Method for monitoring input and output voltage and current of ac-dc power distribution system of data center

Abstract

The present application relates to the field of power electronics and power supply and distribution technology, and discloses a monitoring method for input and output voltage and current of an AC / DC power distribution system in a data center, which comprises: deploying intelligent sensing units with edge computing capabilities at each monitoring node to sample at a basic frequency and extract real-time characteristics such as effective value, harmonic distortion rate and differential change rate of voltage and current; determining whether the system is in a steady state, quasi-instantaneous state or strong instantaneous state according to multi-level dynamic threshold values, and performing feature abstract upload, local high-frequency backtracking or collaborative full-waveform synchronous acquisition respectively; storing data in a central server and constructing a state evolution model to dynamically optimize threshold parameters. The system comprises intelligent sensing units, a central monitoring server and an industrial Ethernet link. The present application significantly reduces steady-state data transmission volume, completely captures transient details and realizes adaptive adjustment of monitoring sensitivity.

Description

Technical Field

[0001] This invention belongs to the field of power electronics and power supply and distribution technology, specifically relating to a method for monitoring the input and output voltage and current of AC / DC power distribution systems in data centers. Background Technology

[0002] With the continuous expansion of data center scale and the rapid growth of computing power demand, the reliability, energy efficiency, and intelligent operation and maintenance level of AC / DC power distribution systems have become crucial to ensuring business continuity. Modern data center power distribution architectures generally adopt high-density, multi-level AC / DC hybrid power supply modes, and their operational status highly depends on real-time monitoring of electrical parameters such as input and output voltage and current. Traditional monitoring methods typically employ fixed high-frequency sampling strategies to ensure the capture of potential abnormal events under any operating condition. However, such methods continuously generate a large amount of redundant data when the system is in a long-term steady-state operation—statistics show that over 99% of the sampled data reflects stable waveforms with no significant changes, not only consuming significant network bandwidth, storage space, and edge computing resources, but also exacerbating the power consumption and heat dissipation burden on terminal devices.

[0003] Intelligent sensing and adaptive acquisition of the dynamic characteristics of voltage and current have become a core direction for improving the efficiency of power distribution monitoring. This technology aims to dynamically adjust sampling strategies by analyzing the time and frequency domain characteristics of electrical signals, ensuring complete recording of critical events while minimizing the generation of invalid data. Ideally, the solution should automatically reduce the sampling frequency and quantization accuracy in the steady-state range, and combine this with efficient compression algorithms to reduce transmission load; while upon detecting transient events such as abrupt edges, harmonic distortion, or voltage dips, it should quickly switch to a high-fidelity full-rate recording mode to completely capture the waveforms before and after the event.

[0004] A simple threshold triggering mechanism has been introduced, but its sampling strategy switching logic is rigid, lacks a deep understanding of waveform morphology, and struggles to distinguish between noise interference and genuine fault characteristics, easily leading to false triggers or missed detections. Furthermore, existing solutions generally lack a correlation model between steady-state and transient data, failing to utilize differential information for efficient compression, resulting in the need to transmit large amounts of raw data even during periods of low activity. In addition, in multi-channel synchronous monitoring scenarios, independent decision-making by each channel further exacerbates resource fragmentation, making global optimization difficult. Summary of the Invention

[0005] This invention provides a method for monitoring the input and output voltage and current of AC / DC power distribution systems in data centers. It aims to address the problem in existing technologies where network bandwidth, storage space, and computing resources are occupied by a large amount of invalid steady-state data due to the use of fixed high-frequency sampling strategies. Simultaneously, it overcomes the technical deficiency of losing critical dynamic details due to insufficient sampling rate during transient events. This invention achieves high-fidelity, low-redundancy, and high-efficiency monitoring of power distribution system electrical parameters by constructing a multi-level dynamic sampling and data compression architecture based on an adaptive event-triggered mechanism.

