Abnormality detection method and system based on MCC cabinet electrical parameters and event log

Abstract

The application relates to the technical field of electrical equipment state monitoring, and particularly provides an abnormality detection method and system based on MCC cabinet electric parameters and event logs, which comprises the following steps: collecting steady-state electric parameters, three-phase current waveforms in a starting process and millisecond-level event logs; performing hierarchical determination on electrical safety, starting behavior and function consistency through a rule engine, and outputting continuous rule severity scores; reconstructing the steady-state data through a time sequence self-encoder and calculating a steady-state abnormality score, wherein a loss function of the time sequence self-encoder comprises an apparent power conservation regularization term; reconstructing the starting waveforms through a one-dimensional convolution self-encoder and calculating a starting abnormality score; and outputting hierarchical early warnings according to threshold intervals after weighted fusion of the three types of scores. The application does not need to additionally install sensors, fuses deterministic rules and unsupervised learning, can effectively detect motor implicit degradation and MCC function abnormalities, and reduces the false alarm rate.

Description

Technical Field

[0001] This invention belongs to the field of electrical equipment condition monitoring technology, specifically relating to an anomaly detection method and system based on MCC cabinet electrical parameters and event logs. Background Technology

[0002] The MCC (Motor Control Center) cabinet is the core equipment for centralized control and protection of motors in industrial fields. Existing intelligent MCC cabinets are typically equipped with motor protectors or intelligent feeder units, providing basic protection functions such as overload, short circuit, phase loss, and three-phase imbalance. However, these protection functions are all passively triggered, meaning they only activate when electrical parameters exceed preset thresholds. They cannot proactively identify early signs of motor degradation (such as rotor bar breakage, bearing wear, and winding insulation aging) based on long-term trends in operating data, nor can they detect functional abnormalities in the MCC cabinet's own control circuits, protection devices, or measurement units.

[0003] On the other hand, traditional threshold alarm methods are susceptible to false alarms (parameters exceeding limits due to changes in operating conditions but not actual faults) or missed alarms (parameters not exceeding thresholds in the early stages of a fault) due to factors such as load fluctuations and power supply voltage changes. While installing dedicated sensors for vibration, temperature, etc., on the motor side can improve detection accuracy, the retrofit cost is high and the construction is complex, making it difficult to promote in already operational MCC cabinets. Summary of the Invention

[0004] To address the aforementioned shortcomings of existing technologies, this invention provides an anomaly detection method and system based on MCC cabinet electrical parameters and event logs, thereby resolving the aforementioned technical problems.

[0005] In a first aspect, the present invention provides an anomaly detection method based on MCC cabinet electrical parameters and event logs, comprising: For each circuit of the MCC cabinet, monitoring data is collected under steady-state operating conditions and during motor startup. The monitoring data includes electrical parameters and event logs. The electrical parameters in the monitoring data are preprocessed to generate steady-state time window data and startup waveform data. Anomaly detection is performed on the steady-state time window data, startup waveform data, and event logs using a pre-built rule engine, and a rule severity score is output. The rule engine includes electrical safety rules, startup behavior rules, and MCC functional consistency rules. A temporal autoencoder is used to reconstruct the steady-state time window data and calculate the steady-state anomaly score; a one-dimensional convolutional autoencoder is used to reconstruct the startup waveform data and calculate the startup anomaly score; wherein the training loss function of the temporal autoencoder includes a physical constraint regularization term based on apparent power conservation; The rule severity score, steady-state anomaly score, and activation anomaly score are weighted and summed according to preset weights. The anomaly level is determined based on the threshold range in which the weighted sum falls, and the corresponding warning information is output.

[0006] In one optional implementation, for each circuit in the MCC cabinet, monitoring data is collected under steady-state operating conditions and during motor startup, including: Under steady-state operating conditions, the effective values ​​of three-phase current, effective values ​​of three-phase voltage, active power, and power factor are collected at a sampling frequency of 1Hz. When a motor start-up event is detected, the system automatically switches to a sampling frequency of 100Hz to collect the instantaneous waveforms of the three-phase current during the start-up process. The detection condition for the start-up event is that the current of any one phase jumps from less than 0.05 times the rated current to more than 0.2 times the rated current and lasts for more than 20 milliseconds. Action event logs are collected independently with an accuracy of 1 millisecond. The event logs include at least start command events, contactor engagement events, and protection action events, and each event is accompanied by a precise timestamp.

[0007] In one optional implementation, the electrical parameters in the monitoring data are preprocessed to generate steady-state time window data and startup waveform data, including: The electrical parameters and event logs are time-aligned using a unified clock source, with an alignment error of no more than 100 milliseconds; Interpolation compensation is performed on missing values ​​in the time series of electrical parameters. For single missing points, the arithmetic mean of data from adjacent time points is used to fill in the missing values. For multiple consecutive missing points exceeding a set number, the time window is marked as invalid and the subsequent analysis is skipped. Outliers in the electrical parameter time series were removed, and statistical criteria were used to identify outliers that deviated from the mean by more than three standard deviations, which were then replaced with the mean. The preprocessed steady-state electrical parameters are divided into multiple steady-state time window data according to a sliding window of fixed time length; The instantaneous waveforms of the three-phase current acquired at high speed during startup are extracted as independent startup waveform data for each startup event.

[0008] In one optional implementation, a pre-built rule engine is used to perform anomaly detection on the steady-state time window data, startup waveform data, and event logs, and outputs a rule severity score, including: The rule engine is used to match the steady-state time window data, startup waveform data, and event logs with the corresponding rule conditions, and the successfully matched triggering rules and their corresponding severity scores are recorded. The maximum severity score among all triggering rules is used as the final rule severity score in the output.

