PLC electric control cabinet control system and method

Abstract

The invention discloses a PLC electric control cabinet control method, and belongs to the technical field of electrical control. The method specifically comprises the following steps: S1, multi-dimensional parameter fusion acquisition: acquiring multi-dimensional data through a multi-type sensor, setting a hierarchical acquisition period according to parameter types, storing multi-source data in a classified and layered manner, and performing traceability judgment on abnormal data; s2, parameter preprocessing and feature extraction: different noise reduction algorithms are adopted for different signal types, and depth features of data are extracted through multi-step processing; electrical parameters, physical state parameters, environmental parameters and time sequence label parameters are collected through multiple types of sensors, the equipment operation state and external influence factors are comprehensively captured, compared with traditional single parameter collection, the real working condition of the system can be reflected more accurately, and complete data support is provided for subsequent control and diagnosis. And a hierarchical acquisition period is set according to parameter types, so that resource waste or key data omission caused by one-step acquisition is avoided.

Description

Technical Field

[0001] This invention relates to the field of electrical control technology, and in particular to a PLC electrical control cabinet control system and method. Background Technology

[0002] PLC-based electrical control is one of the core technologies in current industrial automation. It uses a programmable logic controller (PLC) as its core, and through the collaboration of hardware modules (such as input / output modules, high-speed counting modules, and analog signal processing modules) and software programs (such as ladder diagrams, function block diagrams, and structured text), it achieves logical, timing, and process control of electrical actuators such as motors, valves, and transmission equipment. This method boasts advantages such as modular hardware design (flexible expansion of input / output points according to control requirements), strong program portability (the same brand of PLC program can be adapted between different models of equipment), excellent anti-interference capabilities (possessing hardware protection such as opto-isolation and electromagnetic shielding, as well as software filtering functions), and fast real-time response speed (mainstream PLC scan cycles can be as low as milliseconds). It has been widely applied in manufacturing production lines, energy generation systems, transportation equipment (such as subway traction systems), and chemical process control, becoming a key technological support for driving industrial automation upgrades. However, existing PLC control technologies have several limitations. Insufficient real-time performance makes it difficult to quickly respond to high-speed dynamic changes in electrical systems, leading to significant control deviations. Summary of the Invention

[0003] Purpose of the invention: The purpose of this invention is to provide a PLC electrical control cabinet control system and method, which can solve the problem of insufficient real-time performance in the prior art.

[0004] Technical Solution: To solve the above-mentioned technical problems, according to one aspect of the present invention, more specifically, a PLC electrical control cabinet control method, the method specifically includes the following steps:

[0005] S1. Multi-dimensional parameter fusion acquisition: Multi-dimensional data is acquired through multiple types of sensors, hierarchical acquisition cycles are set according to parameter type, multi-source data is stored in categories and layers, and abnormal data is traced and judged.

[0006] S2. Parameter preprocessing and feature extraction: Different noise reduction algorithms are used for different signal types, and deep features of the data are extracted through multi-step processing.

[0007] S3. Dynamic control model construction: A dynamic control model is constructed based on deep feature data. The working conditions are identified based on key deep features. The model is dynamically adapted and optimized by combining the working conditions and fault hazard features. At the same time, model verification is strengthened and deep feature deviation verification is added.

[0008] S4. Multi-objective adaptive adjustment: Based on deep features, the control target priority matrix is ​​dynamically adjusted, deep feature refinement calculation is incorporated, a distributed model prediction and spatiotemporal synchronous verification collaborative upgrade strategy is adopted, the adjustment amount increase is combined with early warning, and fault hazard characteristics are associated.

[0009] S5. Real-time fault diagnosis and self-healing control: The cloud-based fault feature library matches deep features to construct deep feature fault templates. It adopts a dual diagnostic algorithm of deep feature matching and mechanism rule judgment. It combines fault type and deep feature level to formulate a graded self-healing control strategy, extends the fault recovery verification cycle, and conducts verification in two stages.

[0010] S6. Closed-loop optimization and iterative upgrade: The system control effect is quantitatively evaluated periodically. The model parameter optimization algorithm is upgraded by adopting a deep feature-oriented particle swarm optimization algorithm. The optimization strategy library is classified and stored according to the working condition type and deep feature mode. The remote upgrade function is expanded and online upgrade of the deep feature extraction algorithm is supported.

[0011] Furthermore, in step S1, the collected data is divided into electrical parameters, physical state parameters, environmental parameters, and time-series label parameters. The electrical parameters include: real-time current value, voltage value, power factor, power, torque, leakage current, and insulation resistance. The physical state parameters include: equipment temperature, vibration frequency, actuator position, pressure value, and torque fluctuation. The environmental parameters include: ambient temperature, humidity, dust concentration, and electromagnetic interference intensity. The time-series label parameters include: collection timestamp, equipment number, and operating condition identifier, which are initialized to be labeled. For instantaneous signals exceeding the preset range, the parameter change trend of three adjacent collection cycles is compared for judgment. If it is determined to be an interference signal, it is directly discarded. If it is determined to be a precursor to a knowledge fault, it is marked as an abnormality to be verified and retained. When packaging the data of a row to be verified, the judgment result is attached and transmitted to the PLC main controller.

[0012] Furthermore, step S2 specifically includes the following steps:

[0013] S21. Optimize the noise reduction algorithm for different signal types: Vibration signal and torque signal adopt 5-layer wavelet noise reduction + empirical mode decomposition, 10 intrinsic mode functions at the boundary, and reconstruct the signal after removing the noisy intrinsic mode functions. Current and voltage signals adopt Kalman filtering + sliding window averaging, with a window size of 5 acquisition cycles. The standardization process is interval standardization + deviation correction. First, the parameters are mapped to the interval [1, 1] according to the standardization value = (actual value, rated value) / (limit value, rated value). Then, the standardization result is corrected according to the equipment aging coefficient calculated from the running time and historical fault data.