[0006] This invention provides a method for monitoring the input and output voltage and current of an AC / DC power distribution system in a data center, comprising:

[0007] Intelligent sensing units with local edge computing capabilities are deployed at each monitoring node of the power distribution system. The intelligent sensing unit is equipped with an analog-to-digital converter, a local buffer, and an embedded processing core. The intelligent sensing unit continuously collects voltage and current signals of the power distribution line at a basic sampling frequency and temporarily stores the raw sampling data in the local buffer.

[0008] Real-time feature extraction is performed on the original sampled data to calculate the effective voltage value, effective current value, power factor, harmonic distortion rate, and first-order derivative rate of change within the current time window; the feature extraction results are compared with preset multi-level dynamic thresholds to determine whether the current system operating state is in steady state, quasi-transient state, or strong transient state.

[0009] If the state is determined to be steady, only the compressed and encoded feature summary data is retained and the original sampled data is discarded. At the same time, the sampling data upload cycle is extended to the first preset interval. If the state is determined to be quasi-transient, a local high-frequency backtracking mechanism is activated to extract the original sampled data within the second preset duration before the event occurs from the local cache. Data is then collected for the third preset duration after the event occurs at the second preset sampling frequency to form a complete transient segment before uploading. If the state is determined to be strong transient, all contents of the current local cache are immediately frozen, and the collaborative sampling instructions of adjacent nodes are triggered synchronously. All associated nodes are forced to record the full waveform data for the fourth preset duration at the highest sampling frequency, and then upload it to the central monitoring server after attaching a timestamp tag.

[0010] After receiving feature summary data or transient segment data from each node, the central monitoring server stores the data in a structured database or a time-series waveform database according to the data type. Based on historical data, it constructs an evolution model of the power distribution system's operating status to dynamically optimize the multi-level dynamic threshold parameters of each node.

[0011] In one embodiment of the present invention, the basic sampling frequency is set to 2000 times per second, the first preset interval is 10 seconds, the second preset duration is 500 milliseconds, the third preset duration is 1.5 seconds, the fourth preset duration is 2 seconds, and the maximum sampling frequency is 200,000 times per second.

[0012] In one embodiment of the present invention, the multi-level dynamic threshold includes a steady-state threshold group, a quasi-transient threshold group, and a strong transient threshold group. The steady-state threshold group is defined as follows: the effective voltage fluctuation is less than 0.5% of the rated voltage, the effective current fluctuation is less than 1% of the rated current, and the absolute value of the first-order derivative rate of change is less than 0.1 volts per millisecond or 0.2 amperes per millisecond. The quasi-transient threshold group is defined as follows: the effective voltage fluctuation is between 0.5% and 3% of the rated voltage, or the effective current fluctuation is between 1% and 5% of the rated current, or the absolute value of the first-order derivative rate of change is between 0.1 volts per millisecond and 1 volt per millisecond or 0.2 amperes per millisecond and 2 amperes per millisecond. The strong transient threshold group is defined as follows: the effective voltage fluctuation is greater than 3% of the rated voltage, or the effective current fluctuation is greater than 5% of the rated current, or the absolute value of the first-order derivative rate of change is greater than 1 volt per millisecond or 2 amperes per millisecond, or the harmonic distortion rate suddenly increases by more than 10%.

[0013] As one embodiment of the present invention, the feature summary data is compressed using differential coding. Specifically, based on the most recently uploaded steady-state feature value, only the difference between the current feature value and the reference value is recorded, and the difference is quantized by integer and Huffman encoded to generate a compact data packet with a length not exceeding 16 bytes.

[0014] As one embodiment of the present invention, the implementation of the local high-frequency backtracking mechanism relies on the local buffer adopting a circular double buffer structure, wherein the main buffer continuously writes the latest sampled data, and the auxiliary buffer immediately locks and switches to read-only state when a quasi-transient event is detected, ensuring that the data before the event is not overwritten. At the same time, the main buffer continues to write the data after the event at a second preset sampling frequency until the third preset duration ends.

[0015] In one embodiment of the present invention, the collaborative sampling instruction is transmitted through a dedicated low-latency control bus. This bus adopts a time-triggered protocol to ensure that the instruction is delivered to all associated nodes within 100 microseconds, and the start time of the highest sampling frequency of each node is aligned with the global synchronization clock source, with a time synchronization accuracy better than 1 microsecond.