[0009] In one optional implementation, the electrical safety category rules include: The severity is determined by the ratio of the maximum three-phase current to the rated current. When the ratio is greater than 1.20, the severity is 1.0, and when the ratio is between 1.05 and 1.20, the severity is 0.7. When the average current is greater than 0.2 times the rated current, the minimum phase current is less than 0.1 times the rated current, and the ratio of the maximum phase current to the minimum phase current is greater than 5, the severity is 1.0. Three-phase imbalance rule: The severity is determined based on the degree of imbalance of the three-phase current. When the degree of imbalance is greater than 0.1, the severity is 0.7. When the degree of imbalance is between 0.08 and 0.1, the severity is 0.4. The degree of imbalance is defined as the difference between the maximum phase current and the minimum phase current divided by the average value of the three-phase current. The startup behavior rules include: The starting time limit is dynamically calculated based on the motor's rated power. When the actual measured starting time exceeds the limit, the severity is 0.7. Based on the statistical mean and standard deviation of the peak current of the most recent 20 successful starts, if the peak current of this start deviates from the mean by more than 3 times the standard deviation, the severity is 0.7; if the deviation is between 2 and 3 times the standard deviation, the severity is 0.4. The MCC functional consistency class rules include: When both the start command event and the contactor engagement event occur, but the time from the command issuance to current establishment exceeds 1 second and the current is still less than 0.1 times the rated current afterward, the severity is 1.0; When no protection action event occurs but the fault current exceeds 1.2 times the setting value and lasts for more than 0.5 seconds, the severity is 0.7; When a protection action event occurs but the maximum current before and after the action is less than 1.05 times the setting value, the severity is 0.7.

[0010] In one optional implementation, a time-series autoencoder is used to reconstruct the steady-state time window data and calculate a steady-state anomaly score, including: The time-series autoencoder consists of a time-series encoder and a time-series decoder connected in series. The time-series encoder compresses the input steady-state time window data sequence into a low-dimensional latent vector, and the time-series decoder reconstructs the low-dimensional latent vector into an output sequence with the same dimension as the input sequence. By minimizing the reconstruction error between the input and output sequences, the time-series autoencoder learns the timing pattern under normal operating conditions. Both the time-series encoder and the time-series decoder use long short-term memory networks as basic units. The training method adopts an unsupervised learning approach, and the training data only uses historical steady-state window data that has been verified by the rule engine to be free of anomalies. The training objective is to minimize the reconstruction error between the input and the reconstructed output. The loss function of the time-series self-time encoder adds a physical constraint regularization term to the conventional reconstruction error, and the physical constraint regularization term is based on the principle of apparent power conservation. The steady-state time window data collected in real time is input into the trained time-series self-time encoder. The reconstruction error between the input and the reconstruction output is calculated. The reconstruction error is normalized according to the specified quantile of the reconstruction error distribution of the training set and mapped to the interval between zero and one to obtain the steady-state anomaly score.

[0011] In one optional implementation, the loss function of the time-series self-time encoder adds a physical constraint regularization term to the conventional reconstruction error, including: Based on the principle of apparent power conservation, the reconstructed active power and reactive power are extracted at the output of the time-series decoder of the time-series autoencoder, and the reconstructed apparent power is calculated based on the reconstructed active power and reactive power. Calculate the measured apparent power based on the measured three-phase voltage and three-phase current within the same time window; The absolute value of the difference between the reconstructed apparent power and the measured apparent power is used as the physical constraint regularization term. The larger the physical constraint regularization term is, the more the reconstruction result violates the physical law of apparent power conservation, and the heavier the penalty on the loss function. The total loss function of the temporal autoencoder is a weighted sum of the conventional reconstruction error and the physical constraint regularization term.

[0012] In one optional implementation, a one-dimensional convolutional autoencoder is used to reconstruct the startup waveform data and calculate a startup anomaly score, including: The one-dimensional convolutional autoencoder consists of a convolutional encoder and a convolutional decoder connected in series. The convolutional encoder contains several one-dimensional convolutional layers and pooling layers, which are used to extract local features from the input start waveform data layer by layer and reduce the time dimension, and output a low-dimensional latent feature vector. The convolutional decoder contains several one-dimensional deconvolutional layers and upsampling layers, which are used to gradually restore the latent feature vector into a reconstructed waveform with the same length and number of channels as the input waveform. The input of the one-dimensional convolutional autoencoder is the extracted three-phase current instantaneous waveform data of a single startup process. Each waveform contains three channels corresponding to the three phases, and the length of each channel is the number of sampling points within a fixed time period. The training method adopts an unsupervised learning approach, and the training data only uses historical startup waveform data that has been verified by the rule engine as normal startup. The training objective is to minimize the reconstruction error between the input waveform and the reconstructed waveform. The real-time acquired startup waveform data is input into the trained one-dimensional convolutional autoencoder. The reconstruction error between the input waveform and the reconstructed waveform is calculated. The reconstruction error is normalized according to the specified quantile of the reconstruction error distribution of the training set and mapped to the interval between zero and one to obtain the startup anomaly score.

[0013] In one optional implementation, the rule severity score, steady-state anomaly score, and activation anomaly score are weighted and summed according to preset weights, and the anomaly level is determined based on the threshold range in which the weighted sum falls, including: Weight coefficients were set for rule severity score, steady-state anomaly score and initiation anomaly score respectively, with the sum of the weights of the three being 1. The rule severity score had the largest weight, followed by the steady-state anomaly score, and the initiation anomaly score had the smallest weight. The rule severity score, steady-state anomaly score, and initiation anomaly score are multiplied by their respective weights, and then added together to obtain the comprehensive score. The comprehensive score range is divided into four levels by three preset thresholds: when the comprehensive score is greater than or equal to the first threshold, it is determined as a level 1 alarm; when the comprehensive score is greater than or equal to the second threshold but less than the first threshold, it is determined as a level 2 alarm; when the comprehensive score is greater than or equal to the third threshold but less than the second threshold, it is determined as a level 3 alarm; when the comprehensive score is less than the third threshold, it is determined as a normal state. The first threshold, the second threshold, and the third threshold are preset values, and the first threshold is greater than the second threshold, which is greater than the third threshold.

