Frequency converter abnormal state detection method based on time sequence data analysis

Abstract

The invention discloses a frequency converter abnormal state detection method based on time series data analysis, which belongs to the technical field of electric power, and comprises the steps of: finally generating a structured abnormal event through acquisition, preprocessing, feature extraction, anomaly detection, anomaly level judgment and trend analysis of data of a multi-source sensor; accurate prediction and early warning of the operation state of a complex system are achieved, the technical problems of high real-time performance, low data redundancy and multi-source feature fusion of abnormal monitoring of the frequency converter are solved, redundant data are reduced, the data volume is reduced, the communication bandwidth pressure is reduced, it is ensured that high-frequency, low-frequency and event data are aligned on the same time axis, and the accuracy of abnormal monitoring of the frequency converter is improved. Abnormal data are transmitted preferentially, abnormal information is guaranteed to be complete and reliable, suspicious anomalies are screened rapidly, local and remote alarms are generated according to the abnormal levels, an abnormal database and a periodic summary report are stored, and operation and maintenance and remote monitoring are facilitated.

Description

Technical Field

[0001] This invention belongs to the field of power technology, and in particular relates to a method for detecting abnormal states of frequency converters based on time-series data analysis. Background Technology

[0002] With the development of industrial automation and intelligent manufacturing, frequency converters are widely used in motor drive systems to regulate motor speed and save energy. However, during long-term operation, frequency converters may experience abnormal states such as overvoltage, overcurrent, overheating, abnormal motor vibration, or protection activation. If these abnormalities are not detected in time, they may lead to equipment damage, production shutdowns, or even safety accidents.

[0003] Existing methods for monitoring frequency converter anomalies mainly rely on the following approaches:

[0004] Single threshold alarm: This method sets a fixed threshold for voltage, current, or temperature, and an alarm is triggered once the threshold is exceeded. However, it cannot distinguish between instantaneous fluctuations and continuous anomalies, making it prone to false alarms or missed alarms.

[0005] Periodic acquisition and analysis: Sampling and offline analysis of key sensor data at fixed intervals is slow and makes it difficult to respond to anomalies in real time.

[0006] Lack of multi-source data fusion: Existing methods typically analyze voltage, current or temperature data separately, lacking a comprehensive evaluation of high-frequency, low-frequency and state event data, and thus failing to accurately determine the level and trend of anomalies.

[0007] Large and redundant data transmission: High-frequency data is usually uploaded in its entirety, resulting in high communication bandwidth usage and increased load on processing units. Summary of the Invention

[0008] The purpose of this invention is to provide a method for detecting abnormal states of frequency converters based on time-series data analysis, which solves the technical problems of high real-time performance, low data redundancy, and multi-source feature fusion in frequency converter abnormal monitoring.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] A method for detecting abnormal states of frequency converters based on time-series data analysis includes the following steps:

[0011] Step 1: The sensor data acquisition module collects key sensor data from the frequency converter and adds a timestamp and sampling frequency marker to each key sensor data; thus obtaining the raw data with the timestamp and sampling frequency markers.

[0012] Step 2: The feature extraction module retrieves the raw data and classifies it into high-frequency data, low-frequency data, and state signals according to the data type. It performs sliding window feature extraction on the high-frequency data; performs periodic averaging and trend analysis on the low-frequency data; and records only change events for the state signals using an event-triggered method to obtain state events. It also marks abnormal segments in the high-frequency and low-frequency data and marks abnormal time windows in advance.

[0013] Step 3: The data construction module retrieves the high-frequency features, low-frequency features, status events, and abnormal segment markers obtained in Step 2, and constructs three types of data packets: basic data packets for routine monitoring, sent periodically with low bandwidth; abnormal data packets containing high-frequency features, abnormal segment markers, and event-triggered transmission; and adaptive data packets for dynamically adjusting the sampling rate and data volume based on network bandwidth, load, or remote policies. A data compression strategy is established, including differential representation of high-frequency data; periodic transmission of low-frequency data; mandatory transmission of complete data for abnormal or threshold-exceeding segments; and transmission of only the beginning and end timestamps for continuous segments.

[0014] Step 4: The data transmission module retrieves the data packet generated in Step 3, sets the sending priority, and then sends it to the processing unit via the communication bus according to the priority.

[0015] Step 5: After receiving the data packet, the processing unit decodes the data packet, performs timeline alignment on the decoded data, performs fast filtering, classifies each suspicious abnormal data segment into anomaly levels, adjusts the sampling rate adjustment strategy using anomaly markers with level information, and notifies the sensor data acquisition module to adjust the sampling rate according to the adjusted sampling rate adjustment strategy.

[0016] Step 6: The processing unit performs feature fusion on the suspicious abnormal data, and performs sliding window probability scoring and trend analysis to obtain the abnormal events;

[0017] Step 7: The processing unit processes abnormal events, generates alarm signals, records abnormal data, obtains abnormal event database logs, and periodically summarizes them into summary reports.

[0018] Preferably, step 1 includes the following steps:

[0019] Step 1-1: The sensor data acquisition module receives key sensor data sent by various sensors, including voltage, current, temperature, vibration data, and status signals;

[0020] Steps 1-2: Add a timestamp and sampling frequency mark to each key sensor data point;

[0021] Steps 1-3: Preprocess the key sensor data, including filtering or noise reduction;

[0022] Steps 1-4: Output the raw data with timestamps and sampling frequencies.