[0014] S22. For electrical parameters: calculate the average, peak, valley, variance, root mean square, and crest factor over 10 acquisition cycles. For current parameters, additionally calculate the harmonic content and the proportion of the 210th harmonic. For insulation resistance parameters, calculate the attenuation rate over 1 minute. For physical state parameters: calculate the average rate of change, cumulative temperature rise, and temperature difference between different parts of the equipment over 10 seconds. For vibration parameters, calculate the peak acceleration and the root mean square multiplied by time. Output a 32-dimensional basic statistical feature vector.

[0015] S23. Perform sliding window trend analysis on basic statistical features. Set the window size to 3 statistical periods, calculate the slope of feature changes, such as the root mean square slope of current, the slope of temperature change rate, and trend consistency. Compare whether the three consecutive windows change in the same direction. Mark the points with an absolute slope value > 0.5 / period as "trend breakpoints". Spatial correlation features: calculate the correlation coefficient and coordination deviation of parameters between different devices, output an 18-dimensional spatiotemporal correlation feature vector, and merge it with the basic statistical feature vector to form a 50-dimensional feature set.

[0016] S24. Based on the equipment operation mechanism, extract fault precursor features from the 50-dimensional feature set. For motor equipment: extract the bearing characteristic frequency harmonics as the proportion of the 23rd harmonic in the 100-500Hz frequency band, the stator current imbalance as the ratio of the three-phase current deviation to the average value, and the coupling coefficient of torque fluctuation and speed. For circuit equipment: extract the cable leakage current growth rate as the growth rate every 10 seconds, the insulation resistance temperature coefficient as the resistance change rate for every 1℃ increase in temperature, and the power factor mutation frequency. For actuator equipment: extract the lag coefficient of position deviation and voltage fluctuation as the ratio of position deviation response time to voltage change time, and the valve opening adjustment dead zone as the minimum difference between the command change and the actual opening change. Output a 22-dimensional fault hazard feature vector.

[0017] S25. Using random forest, the correlation between 72-dimensional features and control objectives, including control accuracy, energy consumption, equipment lifespan, and failure risk, is calculated. Features with a correlation greater than 0.6 are retained. The mutual information entropy between the retained features is calculated, and redundant features are eliminated. For features with a mutual information entropy greater than 0.8, the one with the higher correlation is retained. Finally, 3040-dimensional key deep features are retained.

[0018] Furthermore, step S3 specifically includes the following steps:

[0019] S31. The basic model library includes: PID control model, fuzzy control model, predictive control model, spatiotemporal correlation compensation model, and fault hazard suppression model. When loading, the initial parameters of the model with the highest matching degree with the current operating condition are called first.

[0020] S32. Using K-means, the deep features of historical operating data are clustered into 5 typical operating conditions: Light load steady state (load rate <30%, parameter fluctuation <5%), Medium load transition state (load rate 30%-70%, parameter fluctuation 5%-15%), Heavy load steady state (load rate >70%, parameter fluctuation <5%), Abrupt load state (load change rate >20% / s, torque abrupt change rate >30% / s, current harmonic abrupt change (second harmonic proportion >5% / cycle), and Abnormal approach state (key parameters approaching alarm threshold). An SVM-trained operating condition classifier is used to classify the real-time deep feature vectors, outputting the operating condition type and confidence level. When the confidence level is <80%, a secondary identification is initiated: 10 additional cycles of data are collected and recalculated.

[0021] S33. Adjust the model based on the fault hazard characteristics in the working condition type and depth characteristics: Light load steady state: adopt a simplified PID model with low energy consumption optimization; Medium load transition state: adopt a fuzzy PID spatiotemporal correlation compensation composite model; Heavy load steady state: adopt a PID model with feedforward with fault hazard suppression; Abrupt load state: adopt a predictive control model with torque compensation; Abnormal approach state: adopt a fault hazard suppression model with conservative control.

[0022] S34. Calculate the change in key depth features after the model output. If the control deviation is greater than the threshold or the depth feature deviation threshold for three consecutive cycles, reload the basic model and adapt the parameters.

[0023] Furthermore, in step S4, the priority optimization matrix includes: normal operating conditions: priority from high to low is control accuracy, energy consumption, and response speed; high temperature operating conditions: priority from high to low is equipment protection, control accuracy, and energy consumption; fault-prone operating conditions: priority from high to low is fault suppression, equipment protection, and accuracy control; multi-device collaborative operating conditions: priority from high to low is spatiotemporal synchronization, control accuracy, and energy consumption.

[0024] Furthermore, in step S4, if the adjustment amount exceeds the safety threshold and the corresponding potential fault characteristics are within the normal range, then only the amplitude is limited and recorded; if the adjustment amount exceeds the safety threshold and the corresponding potential fault characteristics are not within the normal range, then while limiting the amplitude, a potential fault warning is triggered and pushed to the monitoring terminal for early intervention.

[0025] Furthermore, step S5 specifically includes the following steps:

[0026] The S51 and PLC controllers synchronize the latest fault feature library from the cloud server every hour, and combine it with the fault records of the past 24 hours in the local operation history data to supplement and match deep fault features to build deep feature fault templates. Each fault type corresponds to 35 core deep feature thresholds.

[0027] S52. Calculate the cosine similarity between the real-time 22-dimensional fault hazard features and the fault template, and mark those with a cosine similarity greater than the threshold as suspected faults.

[0028] S53. Verification based on equipment mechanism rules: if the verification passes, it is determined to be a confirmed fault; otherwise, it is marked as pending observation. The diagnosis cycle is synchronized with the data acquisition cycle.

[0029] S54. Develop a self-healing strategy based on the fault type and depth of the fault characteristics, and verify the fault recovery in two stages: monitor whether the basic parameters have returned to the normal range; monitor whether the fault hazard characteristics have returned to the safe range. If both stages are met, the self-healing is considered successful; otherwise, initiate secondary processing, such as remote operation and maintenance intervention.

[0030] Furthermore, step S6 specifically includes the following steps:

[0031] S61. Based on deep features, evaluation indicators are formed, including: control accuracy, stability, energy consumption, failure rate, spatiotemporal synchronization, and failure hazard rate. Evaluation is performed at fixed intervals to generate evaluation reports.

[0032] S62. Based on the evaluation metrics, the model is iterated using the particle swarm optimization algorithm. Each iteration ensures that all evaluation metrics meet the standards and at least two metrics improve to the threshold.