[0014] Secondly, the present invention provides an anomaly detection system based on MCC cabinet electrical parameters and event logs, comprising: The data acquisition module is used to collect monitoring data for each circuit of the MCC cabinet under steady-state operation and during motor startup. The monitoring data includes electrical parameters and event logs. The module also preprocesses the electrical parameters in the monitoring data to generate steady-state time window data and startup waveform data. The first processing module is used to perform anomaly judgment on the steady-state time window data, startup waveform data and event logs using a pre-built rule engine, and output a rule severity score; the rule engine includes electrical safety rules, startup behavior rules and MCC functional consistency rules; The second processing module is used to reconstruct the steady-state time window data using a temporal autoencoder and calculate the steady-state anomaly score; and to reconstruct the startup waveform data using a one-dimensional convolutional autoencoder and calculate the startup anomaly score; wherein the training loss function of the temporal autoencoder includes a physical constraint regularization term based on apparent power conservation. The early warning output module is used to sum the rule severity score, steady-state anomaly score and activation anomaly score according to preset weights, determine the anomaly level based on the threshold range of the weighted sum, and output the corresponding early warning information.

[0015] The beneficial effects of this invention are as follows: The anomaly detection method and system based on MCC cabinet electrical parameters and event logs provided by this invention utilize only the three-phase electrical parameters and event logs collected by the MCC cabinet itself, without the need to install motor-side sensors, to detect early motor degradation and MCC functional abnormalities. A rule engine performs deterministic verification of electrical safety, startup behavior, and control protection consistency, ensuring rapid response and safety in the face of electrical faults. Unsupervised reconstruction of steady-state data and startup waveforms is performed using a time-series autoencoder and a one-dimensional convolutional autoencoder, respectively, combined with physical constraint regularization, which can uncover hidden anomaly patterns under load fluctuations, reducing false alarms caused by changes in operating conditions. Finally, the rule severity score and dual AI score are weighted and fused for graded early warning, balancing determinism and sensitivity, effectively improving the accuracy and engineering practicality of anomaly detection. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic flowchart of a method according to an embodiment of the present invention.

[0018] Figure 2 This is a schematic flowchart illustrating a method of generating rule severity scores using a rule engine according to an embodiment of the present invention.

[0019] Figure 3 This is a schematic block diagram of a system according to an embodiment of the present invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0022] The anomaly detection method based on MCC cabinet electrical parameters and event logs provided in this embodiment of the invention is executed by a computer device, and correspondingly, the anomaly detection system based on MCC cabinet electrical parameters and event logs runs in the computer device.

[0023] Figure 1 This is a schematic flowchart illustrating a method according to an embodiment of the present invention. Wherein, Figure 1 The executing entity can be an anomaly detection system based on MCC cabinet electrical parameters and event logs. Depending on different requirements, the order of steps in this flowchart can be changed, and some steps can be omitted.

[0024] like Figure 1 As shown, the method includes: S1. For each circuit of the MCC cabinet, monitoring data is collected under steady-state operating conditions and during motor startup. The monitoring data includes electrical parameters and event logs. The electrical parameters in the monitoring data are preprocessed to generate steady-state time window data and startup waveform data. S2. Use a pre-built rule engine to perform anomaly detection on the steady-state time window data, startup waveform data, and event logs, and output a rule severity score; the rule engine includes electrical safety rules, startup behavior rules, and MCC functional consistency rules; S3. The steady-state time window data is reconstructed using a temporal autoencoder, and a steady-state anomaly score is calculated; the startup waveform data is reconstructed using a one-dimensional convolutional autoencoder, and a startup anomaly score is calculated; wherein the training loss function of the temporal autoencoder includes a physical constraint regularization term based on apparent power conservation; S4. The rule severity score, steady-state anomaly score, and activation anomaly score are weighted and summed according to preset weights. The anomaly level is determined based on the threshold range in which the weighted sum falls, and the corresponding warning information is output.

[0025] In one embodiment of the present invention, based on step S1, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0026] S101: Variable Sampling Rate Data Acquisition Each circuit in the MCC cabinet is equipped with an intelligent measurement unit (such as an intelligent motor protector) to collect monitoring data in the following manner: Steady-state operation condition data acquisition: The following electrical parameters are continuously acquired at a sampling frequency of 1Hz (i.e., one set of data is acquired per second): RMS value of three-phase current. Three-phase voltage RMS value Active power Power factor .

[0027] High-speed data acquisition during startup: Real-time monitoring of three-phase current. When a motor startup event is detected, the sampling frequency is automatically switched to 100Hz (i.e., a set of instantaneous values ​​is acquired every 10 milliseconds), acquiring the instantaneous waveforms of the three-phase current during startup. .

[0028] The detection conditions for the startup event are:

[0029] in This is the motor's rated current. High-speed data acquisition continues until the current stabilizes within ±10% of the rated current for more than 2 seconds, or the total high-speed data acquisition time reaches 20 seconds (to prevent infinite data acquisition).

[0030] Event log collection: Action event logs are collected independently with a precision of 1 millisecond, including: Startup command event (The closing command issued by the controller); Contactor engagement event (Contactor auxiliary contact feedback); Protective Action Event (Action signal of thermal relay or electronic protector). Each event is accompanied by a precise timestamp (accuracy ≤ 1ms), which can be read through the digital input channel of the intelligent protector or fieldbus (such as Modbus, Profibus).

[0031] S102: Data Preprocessing Using a unified clock source (NTP) as the standard, align electrical parameters with the timestamps in the event log, requiring an alignment error of [missing information]. ms. Data exceeding this error value will be interpolated or discarded.

[0032] For single missing points in the electrical parameter time series, imputation is performed using the arithmetic mean of data from adjacent time points:

[0033] For the start of the sequence ( ) or endpoint ( The missing ) are respectively used or Replacement. For cases where more than 3 consecutive points are missing, mark the time window as invalid and skip it in subsequent analysis.

[0034] The Laida criterion (3σ rule) is used to identify and remove outliers: the mean of the current parameter within the sliding window is calculated. and standard deviation If the sampled value satisfy:

[0035] It is then identified as an outlier and calculated using the current window average. Replace this point. The window length is set to 60 sampling points (corresponding to 60 seconds of steady-state data).

[0036] The preprocessed steady-state electrical parameters are divided into sliding windows of fixed time lengths, with the window length being... There are 60 seconds of data at each time point, and the sliding step is 1 time point (i.e., a new window is generated every second). Each window contains 8 feature dimensions: , forming a tensor .