[0023] Preferably, step 2 includes the following steps:

[0024] Step 2-1: The feature extraction module retrieves the raw data and classifies it.

[0025] High-frequency data includes voltage and current;

[0026] Low-frequency data includes temperature and vibration data;

[0027] Status signals include switch status and protection actions;

[0028] Step 2-2: Use a sliding window to extract features from the high-frequency data to obtain high-frequency features with window time period labels, including mean, variance, peak value and FFT energy; the window length and step size are set to default presets.

[0029] Steps 2-3: Perform periodic averaging and trend analysis on the low-frequency data to output low-frequency characteristics;

[0030] Steps 2-4: Use event-triggered methods to record only change events for status signals and output status events;

[0031] Steps 2-5: Based on high-frequency and low-frequency characteristics, mark the time windows of anomalies in advance and generate anomaly segment markers; the anomaly segment marker information includes: the start and end time of the anomaly segment, the corresponding sensor type, and the feature deviation threshold or anomaly level prompt.

[0032] Preferably, step 3 includes the following steps:

[0033] Step 3-1: The data construction module retrieves the high-frequency features, low-frequency features, status events, and abnormal segment markers generated in Step 2;

[0034] Step 3-2: Construct the following three types of data packets:

[0035] Basic data package: used for routine detection, including high-frequency feature mean, low-frequency feature mean, and state event summary;

[0036] Abnormal data packets: used for transmitting abnormal segment information, including the original high-frequency data of the abnormal segment, abnormal segment markers, and status events;

[0037] Adaptive data packets: used to dynamically adjust the sampling rate and data volume based on network bandwidth, system load, or remote policies, including high-frequency features, low-frequency features, status events, and abnormal segment markers;

[0038] Step 3-3: Set the sending strategy, including: sending basic data packets at low bandwidth and fixed periods, sending abnormal data packets by event triggering or high priority; and sending adaptive data packets by strategy compression.

[0039] Step 3-4: Set the data compression strategy, including: using differential representation for high-frequency features; periodically sending low-frequency features; forcing the transmission of complete data for abnormal segment markers; transmitting only the first and last timestamps for data segments that appear as continuous segments; and compressing the three types of data packets obtained in Step 3-2 using the data compression strategy.

[0040] Steps 3-5: Set packet priority strategy: send abnormal packets first; reduce the sending frequency of basic packets when the network load is high; send adaptive packets according to bandwidth or remote policy.

[0041] Preferably, step 4 includes the following steps:

[0042] Step 4-1: The data transmission module retrieves the data packets generated in Step 3 and sets a sending priority list according to the type of data packets;

[0043] Step 4-2: The data transmission module sends data packets to the processing unit according to the priority list.

[0044] Preferably, step 5 includes the following steps:

[0045] Step 5-1: After receiving the data packet, the processing unit decodes the data packet and restores the differential representation to the original high-frequency data; it also pads the periodically transmitted data with intermediate values; and for complete data in abnormal segments, it retains the original timestamp.

[0046] Step 5-2: Perform a unified timeline alignment operation on the various types of data decoded in Step 5-1 to obtain aligned data;

[0047] Step 5-3: Use lightweight rules or small models to quickly filter the aligned data, including filtering high-frequency features using peak value, variance, or sliding window anomaly detection methods; filtering low-frequency features using trend deviation or periodic anomaly detection methods; and filtering status events using methods to determine anomalies triggered by key switches or protection actions.

[0048] Based on the screening results, mark the suspicious abnormal data segments;

[0049] Step 5-4: Classify each suspicious abnormal data segment into levels and construct anomaly tags with level information. The data format of the anomaly tags with level information includes the anomaly level, the anomaly time window, and the relevant sensor type.

[0050] Step 5-4: Adjust the sampling rate adjustment strategy using anomaly markers with grade information. Specifically, if the high-frequency feature is within the anomaly segment, the sampling rate is increased; the data packet type corresponding to the anomaly segment is adjusted to anomaly data packet or adaptive data packet.

[0051] Step 5-5: Send the adjusted sampling rate adjustment strategy to the sensor data acquisition module, and the sensor data acquisition module updates the sampling rate according to the sampling rate adjustment strategy.

[0052] Preferably, step 6 includes the following steps:

[0053] Step 6-1: Perform feature fusion on the suspicious abnormal data segments, including concatenating the feature vectors after time synchronization, weighting the abnormality, and evaluating the abnormal correlation between sensors; finally, obtain the fused multi-source feature vector or feature matrix;

[0054] Step 6-2: Use a sliding window method to score the probability of anomalous data for each anomalous data segment, and obtain the probability of anomalous data for each time window;

[0055] Anomaly probability scoring includes: scoring the frequency of anomalies within a window and scoring the probability of energy anomalies for high-frequency features; scoring the probability of trend deviation for low-frequency features; and scoring the probability of key action anomalies for state events.

[0056] Step 6-3: Analyze the trend of its abnormal probability over time, determine the persistence and development direction of the abnormal event, determine whether the abnormality is an instantaneous peak, periodic fluctuation or continuous deterioration, and obtain abnormal trend information.