[0016] As one embodiment of the present invention, the power distribution system operation status evolution model constructed by the central monitoring server is a time series prediction model based on long short-term memory network. Its input is the feature summary data sequence uploaded by all nodes in the past 24 hours, and its output is the probability distribution of quasi-transient or strong transient events occurring at each node in the next five minutes. This probability distribution is used to dynamically adjust the quasi-transient and strong transient judgment boundary in the multi-level dynamic threshold of each node, with an adjustment step size of 5% to 15% of the original threshold.

[0017] As one embodiment of the present invention, the embedded processing core of the intelligent sensing unit is a microcontroller with a floating-point arithmetic unit, a main frequency of not less than 400 MHz, a local buffer capacity of not less than 8 megabytes, an analog-to-digital converter resolution of not less than 16 bits, and a sampling channel isolation withstand voltage of not less than 5000 volts.

[0018] This invention provides a monitoring system for the input and output voltage and current of an AC / DC power distribution system in a data center. The system includes: multiple intelligent sensing units deployed at the input and output ports of a power distribution cabinet, each intelligent sensing unit comprising an analog-to-digital converter, a local buffer, and an embedded processing core; a central monitoring server configured with a structured database, a time-series waveform database, and a state evolution model engine; and an industrial Ethernet communication link connecting the intelligent sensing units and the central monitoring server. The intelligent sensing units are used to perform local data acquisition, feature extraction, state determination, adaptive sampling control, and data uploading operations as described in the aforementioned monitoring method. The central monitoring server is used to receive, classify, and store data, run the state evolution model, and issue updated multi-level dynamic threshold parameters to each intelligent sensing unit.

[0019] As one embodiment of the present invention, the industrial Ethernet communication link adopts the Gigabit Ethernet physical layer, supports time-sensitive networking protocols, ensures that the transmission priority of transient fragment data is higher than that of feature summary data, and the maximum end-to-end transmission delay does not exceed five milliseconds.

[0020] As one embodiment of the present invention, the state evolution model engine performs model retraining at 2:00 AM every day. The training dataset contains all labeled transient event fragments and their corresponding preceding feature summary sequences within the past seven days. The model loss function adopts weighted cross-entropy, wherein the weight coefficient of strong transient event samples is set to 10, the weight coefficient of quasi-transient event samples is set to 3, and the weight coefficient of steady-state samples is set to 1.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0022] 1. This invention completely abandons the traditional fixed sampling rate monitoring mode by introducing a state determination mechanism based on real-time feature extraction and multi-level dynamic thresholds, and achieves accurate matching between sampling behavior and the actual operating state of the system.

[0023] 2. When the system is in a steady state for an extended period, only highly compressed feature summary data is uploaded, reducing the data transmission volume to less than 0.5% of that of traditional high-frequency sampling schemes, significantly alleviating network load and storage pressure. When the system experiences abnormal disturbances, it can fully capture the waveform details of transient events through local cache backtracking and collaborative high-frequency sampling mechanisms, achieving a time resolution of five microseconds, ensuring that key fault characteristics are not missed.

[0024] 3. The central monitoring server achieves closed-loop optimization of threshold parameters by constructing a state evolution model, enabling the system to possess self-learning and adaptive capabilities. Monitoring sensitivity dynamically adjusts with changes in the operating environment, effectively improving the identification rate of new or complex power quality problems. The overall solution achieves optimal allocation of computing, storage, and network resources while ensuring monitoring accuracy. Attached Figure Description

[0025] Fig. 1 This is a schematic diagram of the overall technical solution architecture of the present invention;

[0026] Fig. 2 This is a schematic diagram of the core principle framework of the multi-level dynamic sampling and data compression architecture based on the adaptive event triggering mechanism in this invention;

[0027] Fig. 3 This is a flowchart illustrating the logical process of the intelligent sensing unit performing local data acquisition, feature extraction, and state determination in this invention.