[0033] S63. Each working condition is further subdivided into 34 deep feature patterns. When invoking a strategy, the working condition is matched first, and then the corresponding optimization strategy is invoked based on the real-time deep feature pattern.

[0034] S64. The cloud platform pushes a new feature extraction algorithm package, which adopts score update and breakpoint resume technology. After the update is completed, it is verified through offline simulation and tested with historical data. The algorithm is then enabled after the verification is successful.

[0035] According to another aspect of the present invention, a PLC electrical control cabinet control system is provided, characterized in that: the system is used to implement the above-described PLC electrical control cabinet control method, including: a multi-dimensional parameter fusion acquisition module, a parameter preprocessing and feature extraction module, a dynamic identification module, an adaptive adjustment module, a real-time fault diagnosis and self-healing module, and an optimization iteration module;

[0036] Multi-dimensional parameter fusion acquisition module: used to acquire multi-dimensional data through multiple types of sensors, set hierarchical acquisition cycles according to parameter type, classify and store multi-source data in layers, and trace and determine the source of abnormal data;

[0037] The parameter preprocessing and feature extraction module is used to extract deep features of data by employing different noise reduction algorithms for different signal types through multi-step processing.

[0038] Dynamic identification module: used to build dynamic control models based on deep feature data, identify working conditions based on key deep features, dynamically adapt and optimize the model by combining working conditions and fault hazard features, and strengthen model verification by adding deep feature deviation verification.

[0039] Adaptive adjustment module: used to dynamically adjust the control target priority matrix based on deep features, incorporate deep feature refinement calculation, adopt a distributed model prediction and spatiotemporal synchronous verification collaborative upgrade strategy, combine the adjustment amount increase with early warning, and associate fault hazard characteristics;

[0040] Real-time fault diagnosis and self-healing module: It is used to construct deep feature fault templates by matching deep features with a cloud fault feature library. It adopts a dual diagnostic algorithm of deep feature matching and mechanism rule judgment, and formulates a graded self-healing control strategy by combining fault type and deep feature level to extend the fault recovery verification cycle. The verification is carried out in two stages.

[0041] The optimization iteration module is used to perform quantitative evaluation of the system control effect at regular intervals. It adopts a deep feature-guided particle swarm optimization algorithm to upgrade the model parameter optimization algorithm, stores the optimization strategy library according to the working condition type and deep feature mode, expands the remote upgrade function, and supports online upgrade of the deep feature extraction algorithm.

[0042] Beneficial effects:

[0043] 1. By collecting electrical parameters, physical state parameters, environmental parameters, and time-series tag parameters through multiple types of sensors, the system comprehensively captures the equipment's operating status and external influencing factors. Compared to traditional single-parameter acquisition, this method more accurately reflects the system's true operating condition, providing complete data support for subsequent control and diagnosis. Hierarchical acquisition cycles are set according to parameter type to avoid resource waste or missing key data caused by a "one-size-fits-all" approach. Multi-source data is stored in a categorized and layered manner for easy retrieval and traceability analysis, improving data processing efficiency. By comparing the parameter change trends of three adjacent acquisition cycles, the system accurately distinguishes between "interference signals" and "fault precursors"—interference signals are directly eliminated to avoid affecting control decisions; fault precursors are marked as "abnormalities to be verified" and retained, with the judgment result transmitted to the PLC main controller, enabling early detection of potential faults and laying the groundwork for subsequent fault diagnosis.

[0044] 2. Match exclusive noise reduction solutions according to different signal characteristics. Compared with traditional single filtering, it can more efficiently remove noise and ensure the accuracy of the original data. Through multi-step extraction of "basic statistical features + spatio-temporal correlation features + potential fault features", the original data is transformed into a "feature language" that can directly support control and diagnosis, solving the problem of "redundant but low-value data" caused by traditional PLCs relying only on the original data. Calculate the correlation between features and control objectives through random forests, and remove redundant features through mutual information entropy. Finally, 3,040 key features are retained. This not only reduces the complexity of subsequent model calculations but also ensures the core control value of the retained features, providing accurate input for the dynamic control model.

[0045] 3. The basic model library covers 5 core models such as PID, fuzzy control, and predictive control. Combined with the Kmeans clustering and SVM working condition classifier, it can judge the current working condition type in real time, avoiding control deviations caused by traditional PLCs using "a single model for all working conditions". When adapting the model, combine the potential fault features in the deep features. Compared with traditional PLCs that only focus on "control objectives" and ignore "fault risks", it can suppress potential faults synchronously during the control process, achieving a double balance of "control accuracy" and "equipment protection". Calculate the change amount of the key deep features after the model outputs. If the control quantity deviation or feature deviation exceeds the threshold for 3 consecutive cycles, immediately reload the basic model and adapt the parameters to form "real-time monitoring and error correction" of the model operation, avoiding control failures caused by model drift and improving the robustness of the control system.

[0046] 4. Based on the deep features, update the priority matrix in real time, solving the problem that traditional PLCs with "fixed priorities" cannot handle dynamic scenarios. Adopt the collaborative strategy of "distributed model prediction" and "spatio-temporal synchronization verification", which is especially suitable for multi-device linkage scenarios, avoiding the "spatio-temporal asynchronization" problem during multi-device control by traditional PLCs and improving the overall control accuracy of the system. If the adjustment amount exceeds the safety threshold, combined with the potential fault feature status, perform differential processing: when the feature is normal, only limit the amplitude and record; when the feature is abnormal, trigger a potential fault warning synchronously. This not only avoids excessive warnings from interfering with operation and maintenance but also can intervene in a timely manner when abnormal adjustments are accompanied by fault risks, preventing small problems from expanding into major faults.

[0047] 5. The PLC synchronizes the latest fault feature database to the cloud every hour and combines it with local 24-hour historical fault records to construct a "deep feature fault template." Compared to traditional local fault databases, this allows for more accurate matching of new faults, reducing missed diagnoses and misdiagnoses. It employs a dual algorithm of "deep feature cosine similarity matching + equipment mechanism rule verification," leveraging data-driven rapid diagnosis while avoiding misjudgments caused by "mismatched data" through mechanism rules, ensuring accurate fault diagnosis. A self-healing strategy is formulated based on fault type and feature level, and the recovery effect is verified in two stages. Compared to traditional PLCs that "only alarm and wait for manual intervention after a fault," this can achieve automatic repair of over 80% of minor faults, shortening downtime and reducing manual maintenance costs. Secondary remote intervention is only initiated when self-healing fails, optimizing the allocation of maintenance resources.