[0057] Step 6-4: Divide the anomaly probability, anomaly trend information and anomaly level obtained in Step 5 into anomaly events, including anomaly start time, anomaly start time, sensor type involved, anomaly level and anomaly trend.

[0058] Preferably, when performing step 7, the processing unit handles the abnormal event, specifically including the following steps:

[0059] Step 7-1: Generate different types of alarms based on the anomaly level, including local alarms and remote alarms;

[0060] Step 7-2: Save the high-frequency features, low-frequency features, and status events corresponding to the abnormal events to obtain the abnormal event database log;

[0061] Step 7-3: Based on the abnormal event database logs, perform periodic statistics and trend analysis on historical abnormal events to obtain statistics on the number of abnormal events, frequency, duration, high-risk sensors and key links, and generate a summary report.

[0062] This invention discloses a method for detecting abnormal states of frequency converters based on time-series data analysis. This method addresses the technical challenges of high real-time performance, low data redundancy, and multi-source feature fusion in frequency converter anomaly monitoring. The invention classifies and processes high-frequency data, low-frequency data, and status signals. It employs sliding window feature extraction, periodic averaging, and event triggering strategies to extract effective features, reduce redundant data, and construct basic data packets, abnormal data packets, and adaptive data packets. Combined with differential representation and periodic transmission strategies, it reduces data volume and communication bandwidth pressure, ensuring that high-frequency, low-frequency, and event data are aligned on the same timeline. It prioritizes the transmission of abnormal data, guaranteeing the integrity and reliability of anomaly information, and quickly filters suspicious anomalies. By combining sliding window probability scoring and trend analysis, it generates abnormal events and simultaneously triggers adjustments to the sampling rate and data packet strategy, achieving closed-loop optimization. Local and remote alarms are generated based on the anomaly level, and an anomaly database and periodic summary reports are saved for convenient operation and maintenance and remote monitoring. Attached Figure Description

[0063] Figure 1 This is the main flowchart of the present invention;

[0064] Figure 2 This is a flowchart of step 1 of the present invention;

[0065] Figure 3 This is a flowchart of step 2 of the present invention;

[0066] Figure 4 This is a flowchart of step 3 of the present invention;

[0067] Figure 5 This is a flowchart of step 5 of the present invention;

[0068] Figure 6 This is a flowchart of step 6 of the present invention. Detailed Implementation

[0069] Depend on Figures 1-6 The present invention discloses a method for detecting abnormal states of frequency converters based on time-series data analysis, comprising the following steps:

[0070] Step 1: The sensor data acquisition module collects key sensor data from the frequency converter and adds a timestamp and sampling frequency marker to each key sensor data; thus obtaining the raw data with the timestamp and sampling frequency markers.

[0071] Step 1 includes the following steps:

[0072] Step 1-1: The sensor data acquisition module receives key sensor data sent by various sensors, including voltage, current, temperature, vibration data, and status signals;

[0073] In this embodiment, each type of data can be collected through IoT sensors (such as voltage transformers / transmitters, Hall current sensors / transmitters, thermocouples / transmitters, acceleration sensors / transmitters, digital inputs / transmitters, etc.).

[0074] Voltage data includes input / output phase voltages and bus voltages;

[0075] The current data includes the current of each phase and the DC bus current;

[0076] Temperature data includes IGBT, heat sink, and motor bearing temperatures;

[0077] Vibration data includes motor bearing and fan vibration acceleration;

[0078] Status signals include switch status and protection action triggering.

[0079] Steps 1-2: Add a timestamp and sampling frequency mark to each key sensor data point;

[0080] In this embodiment, the timestamp is generated using a microprocessor system clock or a network time protocol (such as NTP) synchronized clock. The timestamp format is generally ISO8601 or UNIX timestamp.

[0081] The default values ​​for the sampling frequency flag are as follows:

[0082] High-frequency data (such as voltage / current): Default value 1kHz-10kHz;

[0083] Low-frequency data (such as temperature / vibration): 1Hz-100Hz;

[0084] Status signal: Records the time when the event is triggered.

[0085] The basic structure of each key sensor data point should include at least: [unique sensor identifier] - [timestamp] - [current sampling frequency] - [raw measurement value] - [voltage / current / temperature / vibration / state].

[0086] Steps 1-3: Preprocess the key sensor data, including filtering or noise reduction;

[0087] In this embodiment, high-frequency data can be filtered using a digital low-pass filter or a band-pass filter (e.g., an FIR filter or an IIR filter) to remove noise or interference frequencies.

[0088] For filtering low-frequency data, moving average or exponential smoothing filters can be used to remove random fluctuations.

[0089] Small-amplitude random noise can be addressed using wavelet denoising or sliding window midpoint filtering; sudden outliers can be identified and corrected through outlier detection (e.g., exceeding the mean ± 3σ).

[0090] Steps 1-4: Output the raw data with timestamps and sampling frequencies.

[0091] In this embodiment, the original data contains at least the following bytes:

[0092] [Unique data packet number]-[Data start timestamp]-[Number of sensor data in the current data packet]-[Array of multiple sensor data].