[0028] Fig. 4 This is a logical flowchart of the adaptive sampling and data uploading strategy under the three operating states of steady state, quasi-transient state and strong transient state in this invention;

[0029] Fig. 5 This is a schematic diagram of the multi-level interaction relationship and data flow between the intelligent sensing unit and the central monitoring server in this invention;

[0030] Fig. 6 This is a flowchart illustrating the logical flow of the central monitoring server in this invention, which constructs a power distribution system operation status evolution model and implements closed-loop optimization of threshold parameters. Detailed Implementation

[0031] Please refer to Figs. 1 to 6This invention provides a method for monitoring the input and output voltage and current of a data center AC / DC power distribution system. Its core lies in deploying intelligent sensing units with local edge computing capabilities to achieve multi-level dynamic sampling and data compression processing at each power distribution node based on an adaptive event triggering mechanism. This significantly reduces data redundancy during steady-state operation while ensuring high-fidelity transient capture capabilities. The method strictly follows a preset S-step process, including operation stages S1, S2, S3, S4, S5, S6, and S7, with a closed-loop feedback and collaborative linkage mechanism between each stage.

[0032] Step S1 involves deploying intelligent sensing units with local edge computing capabilities at each monitoring node of the power distribution system. Each intelligent sensing unit is equipped with an analog-to-digital converter, a local buffer, and an embedded processing core. Each intelligent sensing unit is installed at an input or output port of the data center power distribution cabinet, directly connected to the voltage and current signal loop of the line under test. The analog-to-digital converter adopts a 16-bit resolution architecture, and its analog input channel is designed with 5000V isolation withstand voltage to ensure safe acquisition of raw electrical parameter signals in a high-voltage AC / DC mixed environment.

[0033] The embedded processing core is a microcontroller with a main frequency of no less than 400 MHz, and a built-in floating-point unit for executing subsequent real-time feature extraction and status determination logic. The local buffer capacity is no less than 8 megabytes, employing a circular double-buffered structure. The main buffer continuously receives the latest sampled data stream, while the secondary buffer is locked under specific event triggering conditions to retain historical data fragments. All intelligent sensing units are connected to the central monitoring server via a gigabit industrial Ethernet link. This link supports time-sensitive networking protocols to ensure low-latency transmission of high-priority transient data.

[0034] Step S2 involves continuously acquiring voltage and current signals of the power distribution line at a basic sampling frequency using the intelligent sensing unit, and temporarily storing the raw sampling data in the local buffer. The basic sampling frequency is set to 2000 times per second, i.e., a sampling period of 500 microseconds. At this frequency, the analog-to-digital converter synchronously performs digital conversion on the voltage and current channels, generating raw sampling data frames containing timestamps, channel identifiers, and numerical amplitudes.

[0035] Each frame of data is immediately written to the end of the main buffer in the local buffer. When the main buffer is full, the oldest data is automatically overwritten, maintaining a first-in, first-out sliding window mechanism. This basic sampling frequency is sufficient to cover the power frequency and its main harmonic components, while avoiding unnecessary data redundancy due to oversampling. All raw sampled data only resides temporarily in the local buffer and must not be uploaded to the central monitoring server without further processing.

[0036] Step S3 involves performing real-time feature extraction on the original sampled data, calculating the RMS voltage, RMS current, power factor, harmonic distortion rate, and first-order differential rate of change within the current time window. The feature extraction operation is initiated immediately after each basic sampling cycle by the embedded processing core. The current time window is defined as the set of continuous sampling points within the most recent 100 milliseconds, corresponding to 200 sampling points. The RMS voltage is calculated using the root mean square formula, and the RMS current is calculated similarly.

[0037] The power factor is determined by the cosine of the phase difference between voltage and current, which is solved in the time domain using a cross-correlation algorithm. The harmonic distortion rate is calculated by performing a fast Fourier transform on the signal within the window to determine the ratio of total harmonic content to the fundamental amplitude. The first-order differential rate of change is obtained by subtracting the voltage or current values ​​from two adjacent sampling points and dividing by the sampling period, denoted as dV / dt and dI / dt, respectively. All characteristic values ​​are stored in local registers in floating-point format for subsequent state determination.