[0037] The instantaneous waveforms of the three-phase current are collected at high speed during each startup process. Extract the startup waveform data into independent startup events based on a single startup event. Each waveform has a length of [length missing]. Each sampling point (corresponding to 10 seconds, 100Hz sampling rate) will be used. If the startup process lasts less than 10 seconds, zeros will be padded to the end of the waveform to a fixed length. If it lasts more than 10 seconds, the first 10 seconds will be truncated.

[0038] In one embodiment of the present invention, based on step S2, the following will provide a possible embodiment and its specific implementation will be described in a non-limiting manner. Please refer to [the relevant documentation]. Figure 2 .

[0039] S201: Rule-based matching and severity score recording The pre-built rule engine contains three categories of rules: electrical safety, startup behavior, and MCC functional consistency. For each MCC loop, real-time collected or generated steady-state time window data, startup waveform data, and event logs are input into the rule engine, and each rule is matched against the rule conditions. Each rule has a pre-set severity score. When the rule conditions are met, the severity score of the rule is recorded.

[0040] (a) Electrical safety rules Overload rules define load rate The maximum value of the three-phase current and the rated current The ratio:

[0041] in This represents the effective value of the three-phase current within the current steady-state window. According to... The numerical range determines the severity:

[0042] when When a Level 1 alert (severity 1.0) is triggered, A level 2 warning (severity 0.7) is triggered.

[0043] Phase loss rule: First calculate the average current. The rule is triggered under the following conditions:

[0044] When all of the above conditions are met, it is determined to be a severe phase loss, and the severity is... This rule is used to detect motors operating with a single phase, which is considered a high-risk fault.

[0045] Three-phase imbalance rule: Defines the degree of three-phase current imbalance. for:

[0046] according to The numerical range determines the severity:

[0047] An imbalance exceeding 10% is considered a significant anomaly with a severity of 0.7; an imbalance between 8% and 10% is considered a minor anomaly with a severity of 0.4. For lightly loaded conditions (… The above thresholds can be appropriately relaxed (for example, relaxed to 0.25 and 0.20 respectively) to avoid false alarms.

[0048] (ii) Initiation behavior rules The startup behavior rules are only determined during the motor startup process (i.e., the high-speed sampling period).

[0049] Start-up timeout rule: based on motor rated power (Unit: kW) Dynamically calculated start-up time limit:

[0050] The minimum limit is 2 seconds, and the maximum is 20 seconds. Actual startup time. Defined as the time from the start of current rise (beyond 0.2) The time elapsed until the current stabilizes within ±10% of the rated current. If this occurs, the startup timeout rule will be triggered, with the severity level set to [missing information]. .

[0051] Initiate peak anomaly rules: Maintain a sliding window to record the most recent peak anomalies. Peak current sequence of the first successful start (i.e., a start-up without triggering protection and operating normally). Calculate the mean of the sequence. and sample standard deviation :

[0052] Peak current during startup Deviation from the mean is expressed as a multiple of the standard deviation:

[0053] according to The scope determines the severity:

[0054] If the number of historical startups is less than 20, the rule will not be triggered temporarily, or 1.2 times the device nameplate value will be used as a temporary upper limit for alternative judgment.

[0055] (III) MCC Functional Consistency Rules These rules rely on a combination of event logs and electrical parameters to detect functional abnormalities in control loops, protection devices, or measurement units.

[0056] Contactor fault rules: Define the following logical conditions: Control command events (That is, the controller issued a start command); Contactor engagement event (The contactor auxiliary contact feedback has been engaged); Time from command issuance to current establishment More than 1 second, of which ; exist At that moment, the maximum value of the three-phase current is still less than .

[0057] If all of the above conditions are met, the contactor is deemed faulty (e.g., main contacts are not conducting, contacts are burned or stuck), and the severity is determined accordingly. .

[0058] Protection failure to operate rule: Define fault current threshold This refers to the set current value of the protection device. Real-time monitoring of the maximum three-phase current value is also required. If the following conditions are met simultaneously: Protective Action Event (No protective action was taken); And duration If the time is less than one second, it is determined to be a protective refusal to move, severity level. .

[0059] Protection against malfunctions: If the following conditions are met simultaneously: Protective Action Event ; During the protective action Within 200 milliseconds before and after (i.e. ), maximum current If it is determined to be a false alarm, the severity will be... .

[0060] S202: Rule Severity Score Synthesis For each data collection cycle (or each startup event), the rule engine iterates through all the rules mentioned above, recording the severity score corresponding to each triggered rule. The final rule severity score... Take the maximum value of the severity scores for all triggering rules:

[0061] If no rule is triggered, then .

[0062] This synthesis strategy ensures that the most severe anomalies (such as severe overload, phase loss, and contactor anomalies, all with a severity of 1.0) can directly dominate the output of the rule engine, avoiding the underestimation of multiple low-severity anomalies when they are superimposed.

[0063] In one embodiment of the present invention, based on step S3, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0064] S301: Temporal Autoencoder Construction, Training, and Anomaly Scoring Calculation (I) Structure of a Timing Autoencoder The time-series autoencoder consists of a time-series encoder and a time-series decoder connected in series. Its function is to compress the input steady-state time window data sequence into a low-dimensional latent vector, and then reconstruct the output sequence of the same dimension as the input sequence from this latent vector. By minimizing the reconstruction error between the input and output sequences, the time-series autoencoder learns the temporal dependencies and inherent patterns of electrical parameters under normal operating conditions.

[0065] The timing encoder uses a multilayer long short-term memory (LSTM) network as its basic unit. Let the input sequence be... ,in Vector at each time point (Include The encoder reads the input sequence step by step and outputs the hidden state of the last time step. Alternatively, the aggregated result of the hidden states at all time steps can be used as a low-dimensional latent vector. Typically, the encoder uses a two-layer LSTM, with 128 hidden units in the first layer and 64 hidden units in the second layer, ultimately outputting a latent vector. .