[0093] Step 2: The feature extraction module retrieves the raw data and classifies it into high-frequency data, low-frequency data, and state signals according to the data type. It performs sliding window feature extraction on the high-frequency data; performs periodic averaging and trend analysis on the low-frequency data; and records only change events for the state signals using an event-triggered method to obtain state events. It also marks abnormal segments in the high-frequency and low-frequency data and marks abnormal time windows in advance.

[0094] Step 2 includes the following steps:

[0095] Step 2-1: The feature extraction module retrieves the raw data and classifies it.

[0096] High-frequency data includes voltage and current;

[0097] Low-frequency data includes temperature and vibration data;

[0098] Status signals include switch status and protection actions;

[0099] Step 2-2: Use a sliding window to extract features from the high-frequency data to obtain high-frequency features with window time period labels, including mean, variance, peak value and FFT energy; the window length and step size are set to default presets.

[0100] In this embodiment, the window length L and step size S are controlled by preset parameters or dynamically adjusted parameters. For example, the voltage / current high-frequency data window length is 100ms and the step size is 50ms (50% overlap).

[0101] The calculation and processing techniques for mean, variance, peak value, and FFT energy are all existing technologies, so they will not be described in detail.

[0102] The data structure for high-frequency features with window time period markers is as follows:

[0103] [Corresponding Sensor ID]-[Window Start Timestamp]-[Window End Timestamp]-[Mean]-[Variance]-[Peak]-[Frequency Domain Energy].

[0104] Steps 2-3: Perform periodic averaging and trend analysis on the low-frequency data to output low-frequency characteristics;

[0105] Low-frequency data is periodically averaged by taking the average value over a preset fixed time period.

[0106] Trend analysis can be performed using linear regression or exponentially weighted moving average (EWMA), such as analyzing temperature or vibration trends. The final output can be a slope or trend indicator to determine whether a gradual anomaly is occurring or a continuous deterioration is occurring.

[0107] Steps 2-4: Use event-triggered methods to record only change events for status signals and output status events;

[0108] Status signals using an event-triggered method only record status change events, for example:

[0109] The switch changes from 0 to 1 or from 1 to 0.

[0110] The data structure for output status events is as follows:

[0111] [Corresponding Sensor ID]-[Time of Change]-[Old State]-[New State].

[0112] Steps 2-5: Based on high-frequency and low-frequency characteristics, mark the time windows of anomalies in advance and generate anomaly segment markers; the anomaly segment marker information includes: the start and end time of the anomaly segment, the corresponding sensor type, and the feature deviation threshold or anomaly level prompt.

[0113] When marking the time window for anomalies in advance, this embodiment can adopt the following method:

[0114] For high-frequency anomaly detection, anomalies are marked when the sliding window features exceed the threshold range (e.g., mean ±3σ, peak value exceeds the safety limit).

[0115] For low-frequency anomaly detection, anomalies are marked when the trend slope exceeds a preset threshold or the periodic average deviates from the normal range.

[0116] The specific data structure for the anomaly segment marking information is as follows:

[0117] [Corresponding Sensor ID]-[Window Start Timestamp]-[Window End Timestamp]-[High Frequency / Low Frequency / Status]-[Degree of Deviation]-[Abnormality Level].

[0118] Step 3: The data construction module retrieves the high-frequency features, low-frequency features, status events, and abnormal segment markers obtained in Step 2, and constructs three types of data packets: basic data packets for routine monitoring, sent periodically with low bandwidth; abnormal data packets containing high-frequency features, abnormal segment markers, and event-triggered transmission; and adaptive data packets for dynamically adjusting the sampling rate and data volume based on network bandwidth, load, or remote policies. A data compression strategy is established, including differential representation of high-frequency data; periodic transmission of low-frequency data; mandatory transmission of complete data for abnormal or threshold-exceeding segments; and transmission of only the beginning and end timestamps for continuous segments.

[0119] Step 3 includes the following steps:

[0120] Step 3-1: The data construction module retrieves the high-frequency features, low-frequency features, status events, and abnormal segment markers generated in Step 2;

[0121] In this embodiment, high-frequency features generally include the mean, variance, peak value, and FFT energy calculated by the sliding window; low-frequency features include periodic average and trend slope; status events include switch status changes and protection action events; and abnormal segment markers include abnormal start and end times, abnormal level, and abnormal type.

[0122] Step 3-2: Construct the following three types of data packets:

[0123] Basic data package: used for routine detection, including high-frequency feature mean, low-frequency feature mean, and state event summary; for example:

[0124] [Corresponding Sensor ID]-[Timestamp]-[High Frequency Sensor Mean Set]-[Low Frequency Sensor Mean Set]-[State Event Count]-[State Event Summary].

[0125] Abnormal data packets: used for transmitting abnormal segment information, including the original high-frequency data of the abnormal segment, abnormal segment markers, and status events; for example:

[0126] [Corresponding Sensor ID]-[Window Start Timestamp]-[Window End Timestamp]-[Original Data of Abnormal Segment]-[Abnormal Related Events]-[Abnormal Level].

[0127] Adaptive data packets: used to dynamically adjust the sampling rate and data volume based on network bandwidth, system load, or remote policies, including high-frequency features, low-frequency features, status events, and abnormal segment markers; for example:

[0128] [Corresponding Sensor ID]-[Timestamp]-[High-Frequency Sensor Mean Set]-[Low-Frequency Sensor Mean Set]-[State Event Count]-[State Event Summary]-[Current Compression Strategy].