[0038] Step S4 involves comparing the feature extraction results with preset multi-level dynamic thresholds to determine whether the current system operating state is in a steady state, quasi-transient state, or strong transient state. The multi-level dynamic thresholds comprise three levels: a steady-state threshold group, a quasi-transient threshold group, and a strong transient threshold group. The steady-state threshold group is defined as follows: the effective voltage fluctuation is less than 0.5% of the rated voltage, the effective current fluctuation is less than 1% of the rated current, and the absolute value of the first-order derivative rate of change is less than 0.1 volts per millisecond or 0.2 amperes per millisecond. If all feature values ​​meet the steady-state threshold group conditions, the system is determined to be in a steady state.

[0039] The quasi-transient threshold group is defined as follows: the fluctuation range of the effective voltage value is between 0.5% and 3% of the rated voltage, or the fluctuation range of the effective current value is between 1% and 5% of the rated current, or the absolute value of the first derivative rate of change is between 0.1 volts per millisecond and 1 volt per millisecond or between 0.2 amperes per millisecond and 2 amperes per millisecond.

[0040] If any characteristic value falls within the quasi-transient threshold group and does not trigger a strong transient condition, it is determined to be a quasi-transient. The strong transient threshold group is defined as follows: the effective voltage fluctuation is greater than 3% of the rated voltage, or the effective current fluctuation is greater than 5% of the rated current, or the absolute value of the first derivative rate of change is greater than 1 volt per millisecond or 2 amperes per millisecond, or the harmonic distortion rate suddenly increases by more than 10%. If any characteristic value meets the strong transient threshold group conditions, it is immediately determined to be a strong transient. The state determination logic adopts a hierarchical priority strategy, with strong transient determination having the highest priority, followed by quasi-transient, and finally steady state.

[0041] Step S5 is as follows: If the state is determined to be steady, only the compressed feature summary data is retained and the original sampled data is discarded. Simultaneously, the sampling data upload cycle is extended to a first preset interval. The feature summary data is compressed using differential coding, specifically as follows: Using the most recently successfully uploaded steady-state feature value to the central monitoring server as the reference vector, the difference between the current feature value and each dimension of the reference vector is calculated; each difference is integer-quantized, with the quantization step size set according to the physical dimensions and typical fluctuation range of each feature; the quantized integer sequence is Huffman-coded to generate variable-length codewords, and finally packaged into a compact data packet with a length not exceeding 16 bytes. This data packet contains a checksum, timestamp, and compressed feature differences, and is uploaded via an industrial Ethernet link with the lowest priority. The original sampled data is immediately cleared from the local cache after feature extraction is completed, without occupying persistent storage space. The upload cycle is set to 10 seconds, meaning that every ten basic sampling windows are merged to generate a feature summary data packet and uploaded once.

[0042] Step S6 is as follows: If the event is determined to be quasi-transient, a local high-frequency backtracking mechanism is activated to extract the original sampled data within a second preset duration before the event occurs from the local cache. Data is then collected for a third preset duration after the event occurs at a second preset sampling frequency, forming a complete transient segment before uploading. The second preset duration is 500 milliseconds, corresponding to 1000 basic sampling points. Because the local cache uses a circular double-buffered structure, when a quasi-transient event is detected, the embedded processing core immediately locks the secondary buffer, making its contents read-only to ensure that the data within the first 500 milliseconds of the event is not overwritten by new sampling. Simultaneously, the main buffer switches to high-frequency sampling mode, increasing the sampling frequency to 20,000 times per second, i.e., shortening the sampling period to 50 microseconds, continuously collecting data for 1.5 seconds after the event occurs, for a total of 30,000 sampling points.

[0043] During high-frequency sampling, the analog-to-digital converter operates at maximum throughput, with all channels acquiring data synchronously. A complete transient segment is composed of the base sampling data from 500 milliseconds prior to the event and the high-frequency sampling data from 1.5 seconds after the event. After adding a globally unique event identifier and a precise timestamp, it is encapsulated into a high-priority data packet and uploaded to the central monitoring server. Once uploaded, the local buffer reverts to base sampling mode, and the secondary buffer is unlocked and rejoins the circular write operation.