[0066] Temporal decoder: Employs an LSTM structure symmetric to the encoder. The decoder uses the latent vector... As the initial hidden state, a reconstruction sequence is generated step by step. Each The decoder is also a two-layer LSTM (64→128), and finally a fully connected layer maps the output to an 8-dimensional feature space.

[0067] The reconstruction error between the input sequence and the reconstructed sequence is defined as the mean squared error (MSE):

[0068] (ii) Unsupervised training strategies An unsupervised learning approach was used to train the temporal autoencoder. The training data consisted only of historical steady-state window data (i.e., time periods where no rules were triggered) verified by the rule engine to be anomaly-free. Normal operating condition data for more than 7 consecutive days was extracted from the historical database and divided into several segments of length [missing information]. The windows form the training set.

[0069] The training objective is to minimize the following loss function:

[0070] in This is the weighting coefficient for the physical constraint regularization term (0.1 is recommended). This is the physical constraint regularization term (see the next section for details). The optimizer can be Adam, with a learning rate of 0.001, a batch size of 32, and training epochs until the loss converges.

[0071] (iii) Physical constraint regularization term To ensure that the features learned by the model conform to the laws of electrical physics, a physical constraint regularization term based on apparent power conservation is introduced into the loss function. Its calculation method is as follows: For each time point The active power of the decoder reconstruction output and power factor It can calculate the reconfigured reactive power. :

[0072] The reconfigurable apparent power is:

[0073] Simultaneously, based on the measured three-phase voltage and three-phase current within the same time window, the measured apparent power is calculated. Let the average effective value of the line voltage at that time point be... Average value of line current RMS value They are respectively:

[0074] (If the system only provides phase voltage, then line voltage = phase voltage × The measured apparent power is:

[0075] The physical constraint regularization term is defined as the average of the absolute values ​​of the differences between the reconstructed apparent power and the measured apparent power at all time points:

[0076] This regularization term penalizes the behavior of the reconstruction result that violates the apparent power conservation, forcing the model to output physically reliable electrical parameters while maintaining low reconstruction error.

[0077] (iv) Calculation of steady-state anomaly score For real-time acquired steady-state time window data Input the trained temporal autoencoder to obtain the reconstructed sequence. Calculate the reconstruction error of this window:

[0078] In order to map the reconstruction error to a unified To determine the interval, we need to utilize the reconstruction error distribution on the training set. We need to collect the reconstruction errors of all normal windows in the training set. Calculate its 0.01 quantile. and 0.99 quantile :

[0079] Steady-state anomaly score Defined as:

[0080] That is, the score is 0 when the reconstruction error is below the lower limit of the normal distribution, and 1 when it is above the upper limit, with a linear mapping in between.

[0081] To adapt the model to long-term load drift and motor aging, the calculation is performed every 24 hours using data verified by the rule engine within the past week that has no anomalies. and This enables dynamic threshold updates.

[0082] S302: Construction, Training, and Startup Anomaly Score Calculation of One-Dimensional Convolutional Autoencoders (I) One-dimensional convolutional autoencoder structure The one-dimensional convolutional autoencoder consists of a convolutional encoder and a convolutional decoder connected in series. Its function is to compress the input three-phase current startup waveform into a low-dimensional latent feature vector, and then progressively reconstruct a reconstructed waveform with the same length and number of channels as the input waveform from this latent vector. By minimizing the reconstruction error between the input waveform and the reconstructed waveform, the model learns the waveform pattern of a normal startup process.

[0083] (1) Convolutional encoder The encoder comprises several one-dimensional convolutional layers and pooling layers to extract local features layer by layer and reduce the temporal dimension. The input data is the instantaneous waveform of the three-phase current during a single startup process, denoted as... ,in Indicates the number of channels (corresponding to phases A, B, and C). This indicates the number of sampling points for each channel (corresponding to 10 seconds and a 100Hz sampling rate). If the actual startup process lasts less than 10 seconds, zeros are padded to the end of the waveform to a fixed length; if it exceeds 10 seconds, the first 10 seconds are truncated.

[0084] A typical encoder structure is as follows: First layer: A one-dimensional convolutional layer with 3 input channels and 16 output channels, a kernel size of 5, a stride of 1, and padding of 2 (to preserve length), followed by a ReLU activation function and a max-pooling layer (pooling window 2, stride 2). The output size becomes... .

[0085] The second layer is a one-dimensional convolutional layer with 16 input channels and 32 output channels. The kernel size is 5, the stride is 1, and the padding is 2. It is followed by ReLU and max pooling (pooling window 2, stride 2). The output size becomes... .

[0086] The third layer is a one-dimensional convolutional layer with 32 input channels and 64 output channels. The kernel size is 5, the stride is 1, and the padding is 2. It is followed by ReLU and max pooling (pooling window 2, stride 2). The output size becomes... .

[0087] Flatten: to The feature map is flattened into a one-dimensional vector with a length of .

[0088] Fully connected layer: maps 8000-dimensional vectors to low-dimensional latent vectors ,in The potential spatial dimension is 128 (recommended). This layer is followed by Dropout (scale 0.2) to prevent overfitting.

[0089] (2) Convolutional Decoder The decoder consists of several one-dimensional deconvolution layers and upsampling layers to progressively reconstruct the latent vectors into a reconstructed waveform of the same size as the input. The decoder structure is symmetrical to the encoder. Fully connected layer: connects latent vectors Map back The feature map of dimension, and reshaped into The shape.

[0090] First-layer upsampling: Use one-dimensional deconvolution with 64 input channels, 32 output channels, kernel size 5, stride 2, padding 2, and output size becomes... , followed by ReLU.

[0091] Second-layer upsampling: one-dimensional deconvolution, 32 input channels, 16 output channels, kernel size 5, stride 2, padding 2, output size becomes... , followed by ReLU.

[0092] Third-layer upsampling: one-dimensional deconvolution, 16 input channels, 3 output channels, kernel size 5, stride 2, padding 2, output size becomes... This is followed by Tanh or linear activation (depending on the data normalization range). The final output is the reconstructed waveform. .

[0093] The reconstruction error between the input waveform and the reconstructed waveform is expressed as mean square error (MSE).