[0129] Step 3-3: Set the sending strategy, including: sending basic data packets at low bandwidth and fixed periods, sending abnormal data packets by event triggering or high priority; and sending adaptive data packets by strategy compression.

[0130] Step 3-4: Set the data compression strategy, including: using differential representation for high-frequency features; periodically sending low-frequency features; forcing the transmission of complete data for abnormal segment markers; transmitting only the first and last timestamps for data segments that appear as continuous segments; and compressing the three types of data packets obtained in Step 3-2 using the data compression strategy.

[0131] In this embodiment, the transmission and compression strategies are designed to reduce communication volume. Differential representation of high-frequency features means only the change from the previous window mean / peak value is transmitted; periodic transmission of low-frequency features can omit periods without change, reducing redundancy; abnormal segments are marked for complete transmission to ensure their integrity and facilitate rapid response. For continuous non-abnormal data segments, only the beginning and end timestamps are transmitted to reduce invalid data transmission.

[0132] Steps 3-5: Set packet priority strategy: send abnormal packets first; reduce the sending frequency of basic packets when the network load is high; send adaptive packets according to bandwidth or remote policy.

[0133] Step 4: The data transmission module retrieves the data packet generated in Step 3, sets the sending priority, and then sends it to the processing unit via the communication bus according to the priority.

[0134] Step 4 includes the following steps:

[0135] Step 4-1: The data transmission module retrieves the data packets generated in Step 3 and sets a sending priority list according to the type of data packets;

[0136] In this embodiment, abnormal data packets have a high priority, adaptive data packets have a medium priority, and basic data packets have a low priority. If multiple abnormal data packets occur simultaneously, they can be further sorted according to their abnormality level: severe abnormality > medium-level abnormality > minor abnormality.

[0137] Step 4-2: The data transmission module sends data packets to the processing unit according to the priority list.

[0138] Data packet priority information is marked by default through fields in the data packet header, such as:

[0139] [Corresponding Sensor ID]-[Basic / Abnormal / Adaptive]-[Priority: High / Medium / Low]-[Window End Timestamp]....

[0140] Step 5: After receiving the data packet, the processing unit decodes the data packet, performs timeline alignment on the decoded data, performs fast filtering, classifies each suspicious abnormal data segment into anomaly levels, adjusts the sampling rate adjustment strategy using anomaly markers with level information, and notifies the sensor data acquisition module to adjust the sampling rate according to the adjusted sampling rate adjustment strategy.

[0141] Step 5 includes the following steps:

[0142] Step 5-1: After receiving the data packet, the processing unit decodes the data packet and restores the differential representation to the original high-frequency data; it also pads the periodically transmitted data with intermediate values; and for complete data in abnormal segments, it retains the original timestamp.

[0143] In this embodiment, the following algorithm can be used to recover high-frequency features represented by difference:

[0144] Original value [t] = previous value [t-1] + increment [t];

[0145] Where, the increment [t] represents the difference between data at adjacent time points, used for time series data accumulation calculation. t represents time point.

[0146] For low-frequency features transmitted periodically, if missing windows exist, linear interpolation or intermediate values ​​are used to fill in the gaps, such as:

[0147] The fill value [t] = (the value of the previous time step [t-1] + the value of the next time step [t+1]) / 2.

[0148] Step 5-2: Perform a unified timeline alignment operation on the various types of data decoded in Step 5-1 to obtain aligned data;

[0149] In this embodiment, the smallest common time unit (such as milliseconds) can be used as the unified time axis step size. High-frequency windows, low-frequency periods, and discrete events are aligned according to timestamps. The result is essentially a unified time series matrix, as shown in Table 1.

[0150]

[0151]

[0152] Table 1

[0153] Step 5-3: Use lightweight rules or small models to quickly filter the aligned data, including filtering high-frequency features using peak value, variance, or sliding window anomaly detection methods; filtering low-frequency features using trend deviation or periodic anomaly detection methods; and filtering status events using methods to determine anomalies triggered by key switches or protection actions.

[0154] Based on the screening results, mark the suspicious abnormal data segments;

[0155] In this embodiment, the abnormal condition for high-frequency features is if the high-frequency peak value [t] > the peak value threshold or the high-frequency variance [t] > the variance threshold: it is marked as suspicious.

[0156] The abnormal condition for low-frequency features is if |current low-frequency value - low-frequency trend value| > trend threshold: then it is marked as suspicious.

[0157] Step 5-4: Classify each suspicious abnormal data segment into levels and construct anomaly tags with level information. The data format of the anomaly tags with level information includes the anomaly level, the anomaly time window, and the relevant sensor type.

[0158] In this embodiment, the level division is as follows:

[0159] A significant deviation from the threshold or a series of abnormal windows is considered high.

[0160] Short-term abnormalities or slight deviations are considered neutral.

[0161] Single-point instantaneous anomaly is low.

[0162] Step 5-4: Adjust the sampling rate adjustment strategy using anomaly markers with grade information. Specifically, if the high-frequency feature is within the anomaly segment, the sampling rate is increased; the data packet type corresponding to the anomaly segment is adjusted to anomaly data packet or adaptive data packet.