[0044] Step S7 is as follows: If a strong transient event is determined, the entire contents of the current local buffer are immediately frozen, and the cooperative sampling instructions of adjacent nodes are triggered synchronously. This forces all associated nodes to synchronously record the full waveform data for a fourth preset duration at the highest sampling frequency, and uploads it to the central monitoring server after attaching a timestamp. After the strong transient event is triggered, the intelligent sensing unit immediately marks the entire contents of the main buffer and the auxiliary buffer as unwrapable, retaining all the original sampled data within the time window covered by at least 8 megabytes before the event occurred. At the same time, the node broadcasts the cooperative sampling instructions to all other intelligent sensing units directly connected to the physical topology via a dedicated low-latency control bus. This control bus uses a time-triggered protocol, and the instruction transmission delay is strictly controlled within 100 microseconds.

[0045] Upon receiving the instruction, adjacent nodes immediately interrupt their current sampling task, forcibly increasing their sampling frequency to 200,000 times per second (i.e., a sampling period of five microseconds), and simultaneously initiate a fourth, preset two-second full waveform recording. All nodes participating in collaborative sampling align their sampling start times based on the same global synchronization clock source, with a time synchronization accuracy better than 1 microsecond, ensuring strict alignment of multi-node waveform data on the time axis. After recording is complete, each node uploads a data packet containing the complete two-second waveform, node identifier, event type, and nanosecond-level timestamp to the central monitoring server with the highest network priority.

[0046] After receiving data from each intelligent sensing unit, the central monitoring server classifies, stores, and processes the data according to its data type. Feature summary data is parsed and stored in a structured database, indexed by time, node, and feature dimensions; transient segment data is stored in a dedicated time-series waveform database, supporting millisecond-level time range queries and waveform visualization. The server starts the state evolution model engine at 2 AM daily, retraining the Long Short-Term Memory network using all labeled transient event segments from the past seven days and their corresponding preceding feature summary sequences.

[0047] The model takes as input the feature summary data sequence uploaded by each node over the past 24 hours and outputs the probability distribution of quasi-transient or strong transient events occurring at each node within the next five minutes. During training, a weighted cross-entropy loss function is used, with a weight coefficient of 10 for strong transient event samples, 3 for quasi-transient event samples, and 1 for steady-state samples to address class imbalance. After training, the model dynamically adjusts the multi-level dynamic threshold parameters for each node based on the predicted probabilities. Specifically, if the probability of a strong transient event occurring at a node within the next five minutes exceeds 30%, the voltage fluctuation amplitude boundary in its strong transient threshold is lowered from 3% to 2.55% (85% of the original value); if the probability of a quasi-transient event exceeds 50%, the upper limit of the current fluctuation amplitude in the quasi-transient threshold is lowered from 5% to 4.25% (85% of the original value). The adjusted threshold parameters are then transmitted to the corresponding intelligent sensing unit via an industrial Ethernet link, completing closed-loop optimization.

[0048] The above method, through the coordinated execution of steps S1 to S7, constructs a resource-efficient, responsive, and self-learning power distribution monitoring system. During steady-state operation, the data upload volume is reduced to less than five-thousandths of that of traditional fixed high-frequency sampling schemes. In the event of transient events, the system can capture the full waveform details with a time resolution of five microseconds and ensure the accuracy of fault location through multi-node collaboration. The entire solution achieves optimal allocation of computing, storage, and network resources without sacrificing monitoring accuracy, making it suitable for long-term reliable operation in complex AC / DC power distribution environments of large data centers.