[0094] (ii) Unsupervised training strategies A one-dimensional convolutional autoencoder is trained using an unsupervised learning approach. The training data consists only of historical startup waveforms verified by the rule engine as normal startups (i.e., startup events where rules such as startup timeout, peak abnormality, and contactor abnormality were not triggered, and the motor was running normally). At least 100 normal startup waveforms are extracted from the historical database to form the training set.

[0095] The training objective is to minimize the reconstruction error. No additional physical constraints are used (because the startup waveform contains rich transient information, and the current waveform itself follows the laws of electromagnetic induction). The optimizer is Adam, with a learning rate of 0.001, a batch size of 16, and training epochs until the loss converges (e.g., 50 epochs). Early stopping can be used to prevent overfitting, and the validation set is 10% of the training set.

[0096] (III) Initiating the anomaly scoring calculation For real-time acquired startup waveform data Inputting a trained one-dimensional convolutional autoencoder yields a reconstructed waveform. Calculate the reconfiguration error for this startup:

[0097] To map the reconstruction error to The interval is determined by utilizing the reconstruction error distribution of all normal startup waveforms in the training set. The reconstruction error of each normal waveform in the training set is collected. Calculate its 0.01 quantile. and 0.99 quantile :

[0098] Then start the abnormal scoring Defined as:

[0099] The score is 0 (completely normal) when the reconstruction error is below the lower limit of the normal distribution, and 1 (severely abnormal) when it is above the upper limit of the normal distribution, with a linear mapping in between. This score reflects the degree of waveform distortion of the current startup waveform relative to the normal startup mode.

[0100] To enable the model to adapt to normal waveform drift during startup caused by motor aging or load changes, the one-dimensional convolutional autoencoder is retrained (or incrementally learned) every 30 days using waveform data that has been verified as normal startup data by the rule engine within the most recent period, and the waveforms are recalculated. and .

[0101] In one embodiment of the present invention, based on step S4, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0102] S401: Weighted Summation and Threshold Classification (a) Weighting coefficient setting Set the severity score for each rule. Steady-state anomaly score and startup anomaly scoring Weighting coefficients , and All three conditions must be met:

[0103] The rule severity score has the highest weight, ensuring that anomalies related to electrical safety and functional consistency play a leading role in the final decision; the steady-state anomaly score is the next highest, used to reflect latent degradation during long-term operation; and the startup anomaly score has the lowest weight, serving as an important supplement to anomalies during startup.

[0104] The recommended set of weight values ​​is:

[0105] The weights in this group can be adjusted according to different requirements for safety and sensitivity on site, but the weights of the rules should be kept at their maximum.

[0106] (II) Calculation of Comprehensive Score For each evaluation period (the time period corresponding to the steady-state window, or the comprehensive evaluation after each start-up event), the comprehensive score is calculated using the following formula. :

[0107] in Output from the rule engine in Example 2, Output from the timing autoencoder in Example 3, The output is from the one-dimensional convolutional autoencoder in Example 4. The dimensions of all three have been normalized to the same interval, and after weighted summation... .

[0108] (III) Threshold Interval Division Three preset thresholds , , ,satisfy:

[0109] The range of values ​​for the comprehensive score It is divided into four level ranges. A recommended set of thresholds is:

[0110] according to The anomaly level is determined by the interval in question:

[0111] The above thresholds can be adjusted on-site based on historical false alarm and false negative rates. For example, if on-site safety requirements are extremely high, the thresholds can be adjusted accordingly. Lowering it to 0.65 will cause more anomalies to trigger Level 1 alarms; if reducing unnecessary downtime is required, it can be... Increase to 0.8. Adjustments to each threshold should remain consistent. The order relationship.

[0112] S402: Warning Information Output Based on the anomaly level determined in step S401, output the corresponding warning information. Specific output methods include, but are not limited to: Level 1 alarm: Triggers local audible and visual alarms in the MCC cabinet, and simultaneously sends an emergency shutdown command to the central control room via the communication network. A pop-up window on the monitoring interface displays "Serious fault: Please shut down for maintenance immediately", along with the specific reason for the abnormality (such as the rule with the highest contribution in the rule engine, or the feature with the greatest contribution in the AI ​​score).

[0113] Level 2 alarm: The corresponding circuit is highlighted in yellow on the monitoring interface, a maintenance work order is generated (including fault type, suggested inspection items, and historical trend chart), and pushed to the maintenance personnel's mobile terminal, requiring a planned maintenance to be arranged within one week.

[0114] Level 3 alarm: Mark the loop in blue on the monitoring interface, increase the data acquisition frequency (e.g., temporarily increase steady-state sampling from 1Hz to 2Hz), record abnormal trends and generate daily reports for maintenance personnel to monitor regularly.

[0115] Normal: No warning output, only the comprehensive score is recorded in the log.

[0116] All warning messages include a timestamp, loop number, overall score, and scores for each sub-item. This facilitates subsequent traceability and analysis. Early warning information can be uploaded to the upper-level management system via industrial communication protocols such as Modbus, OPC UA, or MQTT.

[0117] Through the above steps, this invention achieves the organic integration of deterministic judgment by rule engine and unsupervised detection by artificial intelligence, which not only ensures rapid response to electrical safety faults, but also discovers hidden degradation patterns that are difficult to capture by traditional methods. At the same time, the weighted fusion mechanism reduces the impact of false alarms from a single path on the final decision.

[0118] In some embodiments, the anomaly detection system based on MCC cabinet electrical parameters and event logs may include multiple functional modules composed of computer program segments. The computer programs for each program segment in the anomaly detection system based on MCC cabinet electrical parameters and event logs may be stored in the memory of a computer device and executed by at least one processor to perform (see details). Figure 1 (Description) Function for anomaly detection based on MCC cabinet electrical parameters and event logs.

[0119] In this embodiment, the anomaly detection system based on MCC cabinet electrical parameters and event logs can be divided into multiple functional modules according to its functions, such as... Figure 3 As shown. The module referred to in this invention is a series of computer program segments that can be executed by at least one processor and perform a fixed function, and is stored in memory. In this embodiment, the functions of each module will be described in detail in subsequent embodiments.