[0163] Step 5-5: Send the adjusted sampling rate adjustment strategy to the sensor data acquisition module, and the sensor data acquisition module updates the sampling rate according to the sampling rate adjustment strategy.

[0164] Step 6: The processing unit performs feature fusion on the suspicious abnormal data, and performs sliding window probability scoring and trend analysis to obtain the abnormal events;

[0165] Step 6 includes the following steps:

[0166] Step 6-1: Perform feature fusion on the suspicious abnormal data segments, including concatenating the feature vectors after time synchronization, weighting the abnormality, and evaluating the abnormal correlation between sensors; finally, obtain the fused multi-source feature vector or feature matrix;

[0167] In this embodiment, the alignment result of the unified time axis in step 5 is used to stitch together high-frequency features, low-frequency features, and state events according to a unified time window. For example:

[0168] Eigenvector[t] = [HF1[t],HF2[t],LF1[t],LF2[t],state[t],...].

[0169] Anomaly weighting can assign different weights to features based on the anomaly level in step 5, such as:

[0170] High-frequency features and high-level anomalies → increased weighting;

[0171] Low-frequency features or state events → weights are assigned according to their abnormal importance.

[0172] Feature fusion is the sum of weighted high-frequency and low-frequency features, which is the sum of features multiplied by their weights.

[0173] Step 6-2: Use a sliding window method to score the probability of anomalous data for each anomalous data segment, and obtain the probability of anomalous data for each time window;

[0174] Anomaly probability scoring includes: scoring the frequency of anomalies within a window and scoring the probability of energy anomalies for high-frequency features; scoring the probability of trend deviation for low-frequency features; and scoring the probability of key action anomalies for state events.

[0175] In this embodiment, a sliding window approach is used to perform probability scoring for each suspicious abnormal data segment. The sliding window moves forward segment by segment on the time axis, reading the data within the current window with a fixed window length each time, and calculating the abnormal probability of high-frequency features, low-frequency features, and state events respectively.

[0176] For high-frequency features, the number of abnormal feature points appearing within the window is counted, and the degree of abnormality of the window is assessed by their proportion. At the same time, the deviation between the energy features within the window (such as sudden amplitude increases, frequency domain energy changes, etc.) and the energy range during normal operation is compared, and the degree of deviation is used to determine whether the window has a high probability of being abnormal.

[0177] For low-frequency features, the system will determine whether the average value, trend changes, etc., within the window significantly deviate from the device's historical stable range or reference trend. If the low-frequency data within the window shows a significant increase, decrease, or deviation from the normal trend, it will be assigned a higher probability of anomaly.

[0178] For status events, the system checks whether critical switch actions, protection triggers, or other events that do not belong to the normal cyclical behavior have occurred in the window. If the window contains such abnormal events, the probability of the window being abnormal is increased.

[0179] Finally, the anomaly manifestations from the three dimensions of high-frequency features, low-frequency features, and state events are combined to generate the anomaly probability of the window. Each window corresponds to a probability value, forming a complete time series probability distribution.

[0180] Step 6-3: Analyze the trend of its abnormal probability over time, determine the persistence and development direction of the abnormal event, determine whether the abnormality is an instantaneous peak, periodic fluctuation or continuous deterioration, and obtain abnormal trend information.

[0181] In this embodiment, when analyzing the trend of the abnormal probability over time, we first observe whether these probability values ​​show a continuous increase, periodic fluctuation, or rapid decline over time. If the abnormal probability gradually increases in multiple consecutive windows, it indicates that the abnormality has a continuous worsening trend; if the probability value shows obvious periodic fluctuations, it indicates that the device may have a periodic fluctuation problem; if the probability only has a momentary peak but quickly returns to normal, it is judged as a momentary peak anomaly.

[0182] Simultaneously, the system assesses the duration, fluctuation range, and rate of change of these anomaly probabilities to infer the development direction of the anomalous events. Through this analysis of probability trends over time, the system can distinguish between different types of anomalous characteristics, such as differentiating between "rapidly emerging anomalies" and "slowly escalating anomalies," thereby outputting anomaly trend information to clarify the nature of the anomalous events in subsequent processing.

[0183] Step 6-4: Divide the anomaly probability, anomaly trend information and anomaly level obtained in Step 5 into anomaly events, including anomaly start time, anomaly start time, sensor type involved, anomaly level and anomaly trend.

[0184] In this embodiment, the anomaly probability score from step 6-2, the anomaly trend analysis results from step 6-3, and the anomaly level determined in step 5 are integrated. The system first identifies the time period during which the anomaly probability remains consistently high, and extracts the start and end times of the anomaly accordingly. Then, combining the anomaly level information provided in step 5, the severity of the anomaly is determined, and the type of sensor involved is recorded, such as whether it is a current sensor anomaly, a voltage fluctuation anomaly, or a temperature trend anomaly.

[0185] The processing unit also incorporates the abnormal trend markers obtained in step 6-3 into the event description, such as whether the anomaly is a momentary peak, a periodic fluctuation, or a continuously worsening trend. By integrating information such as anomaly level, anomaly trend, probability range, and time period, a structured anomaly event is finally generated. Each anomaly event contains necessary diagnostic information, which can be referenced by subsequent alarm modules, remote monitoring modules, and operation and maintenance systems.

