradation, systemic coordination abnormality, and early minor degradation, based on the local health index and system consistency score. It outputs the corresponding alarm level and evidence information, and records the normal status of units that meet the normal conditions. The edge computing node uploads local health index and early degradation evidence data packets to the cloud analysis center, which then distributes or displays the system consistency score and comprehensive status alarm results.

Images