[0048] 6. Based on deep features, six quantitative evaluation indicators are constructed, and evaluation reports are generated regularly. The model parameters are upgraded using a "deep feature-guided particle swarm optimization algorithm," addressing the problem of "fixed parameters and performance degradation with equipment aging" in traditional PLCs, ensuring that the system control effect remains optimal in the long term. An optimization strategy library is stored and categorized by "operating condition type + deep feature mode," allowing for precise matching of scenarios during invocation. Simultaneously, it supports cloud-based push of new feature extraction algorithms, enabling algorithm upgrades without on-site disassembly, solving the problem of "fixed algorithms and difficulty in adapting to new scenarios" in traditional PLCs, and extending the system's technical lifecycle. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the method. Detailed Implementation

[0050] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] Example 1

[0052] Step 1: Multi-dimensional parameter fusion and acquisition:

[0053] Electrical parameters: Real-time current value (range 0-500A) is collected by a current sensor, voltage value (380V±10%) is collected by a voltage sensor, power factor (0.8-1.0) and active power are collected by a power analyzer, output torque (0-500N·m) is collected by a torque sensor, leakage current (≤10mA) is collected by a leakage current detector, and insulation resistance (≥100MΩ) is collected by an insulation resistance meter.

[0054] Physical state parameters: Infrared temperature sensor collects the stator and rotor temperature of the motor (normal ≤120℃), vibration sensor collects the vibration frequency at the motor bearing (101000Hz), encoder collects the position of the actuator (such as the conveyor belt drive wheel) (accuracy ±0.1mm), pressure sensor collects the pressure value of the motor lubrication system (0.2-0.5MPa), and torque fluctuation detector collects the torque fluctuation amplitude (≤5% of the rated value).

[0055] Environmental parameters: Temperature and humidity sensors collect ambient temperature (15-35℃) and humidity (40%-60%) in the workshop; dust sensors collect dust concentration (≤10mg / m³). 3 The electromagnetic interference detector collects the electromagnetic interference intensity (≤50dBμV / m).

[0056] Time stamp parameters: Automatically generate collection timestamp (accurate to milliseconds), device number (e.g., M001), and operating condition identifier (initialized to "to be identified").

[0057] The data acquisition cycle is set according to the importance of the parameters: key electrical parameters such as current, voltage, and torque are acquired once every 100ms; physical state parameters such as temperature and vibration are acquired once every 500ms; environmental parameters are acquired once every 5s; time-series label parameters are generated synchronously with each acquisition. Multi-source data is stored in the PLC local database according to "parameter type, acquisition time, and device number", and is also backed up to the workshop server. Abnormal data tracing and judgment rules: when a parameter exceeds the preset range (such as the current suddenly rising to 600A), the parameter change trend of the three adjacent acquisition cycles (within 300ms) is compared. If the change trend is "sudden rise and fall" (such as 600A→380A→390A), it is judged as an interference signal such as electromagnetic interference and is directly rejected; if the change trend is "continuous deviation" (such as 380A→420A→450A), it is judged as a fault precursor, marked as "abnormal to be verified" and retained. When packaging the data, the judgment result of "suspected overload precursor" is attached and transmitted to the PLC main controller.

[0058] The second step is parameter preprocessing and feature extraction:

[0059] (I) Vibration and Torque Signals: Five-layer wavelet denoising (wavelet basis function selected as db4) combined with empirical mode decomposition was used to obtain 10 intrinsic mode functions (IMFs). Noisy IMFs (such as IMFs with a correlation <0.3 with the original signal) were removed through correlation analysis. The error of the reconstructed signal was controlled within ±2%. Current and Voltage Signals: Kalman filtering (state equation set as linear time-invariant system) combined with sliding window averaging (window size is 5 acquisition cycles, i.e., 500ms) was used. After filtering, the current and voltage fluctuation amplitude was reduced to ±1%. Standardization Processing: All parameters were mapped to the interval [1, 1] according to the formula "standardized value = (actual value / rated value) / (limit value / rated value)". For example, the rated current of the motor is 380A, the limit value is 500A, and when the actual current is 450A, the standardized value = (450 / 380) / (500 / 380) ≈ 0.58. Then, based on the equipment aging factor (0.92 calculated from 5000 hours of motor operation and historical fault data), the final standardized value is 0.58 × 0.92 ≈ 0.53.

[0060] (II) For electrical parameters, calculate the statistical characteristics over 10 acquisition cycles (1s): Current, Voltage: Average value (e.g., average current 385A), peak value (420A), valley value (370A), variance (25A). 2 =), root mean square (386A), crest factor (420 / 386≈1.09); the current additionally calculates the proportion of the 210th harmonic (e.g., the 2nd harmonic accounts for 1.2%, the 3rd harmonic accounts for 0.8%). Insulation resistance: calculate the attenuation rate within 1 minute (e.g., from 120MΩ to 118MΩ, attenuation rate = (120118) / 120×100%≈1.67%). For physical state parameters, the calculation characteristics are as follows: Temperature: average rate of change within 10 seconds (e.g., the stator temperature rises from 80℃ to 82℃, rate of change = (8280) / 10 = 0.2℃ / s), cumulative temperature rise (2℃), temperature difference between stator and rotor (5℃). Vibration: peak acceleration (15m / s²) 2 Vibrational energy (root mean square acceleration 12 m / s²) 2 (×Acquisition time 1s = 12m·s). The final output is a 32-dimensional basic statistical feature vector, covering all the above feature values.