[0186] Step 7: The processing unit processes abnormal events, generates alarm signals, records abnormal data, obtains abnormal event database logs, and periodically summarizes them into summary reports.

[0187] During step 7, the processing unit handles abnormal events, specifically including the following steps:

[0188] Step 7-1: Generate different types of alarms based on the anomaly level, including local alarms and remote alarms;

[0189] Step 7-2: Save the high-frequency features, low-frequency features, and status events corresponding to the abnormal events to obtain the abnormal event database log;

[0190] Step 7-3: Based on the abnormal event database logs, perform periodic statistics and trend analysis on historical abnormal events to obtain statistics on the number of abnormal events, frequency, duration, high-risk sensors and key links, and generate a summary report.

[0191] This invention discloses a method for detecting abnormal states of frequency converters based on time-series data analysis. This method addresses the technical challenges of high real-time performance, low data redundancy, and multi-source feature fusion in frequency converter anomaly monitoring. The invention classifies and processes high-frequency data, low-frequency data, and status signals. It employs sliding window feature extraction, periodic averaging, and event triggering strategies to extract effective features, reduce redundant data, and construct basic data packets, abnormal data packets, and adaptive data packets. Combined with differential representation and periodic transmission strategies, it reduces data volume and communication bandwidth pressure, ensuring that high-frequency, low-frequency, and event data are aligned on the same timeline. It prioritizes the transmission of abnormal data, guaranteeing the integrity and reliability of anomaly information, and quickly filters suspicious anomalies. By combining sliding window probability scoring and trend analysis, it generates abnormal events and simultaneously triggers adjustments to the sampling rate and data packet strategy, achieving closed-loop optimization. Local and remote alarms are generated based on the anomaly level, and an anomaly database and periodic summary reports are saved for convenient operation and maintenance and remote monitoring.

Claims

1. A variable frequency drive abnormal state detection method based on time series data analysis, characterized by: Comprising the following steps: Step 1: The sensor data acquisition module acquires the key sensor data of the frequency converter, adds a timestamp and a sampling frequency label to each key sensor data respectively, and obtains the original data marked with the timestamp and the sampling frequency; Step 2: The feature extraction module calls the original data, classifies the original data into high-frequency data, low-frequency data and state signals according to the data type of the original data, performs sliding window feature extraction on the high-frequency data, performs periodic averaging and trend analysis on the low-frequency data, and records only the change events by using an event triggering mode for the state signals to obtain state events, and labels the abnormal time window in advance by marking the abnormal segments of the high-frequency data and the low-frequency data; Step 3: The data construction module calls the high-frequency features, low-frequency features, state events and abnormal segment labels obtained in step 2, and constructs three types of data packets: a basic data packet for regular monitoring, which is sent at a low bandwidth and at a fixed period; an abnormal data packet containing high-frequency features, abnormal segment labels and event-triggered sending; and an adaptive data packet for dynamically adjusting the sampling rate and data volume according to the network bandwidth, load or remote strategy; a data compression strategy is established, including differential representation of high-frequency data, periodic transmission of low-frequency data, and forced transmission of complete data for abnormal or threshold-exceeding segments; only the start and end timestamps of the continuous segment are transmitted; Step 4: The data transmission module calls the data packets generated in step 3, sets the sending priority, and then sends the data packets to the processing unit through the communication bus according to the priority; Step 5: After receiving the data packets, the processing unit decodes the data packets, aligns the decoded data on the time axis, performs rapid screening, divides each suspicious abnormal data segment into an abnormal level, adjusts the sampling rate adjustment strategy by using the abnormal label with level information, and notifies the sensor data acquisition module to adjust the sampling rate according to the adjusted sampling rate adjustment strategy; Step 6: The processing unit performs feature fusion on the suspicious abnormal data, performs sliding window probability scoring and trend analysis, and obtains an abnormal event; Step 7: The processing unit processes the abnormal event, generates an alarm signal, records the abnormal data, obtains an abnormal event database log, and periodically summarizes the log into a summary report. When step 1 is performed, the following steps are specifically included: Step 1-1: The sensor data acquisition module receives the key sensor data sent by each sensor, including voltage, current, temperature, vibration data and state signals; 2. The inverter abnormal state detection method based on time series data analysis of claim 1, wherein: Step 1-2: A timestamp and a sampling frequency label are added to each key sensor data respectively; Step 1-3: The key sensor data is preprocessed, including filtering or denoising; Step 1-4: The original data marked with the timestamp and the sampling frequency is outputted. When step 2 is performed, the following steps are specifically included: Step 2-1: The feature extraction module calls the original data, and classifies the original data as follows:

3. The method of claim 1, wherein the method comprises: The high-frequency data includes voltage and current; The low-frequency data includes temperature and vibration data; The state signal includes the switch state and the protection action; ​ ​ Step 2-2: Feature extraction of high-frequency data using sliding window, obtaining high-frequency features with window time period labels, including mean, variance, peak value and FFT energy; window length and step size use default preset settings; Step 2-3: Periodic average and trend analysis of low-frequency data, output low-frequency features; Step 2-4: Event-triggered recording of state signal changes, output state events; Step 2-5: Based on high-frequency and low-frequency features, label abnormal time windows in advance, generate abnormal segment labels; Abnormal segment label information includes: abnormal segment start and end time, corresponding sensor type and feature deviation threshold or abnormal level prompt.