[0120] The data acquisition module is used to collect monitoring data for each circuit in the MCC cabinet under steady-state operation and during motor startup. The monitoring data includes electrical parameters and event logs. The module also preprocesses the electrical parameters in the monitoring data to generate steady-state time window data and startup waveform data. The first processing module is used to perform anomaly judgment on the steady-state time window data, startup waveform data and event logs using a pre-built rule engine, and output a rule severity score; the rule engine includes electrical safety rules, startup behavior rules and MCC functional consistency rules; The second processing module is used to reconstruct the steady-state time window data using a temporal autoencoder and calculate the steady-state anomaly score; and to reconstruct the startup waveform data using a one-dimensional convolutional autoencoder and calculate the startup anomaly score; wherein the training loss function of the temporal autoencoder includes a physical constraint regularization term based on apparent power conservation. The early warning output module is used to sum the rule severity score, steady-state anomaly score and activation anomaly score according to preset weights, determine the anomaly level based on the threshold range of the weighted sum, and output the corresponding early warning information.

[0121] The same or similar parts between the various embodiments in this specification can be referred to mutually. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple, and the relevant parts can be referred to the description in the method embodiments.

[0122] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or modules may be electrical, mechanical, or other forms.

[0123] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0124] In addition, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0125] Although the present invention has been described in detail with reference to the accompanying drawings and preferred embodiments, the present invention is not limited thereto. Various equivalent modifications or substitutions can be made to the embodiments of the present invention by those skilled in the art without departing from the spirit and essence of the invention, and such modifications or substitutions should all be within the scope of the present invention. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should also be covered within the protection scope of the present invention.

Claims

1. An anomaly detection method based on MCC cabinet electrical parameters and event logs, characterized in that, include: For each circuit of the MCC cabinet, monitoring data is collected under steady-state operating conditions and during motor startup. The monitoring data includes electrical parameters and event logs. The electrical parameters in the monitoring data are preprocessed to generate steady-state time window data and startup waveform data. Anomaly detection is performed on the steady-state time window data, startup waveform data, and event logs using a pre-built rule engine, and a rule severity score is output. The rule engine includes electrical safety rules, startup behavior rules, and MCC functional consistency rules. A temporal autoencoder is used to reconstruct the steady-state time window data and calculate the steady-state anomaly score; a one-dimensional convolutional autoencoder is used to reconstruct the startup waveform data and calculate the startup anomaly score; wherein the training loss function of the temporal autoencoder includes a physical constraint regularization term based on apparent power conservation; The rule severity score, steady-state anomaly score, and activation anomaly score are weighted and summed according to preset weights. The anomaly level is determined based on the threshold range in which the weighted sum falls, and the corresponding warning information is output.

2. The method according to claim 1, characterized in that, For each circuit in the MCC cabinet, monitoring data is collected under steady-state operating conditions and during motor startup, including: Under steady-state operating conditions, the effective values ​​of three-phase current, effective values ​​of three-phase voltage, active power, and power factor are collected at a sampling frequency of 1Hz. When a motor start-up event is detected, the system automatically switches to a sampling frequency of 100Hz to collect the instantaneous waveforms of the three-phase current during the start-up process. The detection condition for the start-up event is that the current of any one phase jumps from less than 0.05 times the rated current to more than 0.2 times the rated current and lasts for more than 20 milliseconds. Action event logs are collected independently with an accuracy of 1 millisecond. The event logs include at least start command events, contactor engagement events, and protection action events, and each event is accompanied by a precise timestamp.

3. The method according to claim 1, characterized in that, The electrical parameters in the monitoring data are preprocessed to generate steady-state time window data and startup waveform data, including: The electrical parameters and event logs are time-aligned using a unified clock source, with an alignment error of no more than 100 milliseconds; Interpolation compensation is performed on missing values ​​in the time series of electrical parameters. For single missing points, the arithmetic mean of data from adjacent time points is used to fill the gaps. For multiple consecutive missing points exceeding a set number, the time window is marked as invalid and the subsequent analysis is skipped. Outliers in the electrical parameter time series were removed, and statistical criteria were used to identify outliers that deviated from the mean by more than three standard deviations, which were then replaced with the mean. The preprocessed steady-state electrical parameters are divided into multiple steady-state time window data according to a sliding window of fixed time length; The instantaneous waveforms of the three-phase current acquired at high speed during startup are extracted as independent startup waveform data for each startup event.

4. The method according to claim 1, characterized in that, The pre-built rule engine is used to perform anomaly detection on the steady-state time window data, startup waveform data, and event logs, and outputs a rule severity score, including: The rule engine is used to match the steady-state time window data, startup waveform data, and event logs with the corresponding rule conditions, and the successfully matched triggering rules and their corresponding severity scores are recorded. The maximum severity score among all triggering rules is used as the final rule severity score in the output.

5. The method according to claim 4, characterized in that, The electrical safety rules include: The severity is determined by the ratio of the maximum three-phase current to the rated current. When the ratio is greater than 1.20, the severity is 1.0, and when the ratio is between 1.05 and 1.20, the severity is 0.

7. When the average current is greater than 0.2 times the rated current, the minimum phase current is less than 0.1 times the rated current, and the ratio of the maximum phase current to the minimum phase current is greater than 5, the severity is 1.

0. Three-phase imbalance rule: The severity is determined based on the degree of imbalance of the three-phase current. When the degree of imbalance is greater than 0.1, the severity is 0.

7. When the degree of imbalance is between 0.08 and 0.1, the severity is 0.

4. The degree of imbalance is defined as the difference between the maximum phase current and the minimum phase current divided by the average value of the three-phase current. The startup behavior rules include: The starting time limit is dynamically calculated based on the motor's rated power. When the actual measured starting time exceeds the limit, the severity is 0.

7. Based on the statistical mean and standard deviation of the peak current of the most recent 20 successful starts, if the peak current of this start deviates from the mean by more than 3 times the standard deviation, the severity is 0.7; if the deviation is between 2 and 3 times the standard deviation, the severity is 0.