[0061] (III) Using sliding window trend analysis (window size of 3 statistical periods, i.e., 3s), the slopes of characteristic changes were calculated: Root mean square slope of current: The root mean square currents for the first 3s were 386A, 388A, and 390A respectively, with a slope of (390 / 386) / 3 ≈ 1.33A / s. Slope of temperature change rate: The temperature change rates for the first 3s were 0.2℃ / s, 0.25℃ / s, and 0.3℃ / s respectively, with a slope of (0.3 / 0.2) / 3 ≈ 0.033℃ / s. 2Trend Consistency Judgment: If the root mean square current of three consecutive windows shows an upward trend, it is judged as "consistent trend"; points with an absolute slope value > 0.5 / cycle (e.g., the root mean square current slope of a certain cycle is 2.1A / s > 0.5A / s) are marked as "trend abrupt change points". Spatial Correlation Features: Calculate the current correlation coefficient (0.85) and torque coordination deviation (5%) between the motor and the adjacent conveyor belt motor, output an 18-dimensional spatiotemporal correlation feature vector, and merge it with the 32-dimensional basic statistical feature vector to form a 50-dimensional feature set.

[0062] (IV) Based on the motor operating mechanism, 22-dimensional fault hazard features are extracted from the 50-dimensional feature set: Bearing characteristic frequency harmonics: The proportion of second harmonics in the 100-500Hz frequency band is 1.5%, and the proportion of third harmonics is 1.1%. Stator current imbalance: The three-phase currents are 385A, 382A, and 378A respectively, with an average value of 381.7A. The imbalance is approximately 1.83% (385-378) / 381.7 × 100%. Coupling coefficient between torque fluctuation and speed: The torque fluctuation amplitude is 3%, and the speed fluctuation is 1%. The coupling coefficient is 3% / 1% = 3. Cable leakage current growth rate: The leakage current increases from 5mA to 5.2mA every 10 seconds. The growth rate is approximately 4% (5.25) / 5 × 100%.

[0063] (V) A random forest algorithm (100 decision trees) is used to calculate the correlation between 72-dimensional features (50-dimensional feature set + 22-dimensional fault hazard features) and control objectives (control accuracy ±0.5%, energy consumption ≤5kW·h / h, equipment life ≥10000h, fault risk ≤0.1% / h). Features with a correlation >0.6 are retained (such as root mean square current, temperature change rate, bearing harmonic ratio, etc.). The mutual information entropy between the retained features is calculated, and redundant features with a mutual information entropy >0.8 are removed (such as "peak current" and "root mean square current" with a mutual information entropy of 0.85, retaining the "root mean square current" with higher correlation). Finally, 35 key deep features are retained.

[0064] The third step is to construct the dynamic control model:

[0065] (I) The basic model library contains 5 core models: PID control model: proportional coefficient Kp = 2.5, integral time Ti = 0.5s, derivative time Td = 0.1s. Fuzzy control model: input variables are "current deviation" and "temperature deviation", output variable is "control quantity increment", fuzzy rule library contains 25 rules. Predictive control model: 5-step prediction in the time domain, 2-step control in the time domain, constraints are current ≤ 500A, temperature ≤ 120℃. Spatiotemporal correlation compensation model: compensation coefficient is dynamically adjusted according to the parameter deviation of adjacent equipment (range 0.8-1.2). Fault hazard suppression model: for excessive leakage current, the suppression coefficient is set to 0.95. Based on the current initial motor operating condition (load rate 35%), the initial parameters of the fuzzy control model are preferentially called.

[0066] (II) The historical operating data was clustered into 5 typical operating conditions using the K-means algorithm. Based on the actual operating data of this scenario: load rate 35% (30%-70% range), parameter fluctuation 8% (5%-15% range), it was determined to be "medium load transition state". An SVM was used to train the operating condition classifier (kernel function RBF) to classify the real-time 35-dimensional deep feature vector, and the operating condition type "medium load transition state" and confidence level 92% (≥80%, no secondary recognition required).

[0067] (III) To address the characteristics of the "medium load transient state" and potential faults (leakage current growth rate of 4%), a "fuzzy PID spatiotemporal correlation compensation composite model" is adopted: The fuzzy control module handles the nonlinear deviations of current and temperature, and outputs the initial control quantity. The PID module finely adjusts the initial control quantity to reduce static error. The spatiotemporal correlation compensation module dynamically corrects the control quantity based on the parameters of adjacent motors (e.g., if the current of an adjacent motor increases, the compensation coefficient is adjusted to 1.1).

[0068] (iv) Calculate the changes in key depth features after model output: The root mean square current decreases from 390A to 385A (deviation 5A), and the temperature change rate decreases from 0.3℃ / s to 0.22℃ / s (deviation 0.08℃ / s), neither of which exceeds the threshold (current deviation threshold 8A, temperature change rate deviation threshold 0.1℃ / s), and the model is running normally; if the current deviation is 9A for three consecutive cycles (exceeding the threshold), then reload the initial parameters of the fuzzy control model and adapt the new Kp=2.8, Ti=0.4s, Td=0.12s.

[0069] Step 4, Multi-objective adaptive adjustment:

[0070] (I) Control target priority matrix adjustment: The current operating condition is "medium load transition state", with no high temperature or fault risks. Match the "normal operating condition" priority matrix: First priority: control accuracy (ensure motor speed deviation ≤ ±0.5%). Second priority: energy consumption (control motor input power ≤ 45kW). Third priority: response speed (speed adjustment response time ≤ 0.5s).

[0071] (II) Execution of Adjustment Strategy: A collaborative strategy of "distributed model prediction + spatiotemporal synchronization verification" is adopted: Distributed model prediction: The motor control is divided into three sub-modules: "current regulation", "temperature regulation" and "speed regulation". The parameter changes within 5 steps are predicted for each module, and the control quantity of each sub-module is output. Spatiotemporal synchronization verification: The spatiotemporal consistency of the control quantities of the three sub-modules is compared (e.g., current regulation needs to be completed within 0.2s, and speed regulation needs to respond within 0.3s). After correcting the deviation, the total adjustment quantity is output (e.g., current control quantity +0.8A, speed control quantity +5r / min).