Claims

1. A method for monitoring the input and output voltage and current of an AC / DC power distribution system in a data center, characterized in that, include: Intelligent sensing units with local edge computing capabilities are deployed at each monitoring node of the power distribution system. The intelligent sensing unit is equipped with an analog-to-digital converter, a local cache, and an embedded processing core. The intelligent sensing unit continuously collects voltage and current signals of the power distribution line at a basic sampling frequency and temporarily stores the original sampling data in the local buffer. Real-time feature extraction is performed on the original sampled data to calculate the effective voltage value, effective current value, power factor, harmonic distortion rate, and first-order differential rate of change within the current time window; The results of the feature extraction are compared with preset multi-level dynamic thresholds to determine whether the current system operating state is in a steady state, quasi-transient state, or strong transient state. If the state is determined to be steady, only the compressed and encoded feature summary data will be retained and the original sampled data will be discarded. At the same time, the sampling data upload cycle will be extended to the first preset interval. If it is determined to be a quasi-transient, a local high-frequency backtracking mechanism is activated to extract the original sampling data within the second preset duration before the event occurs from the local cache, and to continue to collect data for the third preset duration after the event occurs at the second preset sampling frequency, and then upload the complete transient segment. If it is determined to be a strong transient, the entire contents of the current local buffer are immediately frozen, and the collaborative sampling instructions of adjacent nodes are triggered synchronously. All associated nodes are forced to record the full waveform data of the fourth preset duration at the highest sampling frequency, and upload it to the central monitoring server after attaching a timestamp tag. After receiving feature summary data or transient segment data from each node, the central monitoring server stores the data in a structured database or a time-series waveform database according to the data type. Based on historical data, it constructs an evolution model of the power distribution system's operating status to dynamically optimize the multi-level dynamic threshold parameters of each node.

2. The method for monitoring the input and output voltage and current of a data center AC / DC power distribution system according to claim 1, characterized in that, Real-time feature extraction is performed on the raw sampled data to calculate the effective voltage value, effective current value, power factor, harmonic distortion rate, and first-order differential rate of change within the current time window, including: The continuous sampling points are used as the current time window, corresponding to the sampling points; The effective values ​​of voltage and current are calculated using the root mean square formula. The power factor is obtained by solving the phase difference between voltage and current in the time domain using a cross-correlation algorithm and calculating the cosine value. Fast Fourier Transform is used to perform spectral analysis on the signal within the window, and the ratio of total harmonic content to fundamental amplitude is calculated to obtain the harmonic distortion rate. The first derivative rates of change of voltage and current are obtained by subtracting the voltage or current values ​​of two adjacent sampling points and dividing by the sampling period.

3. The method for monitoring the input and output voltage and current of a data center AC / DC power distribution system according to claim 2, characterized in that, The results of the feature extraction are compared with preset multi-level dynamic thresholds to determine whether the current system operating state is in a steady state, quasi-transient state, or strong transient state, including: If the fluctuation range of the effective voltage value is less than 0.5% of the rated voltage, the fluctuation range of the effective current value is less than 1% of the rated current, and the absolute value of the first derivative rate of change is less than 0.1 volts per millisecond or 0.2 amperes per millisecond, then it is determined to be in a steady state. If the fluctuation range of the effective voltage value is between 0.5% and 3% of the rated voltage, or the fluctuation range of the effective current value is between 1% and 5% of the rated current, or the absolute value of the first derivative rate of change is between 0.1 volts per millisecond and 1 volt per millisecond or between 0.2 amperes per millisecond and 2 amperes per millisecond, it is determined to be a quasi-transient. If the effective voltage fluctuation is greater than 3% of the rated voltage, or the effective current fluctuation is greater than 5% of the rated current, or the absolute value of the first derivative rate of change is greater than 1 volt per millisecond or 2 amperes per millisecond, or the harmonic distortion rate suddenly increases by more than 10%, it is judged as a strong transient.

4. The method for monitoring the input and output voltage and current of a data center AC / DC power distribution system according to claim 3, characterized in that, If the state is determined to be steady, only the compressed and encoded feature summary data will be retained and the original sampled data will be discarded. Simultaneously, the sampled data upload cycle will be extended to a first preset interval, including: Using the most recently successfully uploaded steady-state feature value as the baseline vector, calculate the difference between the current feature value and each dimension of the baseline vector; Each difference is quantized to an integer, and the quantized integer sequence is Huffman encoded to generate a compact data packet with a length not exceeding 16 bytes. The compact data packet is uploaded via an industrial Ethernet link with the lowest priority, and the original sampled data in the local cache is cleared.