4. The MCC functional consistency class rules include: When both the start command event and the contactor engagement event occur, but the time from the command issuance to current establishment exceeds 1 second and the current is still less than 0.1 times the rated current afterward, the severity is 1.0; When no protection action event occurs but the fault current exceeds 1.2 times the setting value and lasts for more than 0.5 seconds, the severity is 0.7; When a protection action event occurs but the maximum current before and after the action is less than 1.05 times the setting value, the severity is 0.

7.

6. The method according to claim 1, characterized in that, The steady-state time window data is reconstructed using a time-series autoencoder, and a steady-state anomaly score is calculated, including: The time-series autoencoder consists of a time-series encoder and a time-series decoder connected in series. The time-series encoder compresses the input steady-state time window data sequence into a low-dimensional latent vector, and the time-series decoder reconstructs the low-dimensional latent vector into an output sequence with the same dimension as the input sequence. By minimizing the reconstruction error between the input and output sequences, the time-series autoencoder learns the timing pattern under normal operating conditions. Both the time-series encoder and the time-series decoder use long short-term memory networks as basic units. The training method adopts an unsupervised learning approach, and the training data only uses historical steady-state window data that has been verified by the rule engine to be free of anomalies. The training objective is to minimize the reconstruction error between the input and the reconstructed output. The loss function of the time-series self-time encoder adds a physical constraint regularization term to the conventional reconstruction error, and the physical constraint regularization term is based on the principle of apparent power conservation. The steady-state time window data collected in real time is input into the trained time-series self-time encoder. The reconstruction error between the input and the reconstruction output is calculated. The reconstruction error is normalized according to the specified quantile of the reconstruction error distribution of the training set and mapped to the interval between zero and one to obtain the steady-state anomaly score.

7. The method according to claim 6, characterized in that, The loss function of the temporal self-temporal encoder adds a physical constraint regularization term to the conventional reconstruction error, including: Based on the principle of apparent power conservation, the reconstructed active power and reactive power are extracted at the output of the time-series decoder of the time-series autoencoder, and the reconstructed apparent power is calculated based on the reconstructed active power and reactive power. Calculate the measured apparent power based on the measured three-phase voltage and three-phase current within the same time window; The absolute value of the difference between the reconstructed apparent power and the measured apparent power is used as the physical constraint regularization term. The larger the physical constraint regularization term is, the more the reconstruction result violates the physical law of apparent power conservation, and the heavier the penalty on the loss function. The total loss function of the temporal autoencoder is a weighted sum of the conventional reconstruction error and the physical constraint regularization term.

8. The method according to claim 1, characterized in that, The startup waveform data is reconstructed using a one-dimensional convolutional autoencoder, and a startup anomaly score is calculated, including: The one-dimensional convolutional autoencoder consists of a convolutional encoder and a convolutional decoder connected in series. The convolutional encoder contains several one-dimensional convolutional layers and pooling layers, which are used to extract local features from the input start waveform data layer by layer and reduce the time dimension, and output a low-dimensional latent feature vector. The convolutional decoder contains several one-dimensional deconvolutional layers and upsampling layers, which are used to gradually restore the latent feature vector into a reconstructed waveform with the same length and number of channels as the input waveform. The input of the one-dimensional convolutional autoencoder is the extracted three-phase current instantaneous waveform data of a single startup process. Each waveform contains three channels corresponding to the three phases, and the length of each channel is the number of sampling points within a fixed time period. The training method adopts an unsupervised learning approach, and the training data only uses historical startup waveform data that has been verified by the rule engine as normal startup. The training objective is to minimize the reconstruction error between the input waveform and the reconstructed waveform. The real-time acquired startup waveform data is input into the trained one-dimensional convolutional autoencoder. The reconstruction error between the input waveform and the reconstructed waveform is calculated. The reconstruction error is normalized according to the specified quantile of the reconstruction error distribution of the training set and mapped to the interval between zero and one to obtain the startup anomaly score.

9. The method according to claim 1, characterized in that, The rule severity score, steady-state anomaly score, and initiation anomaly score are weighted and summed according to preset weights. The anomaly level is determined based on the threshold range in which the weighted sum falls, including: Weight coefficients were set for rule severity score, steady-state anomaly score and initiation anomaly score respectively, with the sum of the weights of the three being 1. The rule severity score had the largest weight, followed by the steady-state anomaly score, and the initiation anomaly score had the smallest weight. The rule severity score, steady-state anomaly score, and initiation anomaly score are multiplied by their respective weights, and then added together to obtain the comprehensive score. The comprehensive score range is divided into four levels by three preset thresholds: when the comprehensive score is greater than or equal to the first threshold, it is determined as a level 1 alarm; when the comprehensive score is greater than or equal to the second threshold but less than the first threshold, it is determined as a level 2 alarm; when the comprehensive score is greater than or equal to the third threshold but less than the second threshold, it is determined as a level 3 alarm; when the comprehensive score is less than the third threshold, it is determined as a normal state. The first threshold, the second threshold, and the third threshold are preset values, and the first threshold is greater than the second threshold, which is greater than the third threshold.

10. An anomaly detection system based on MCC cabinet electrical parameters and event logs, characterized in that, include: The data acquisition module is used to collect monitoring data for each circuit of the MCC cabinet under steady-state operation and during motor startup. The monitoring data includes electrical parameters and event logs. The module also preprocesses the electrical parameters in the monitoring data to generate steady-state time window data and startup waveform data. The first processing module is used to perform anomaly judgment on the steady-state time window data, startup waveform data and event logs using a pre-built rule engine, and output a rule severity score; the rule engine includes electrical safety rules, startup behavior rules and MCC functional consistency rules; The second processing module is used to reconstruct the steady-state time window data using a time-series autoencoder and calculate the steady-state anomaly score; and to reconstruct the startup waveform data using a one-dimensional convolutional autoencoder and calculate the startup anomaly score. The training loss function of the temporal autoencoder includes a physical constraint regularization term based on apparent power conservation. The early warning output module is used to sum the rule severity score, steady-state anomaly score and activation anomaly score according to preset weights, determine the anomaly level based on the threshold range of the weighted sum, and output the corresponding early warning information.

Images