[0072] (III) Handling of Abnormal Adjustment: If the adjustment exceeds the safety threshold (e.g., current adjustment +10A, safety threshold +8A): If the corresponding fault hazard characteristics (e.g., leakage current growth rate 4%, normal range ≤5%) are within the normal range, only a limit is applied (current adjustment reduced to +8A) and recorded in the log. If the leakage current growth rate rises to 6% (exceeding the normal range), a hazard warning is triggered simultaneously with the limit, and pushed to the workshop monitoring terminal, prompting maintenance personnel to check the cable insulation status.

[0073] Step 5: Real-time fault diagnosis and self-healing control:

[0074] (i) The PLC controller synchronizes the latest fault feature library (including 20 types of fault features such as motor overload, bearing wear, and cable insulation aging) from the cloud server every hour, and combines it with local fault records of the past 24 hours (without historical faults) to construct a deep feature fault template. For example, "bearing wear fault" corresponds to 3 core deep feature thresholds: bearing harmonic ratio > 3%, peak vibration acceleration > 20 m / s². 2 Temperature change rate > 0.5℃ / s.

[0075] (II) Calculation of cosine similarity between real-time 22-dimensional fault hazard characteristics and fault template: bearing harmonic proportion 1.5%, peak vibration acceleration 15 m / s² 2 The temperature change rate is 0.22℃ / s, and the cosine similarity with the "bearing wear failure" template is 0.35 (threshold 0.6), so it is not marked as a suspected failure; the cosine similarity with the "cable insulation aging failure" template (leakage current growth rate > 5%, insulation resistance decay rate > 2%) is 0.42, so it is also not marked.

[0076] (III) Verification based on equipment mechanism rules: The current leakage current growth rate is 4% (≤5%) and the insulation resistance decay rate is 1.67% (≤2%). There are no abnormal mechanism phenomena, so it is determined that there is no confirmed fault. If the leakage current growth rate rises to 6.5% and the insulation resistance decay rate is 2.3% at a certain moment, combined with the "cable insulation aging" mechanism (the leakage current increases as the insulation decreases), the verification is passed and it is determined to be a confirmed fault.

[0077] (iv) Upon diagnosis of a "mild cable insulation aging fault," a self-healing strategy should be implemented: reduce the motor load rate to 20%, extend the heat dissipation time, and reduce leakage current. Verification should be conducted in two phases: Phase 1: Monitor basic parameters (leakage current decreases from 6.5mA to 5.8mA; determine if it returns to the normal range ≤5mA? If not, continue with the strategy). Phase 2: Monitor potential fault characteristics (leakage current growth rate decreases from 6.5% to 4.2%, returning to the safe range ≤5%; insulation resistance attenuation rate decreases from 2.3% to 1.8%, returning to the safe range ≤2%). Successful self-healing is determined upon meeting both phases. If the leakage current still does not drop below 5mA after 1 hour, initiate secondary processing, send a remote maintenance command, and arrange for personnel to replace the cable.

[0078] Step 6: Closed-loop optimization and iterative upgrade:

[0079] (I) Based on deep feature analysis, six evaluation indicators are constructed and evaluated every 2 hours: Control accuracy: Speed ​​deviation 0.3% (compliant ≤ 0.5%). Stability: Current fluctuation amplitude 1.2% (compliant ≤ 2%). Energy consumption: Average power 42kW (compliant ≤ 45kW). Fault occurrence rate: 0 times / 2h (compliant ≤ 0.1 times / h). Spatiotemporal synchronization: Multi-motor coordination deviation 0.2s (compliant ≤ 0.5s). Fault hazard rate: 1 hazard (leakage current growth rate 4%, compliant ≤ 2 hazards). An evaluation report is generated, and the "Energy Consumption Indicators" and "Fault Hazard Rate" can be optimized.

[0080] (II) The particle swarm optimization algorithm (50 particles, 30 iterations) was used to optimize the parameters of the fuzzy PID model: Initial parameters: Kp = 2.5, Ti = 0.5s, Td = 0.1s, energy consumption 42kW, hazard rate 1 item. After iteration: Kp = 2.3, Ti = 0.6s, Td = 0.08s, energy consumption reduced to 40kW, hazard rate 0 items, satisfying the requirement that "all evaluation indicators meet the standards and at least two are improved" (energy consumption and hazard rate improved).

[0081] (III) Under the "medium load transition state", three in-depth characteristic modes are further subdivided: Mode 1: Current fluctuation <5%, temperature stable. Mode 2: Current fluctuation 5%-10%, temperature rises slowly. Mode 3: Current fluctuation >10%, temperature rises rapidly. The current motor is in "Mode 2", so the corresponding optimization strategy is invoked (increase the PID integral time and reduce the temperature rise rate).

[0082] (iv) The cloud platform pushes a new vibration signal feature extraction algorithm package (optimizing the wavelet denoising layer to 6 layers), employing differential update (transmitting only the differences from the old algorithm) and breakpoint resume technology (continuing transmission from the breakpoint after network interruption) to complete the update. Offline simulation verification was conducted: the new algorithm was tested using historical vibration data (including 100 sets of fault data), and the fault identification accuracy improved from 92% to 95%. The new algorithm was then activated after successful verification.

[0083] Example 2

[0084] Step 1: Multi-dimensional parameter fusion and acquisition:

[0085] Electrical parameters: Current sensor collects stator current (0-800A) and rotor current (0-200A), voltage sensor collects grid voltage (10kV±5%) and fan output voltage, power analyzer collects power factor (0.9-1.0) and active power (0-1.5MW), torque sensor collects spindle torque (0-10000N·m), leakage current detector collects cabinet leakage current (≤5mA), and insulation resistance meter collects cable insulation resistance (≥500MΩ).

[0086] Physical state parameters: Infrared thermal imager collects generator stator and rotor temperature (normal ≤150℃), vibration sensor collects main shaft and bearing vibration frequency (5-500Hz), angle sensor collects blade pitch angle (0°-90°), pressure sensor collects hydraulic system pressure (10-15MPa), and anemometer collects impeller speed (0-20r / min) and wind speed (0-25m / s).

[0087] Environmental parameters: Temperature and humidity sensors collect data on the cabin temperature (-30℃-60℃) and humidity (30%-70%); dust sensors collect data on the cabin dust concentration (≤5mg / m³). 3 The lightning monitoring instrument collects the lightning interference intensity (≤100dBμV / m).