4. The method of claim 1, wherein the method comprises: When performing step 3, it specifically includes the following steps: Step 3-1: The data construction module calls the high-frequency features, low-frequency features, state events and abnormal segment labels generated in step 2; Step 3-2: Construct the following three kinds of data packets: Basic data packet: used for regular detection, including high-frequency feature mean, low-frequency feature mean, state event summary; Abnormal data packet: used for abnormal segment information transmission, including abnormal segment original high-frequency data, abnormal segment label, state event; Adaptive data packet: used to dynamically adjust sampling rate and data volume according to network bandwidth, system load or remote strategy, including high-frequency features, low-frequency features, state events and abnormal segment labels; Step 3-3: Set the sending strategy, including: basic data packet is sent in a low-bandwidth and periodic manner, abnormal data packet is sent in an event-triggered or high-priority manner; adaptive data packet is sent in a strategy-compressed manner; Step 3-4: Set the data compression strategy, including: high-frequency features use differential representation; low-frequency features are sent periodically; abnormal segment labels send complete data; data that appears as a continuous segment only transmits the start and end timestamps; use data compression strategy to compress the three kinds of data packets obtained in step 3-2; Step 3-5: Set the data packet priority strategy: abnormal data packet is sent first; basic data packet reduces the sending frequency when the network load is high; adaptive data packet is sent according to bandwidth or remote strategy.

5. The method for detecting abnormal state of frequency converter based on time series data analysis according to claim 1, characterized in that: When performing step 4, it specifically includes the following steps: Step 4-1: The data transmission module calls the data packets generated in step 3, and sets the sending priority list according to the type of data packet; Step 4-2: The data transmission module sends the data packets to the processing unit according to the priority list.

6. The method of claim 1, wherein the method further comprises: When performing step 5, it specifically includes the following steps: Step 5-1: After receiving the data packet, the processing unit decodes the data packet, restores the differential representation to the original high-frequency data; completes the intermediate value of the periodically sent data; for the complete data of the abnormal segment, its original timestamp is retained; Step 5-2: Align the decoded data of various types in step 5-1 on the unified time axis to obtain aligned data; Step 5-3: Use lightweight rules or small models to quickly filter the aligned data, including using peak value, variance or sliding window anomaly judgment method to filter high-frequency features; Using trend deviation or periodic anomaly judgment method to filter low-frequency features; Using key switch or protection action triggered anomaly judgment method to filter state events; According to the screening result, mark the suspicious abnormal data segment; Step 5-4: Grade division is performed on each suspicious abnormal data segment, and a graded abnormal mark is constructed, and the data format of the graded abnormal mark includes an abnormal grade, an abnormal time window and a related sensor type; Step 5-4: The graded abnormal mark is used to adjust the sampling rate adjustment strategy, specifically, if the high-frequency feature is in the abnormal segment, the sampling rate is increased; the data packet type corresponding to the abnormal segment is adjusted to an abnormal data packet or an adaptive data packet; Step 5-5: The adjusted sampling rate adjustment strategy is sent to the sensor data acquisition module, and the sensor data acquisition module updates the sampling rate according to the sampling rate adjustment strategy.

7. The method of claim 1, wherein the method further comprises: determining a frequency of the variable frequency drive; and determining a frequency of the variable frequency drive based on the time series data. When step 6 is performed, the following steps are specifically included: Step 6-1: Feature fusion is performed on the data of the suspicious abnormal data segment, including time-synchronized feature vector splicing, abnormal weight weighting, and inter-sensor abnormal correlation evaluation; finally, a fused multi-source feature vector or feature matrix is obtained; Step 6-2: The sliding window method is used to score the abnormal probability of each abnormal data segment, and the abnormal probability of each time window is obtained; The abnormal probability scoring includes: scoring the abnormal frequency and energy abnormal probability of the high-frequency feature within the window; scoring the trend deviation probability of the low-frequency feature; scoring the key action abnormal probability of the state event; Step 6-3: The trend of the change of the abnormal probability with time is analyzed, the abnormal event persistence and development direction are judged, it is judged whether the abnormality is a transient peak, a periodic fluctuation or a continuous deterioration, and abnormal trend information is obtained; Step 6-4: The abnormal probability, abnormal trend information and abnormal grade division obtained in step 5 are combined to generate an abnormal event, including an abnormal start time, an abnormal start time, a related sensor type, an abnormal grade and an abnormal trend.

8. The method of claim 1, wherein: When step 7 is performed, the processing unit processes the abnormal event, specifically including the following steps: Step 7-1: Different types of alarms are generated according to the abnormal grade, including local alarms and remote alarms; Step 7-2: The high-frequency feature, the low-frequency feature and the state event corresponding to the abnormal event are saved, and an abnormal event database log is obtained; Step 7-3: According to the abnormal event database log, periodic statistics and trend analysis are performed on the historical abnormal events, and the abnormal number, frequency, duration, high-risk sensor and key link statistics are obtained, and a summary report is generated.

Images