5. The method for monitoring the input and output voltage and current of a data center AC / DC power distribution system according to claim 4, characterized in that, If determined to be quasi-transient, a local high-frequency backtracking mechanism is activated to extract the original sampling data within a second preset duration before the event occurs from the local cache, and to continue collecting data for a third preset duration after the event occurs at a second preset sampling frequency, forming a complete transient segment before uploading, including: When a quasi-transient event is detected, the secondary buffer of the local buffer is locked and made read-only, retaining the basic sampled data 500 milliseconds prior to the event. Switch the main buffer to high-frequency sampling mode to continuously collect data for 1.5 seconds after the event occurs at a sampling frequency of 20,000 times per second. The basic sampling data from 500 milliseconds before the event and the high-frequency sampling data from 1.5 seconds after the event are concatenated to form a complete transient fragment. After adding a globally unique event identifier and a precise timestamp, it is encapsulated into a high-priority data packet and uploaded.

6. The method for monitoring the input and output voltage and current of a data center AC / DC power distribution system according to claim 5, characterized in that, If a strong transient condition is identified, the entire contents of the current local buffer are immediately frozen, and the collaborative sampling command of adjacent nodes is triggered synchronously. This forces all associated nodes to synchronously record the full waveform data for the fourth preset duration at the highest sampling frequency, and upload it to the central monitoring server after attaching a timestamp tag. This includes: Mark all contents of the primary and secondary buffers of the local cache as unwriteable. Cooperative sampling instructions are broadcast to all other smart sensing units directly connected to the physical topology via a dedicated low-latency control bus. Upon receiving the instruction, the adjacent nodes forcibly increase their sampling frequency to 200,000 times per second and align their sampling start times according to the same global synchronization clock source; Each node synchronously records two seconds of full waveform data, and uploads it with the highest network priority after attaching the node identifier, event type, and nanosecond-level timestamp.

7. The method for monitoring input and output voltage and current of a data center AC / DC power distribution system according to claim 6, characterized in that, After receiving feature summary data or transient segment data from each node, the central monitoring server stores the data in a structured database or a time-series waveform database according to the data type. Based on historical data, it constructs an evolution model of the power distribution system's operating state to dynamically optimize multi-level dynamic threshold parameters for each node, including: After parsing the feature summary data, it is stored in a structured database and indexed by time, node, and feature dimensions. Store transient segment data into a time-series waveform database, supporting millisecond-level time range query and waveform visualization; The Long Short-Term Memory Network was retrained using all labeled transient event fragments from the past seven days and their corresponding preceding feature summary sequences. Using the feature summary data sequence uploaded by each node in the past 24 hours as input, output the probability distribution of quasi-transient or strong transient events occurring on each node in the next five minutes; The multi-level dynamic threshold parameters of each node are dynamically adjusted according to the probability distribution, with the adjustment step size being 5% to 15% of the original threshold.

8. The method for monitoring the input and output voltage and current of a data center AC / DC power distribution system according to claim 7, characterized in that, The loss function of the Long Short-Term Memory Network is weighted cross-entropy, where the weight coefficient of strong transient event samples is set to 10, the weight coefficient of quasi-transient event samples is set to 3, and the weight coefficient of steady-state samples is set to 1.

9. The method for monitoring the input and output voltage and current of a data center AC / DC power distribution system according to claim 1, characterized in that, The basic sampling frequency is set to 2000 times per second, the maximum sampling frequency is 200,000 times per second, the resolution of the analog-to-digital converter is not less than 16 bits, and the isolation withstand voltage of the sampling channel is not less than 5000 volts.

10. The method for monitoring the input and output voltage and current of a data center AC / DC power distribution system according to claim 1, characterized in that, The local buffer adopts a circular double buffer structure. The main buffer continuously writes the latest sampled data, and the secondary buffer is immediately locked and switched to read-only state when a quasi-transient event is detected.

Images