[0088] Time stamp parameters: collection timestamp (accurate to milliseconds), fan number (WT-012), operating condition identifier (initial "to be identified").

[0089] The data acquisition cycle is set as follows: electrical parameters (current, voltage, power) are acquired every 50ms; physical state parameters (temperature, vibration, speed) are acquired every 200ms; environmental parameters are acquired every 10s; time-series label parameters are generated synchronously. Data is stored in the PLC's local storage module (capacity 128GB) according to "parameter type-acquisition time-fan number" and automatically backed up to the power plant's cloud database every morning. Abnormal data judgment: when the wind speed suddenly increases from 12m / s to 30m / s (over-range 25m / s), the trend of the next 3 acquisition cycles (150ms) is compared: 30m / s→18m / s→13m / s, which is judged as a lightning interference signal and directly rejected; if the spindle torque increases from 8000N·m to 8500N·m→9000N·m, it is judged as an overload precursor, marked as "abnormal to be verified", and the result "suspected spindle overload" is added and transmitted to the PLC main controller.

[0090] The second step is parameter preprocessing and feature extraction:

[0091] (I) Vibration and torque signals: 5-layer wavelet denoising (wavelet basis sym5) + empirical mode decomposition, noisy IMFs with correlation <0.25 are removed from 10 IMFs, and the reconstructed vibration signal error is ≤±1.5%, and the torque signal error is ≤±1%. Current and voltage signals: Kalman filtering (state equation considering grid fluctuations) + sliding window averaging (window of 5 periods, 250ms), and the filtered voltage fluctuation is ≤±0.5%, and the current fluctuation is ≤±0.8%. Standardization: Taking a wind turbine with a rated power of 1.2MW and a maximum power of 1.5MW as an example, when the actual power is 1.3MW, the standardized value = (1.3-1.2) / (1.5-1.2)≈0.33. Combined with the equipment aging coefficient (20000h of operation, coefficient 0.88), the corrected value is 0.33×0.88≈0.29.

[0092] (II) Electrical parameters (10 cycles, 500ms): Current: Average 750A, Peak 780A, Valley 730A, Variance 36A 2 RMS 752A, crest factor 1.04; current 2nd-10th harmonic percentage (2nd harmonic 1.0%, 3rd harmonic 0.7%). Insulation resistance: 1-minute attenuation rate (520MΩ→518MΩ, attenuation rate 0.38%). Physical state parameters: Temperature: 10-second average rate of change (stator temperature from 120℃→121℃, 0.1℃ / s), cumulative temperature rise of 1℃, stator-nacelle temperature difference 25℃. Vibration: peak acceleration 8m / s². 2 Vibrational energy (root mean square acceleration 6 m / s²) 2 ×1s=6m·s). Pitch angle: average value of 30° over 10 seconds, with a variation range of ±2°. Output a 32-dimensional basic statistical feature vector.

[0093] (III) Trend Analysis of Sliding Window (3 periods, 600ms): Power root mean square slope: 1.25MW→1.28MW→1.31MW, slope 0.02MW / ms. Temperature change rate slope: 0.1℃ / s→0.12℃ / s→0.14℃ / s, slope 0.013℃ / s. 2 Trend Consistency: Both power and temperature continue to rise. Points with a slope > 0.5 / cycle (e.g., power slope 0.02MW / ms > 0.5MW / cycle) are marked as "mutation points". Spatial Correlation: The power correlation coefficient between this wind turbine and adjacent wind turbines is calculated to be 0.92, and the speed coordination deviation is 3%. An 18-dimensional spatiotemporal correlation feature vector is output and merged into a 50-dimensional feature set.

[0094] (iv) 22-dimensional fault characteristics: Spindle torque fluctuation and speed coupling coefficient: torque fluctuation 4%, speed fluctuation 2%, coupling coefficient 2. Blade pitch angle and wind speed hysteresis coefficient: pitch angle response time 0.3s, wind speed change time 0.2s, hysteresis coefficient 1.5. Cable leakage current growth rate: from 3mA to 3.1mA every 10 seconds, growth rate 3.3%.

[0095] (v) Random forest is used to calculate the correlation degree. Features with >0.6 are retained (root mean square power, temperature change rate, torque coupling coefficient, etc.), and redundant features with mutual information entropy >0.8 are removed (such as "power peak" and "power root mean square" mutual information entropy 0.82, retain "power root mean square"). Finally, 38-dimensional key deep features are retained.

[0096] The third step is to construct the dynamic control model:

[0097] (I) Basic Model Library Parameters: PID: Kp = 3.0, Ti = 0.8s, Td = 0.2s. Fuzzy Control: Input "Power Deviation" and "Speed ​​Deviation", Output "Pitch Angle Adjustment", 36 fuzzy rules. Predictive Control: 8-step prediction in the time domain, 3-step control in the time domain, with constraints of power ≤ 1.5MW and speed ≤ 20r / min. Based on the initial load rate (1.3MW / 1.5MW ≈ 87%), the initial parameters of the PID model with feedforward are preferentially called.

[0098] (ii) K-means clustering of 5 operating conditions, current load rate 87% (>70%), parameter fluctuation 4% (<5%), determined to be "heavy load steady state"; SVM classification confidence 95%, no secondary identification required.

[0099] (III) "Heavy load steady state" + torque coupling coefficient 2 (normal ≤3), adopting "PID model with feedforward + fault hazard suppression model": Feedforward PID: Adjust the pitch angle (+3°) in advance based on wind speed prediction (next wind speed 14m / s) to reduce power fluctuation. Fault hazard suppression: Set the suppression coefficient to 0.98 for leakage current growth rate of 3.3% to reduce cable load.

[0100] (iv) After the model output, the power deviation decreased from 0.1MW to 0.05MW (threshold 0.08MW) and the temperature deviation was 0.5℃ (threshold 1℃), indicating normal operation; if the power deviation is 0.1MW (exceeding the threshold) for 3 consecutive cycles, the PID model is reloaded and Kp = 3.2 and Ti = 0.7s is adjusted.

[0101] Step 4: Multi-objective adaptive adjustment:

[0102] (I) The priority matrix is ​​currently in "heavy load steady state" with no abnormalities. Match the "normal operating condition" priority as follows: 1. Control accuracy (power deviation ≤ ±0.05MW). 2. Energy consumption (wind turbine self-consumption ≤ 50kW). 3. Response speed (pitch angle adjustment response ≤ 0.3s).

[0103] (II) Distributed model prediction of regulation strategy: divided into "power regulation", "pitch angle regulation" and "speed regulation" sub-modules, predicting 5 steps of parameters and outputting sub-control quantities; spatiotemporal synchronization verification: ensure that the regulation timing of sub-modules is consistent (power regulation within 0.2s, pitch angle within 0.3s), and the total regulation quantity (pitch angle +2°, power limit 1.35MW).

[0104] (III) Abnormal Handling: If the pitch angle adjustment is +5° (safety threshold +4°): the leakage current growth rate is 3.3% (normal ≤5%), the amplitude is limited to +4°, and the log is recorded. If the leakage current growth rate is 5.2% (exceeding normal), the amplitude is limited to +4° and an early warning is triggered, which is pushed to the power plant monitoring center.

[0105] Step 5: Real-time fault diagnosis and self-healing control:

[0106] (i) Hourly synchronization of the cloud-based fault database (including 18 categories such as blade faults and generator faults), combined with local 24-hour fault-free records, to construct a template: "Blade jamming fault" corresponding threshold (pitch angle change rate < 0.5° / s, vibration peak value > 15m / s) 2 (Power fluctuation >10%).

[0107] (II) Real-time characteristics: Pitch angle change rate 1.2° / s, peak vibration rate 8m / s 2 The power fluctuation was 3%, and the cosine similarity to the "blade jamming" template was 0.28 (threshold 0.6), indicating no suspected fault.

[0108] (III) If the pitch angle change rate is 0.4° / s and the vibration peak value is 16m / s at a certain moment... 2 Based on the "blade jamming" mechanism (increased mechanical resistance leading to slow adjustment and increased vibration), the verification was successful, and the fault was diagnosed.

[0109] (IV) Self-healing strategy: Reduce wind speed sampling frequency, increase pitch angle adjustment (+5°), and activate hydraulic system lubrication. Phased verification: 1. Basic parameters: Pitch angle recovery rate is 1.0° / s, peak vibration is 12 m / s. 2 2. Fault characteristics: Power fluctuations decrease to 4%, returning to a safe range. Successful completion of both stages indicates successful self-healing; if recovery fails within 10 minutes, remote maintenance is initiated, and personnel are dispatched for on-site repair.

[0110] Step 6: Closed-loop optimization and iterative upgrade:

[0111] (I) Six indicators are evaluated every 4 hours: Control accuracy: Power deviation 0.03MW (compliant). Stability: Speed ​​fluctuation 0.5r / min (compliant). Energy consumption: Self-consumption 48kW (compliant). Failure rate: 0 times (compliant). Spatiotemporal synchronization: Multi-fan coordination deviation 0.15s (compliant). Fault hazard rate: 0 items (compliant).

[0112] (II) Particle Swarm Optimization (60 particles, 40 iterations): After optimization, the PID parameters Kp = 2.9, Ti = 0.85s, power deviation 0.02MW, and energy consumption 45kW, which meet the requirements of "both standards are met and both are improved".

[0113] (III) "Heavy Load Steady State" is further divided into 3 modes: Mode A: Stable wind speed (12-14 m / s). Mode B: Fluctuating wind speed (14-16 m / s). Mode C: Sudden increase in wind speed (16-18 m / s). The current wind speed is 15 m / s (Mode B), and the "Enhanced Predictive Control" strategy is invoked.

[0114] (iv) The cloud platform pushes a new power feature extraction algorithm, and offline verification is performed after differential update: historical power data testing shows that the fault identification accuracy has increased from 90% to 94%, and the verification is successful and enabled.

[0115] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A PLC electrical control cabinet control method, characterized in that, The method specifically includes the following steps: S1. Multi-dimensional parameter fusion acquisition: Multi-dimensional data is acquired through multiple types of sensors, hierarchical acquisition cycles are set according to parameter type, multi-source data is stored in categories and layers, and abnormal data is traced and judged. S2. Parameter preprocessing and feature extraction: Different noise reduction algorithms are used for different signal types, and deep features of the data are extracted through multi-step processing. S3. Dynamic control model construction: A dynamic control model is constructed based on deep feature data. The working conditions are identified based on key deep features. The model is dynamically adapted and optimized by combining the working conditions and fault hazard features. At the same time, model verification is strengthened and deep feature deviation verification is added. S4. Multi-objective adaptive adjustment: Based on deep features, the control target priority matrix is ​​dynamically adjusted, deep feature refinement calculation is incorporated, a distributed model prediction and spatiotemporal synchronous verification collaborative upgrade strategy is adopted, the adjustment amount increase is combined with early warning, and fault hazard characteristics are associated. S5. Real-time fault diagnosis and self-healing control: The cloud-based fault feature library matches deep features to construct deep feature fault templates. It adopts a dual diagnostic algorithm of deep feature matching and mechanism rule judgment. It combines fault type and deep feature level to formulate a graded self-healing control strategy, extends the fault recovery verification cycle, and conducts verification in two stages. S6. Closed-loop optimization and iterative upgrade: The system control effect is quantitatively evaluated periodically. The model parameter optimization algorithm is upgraded by adopting a deep feature-oriented particle swarm optimization algorithm. The optimization strategy library is classified and stored according to the working condition type and deep feature mode. The remote upgrade function is expanded and online upgrade of the deep feature extraction algorithm is supported.

2. The PLC electrical control cabinet control method according to claim 1, characterized in that: In step S1, the collected data is divided into electrical parameters, physical state parameters, environmental parameters, and time-series tag parameters. Electrical parameters include: real-time current value, voltage value, power factor, power, torque, leakage current, and insulation resistance; physical state parameters include: equipment temperature, vibration frequency, actuator position, pressure value, and torque fluctuation; environmental parameters include: ambient temperature, humidity, dust concentration, and electromagnetic interference intensity; time sequence label parameters include: acquisition timestamp, equipment number, and operating condition identifier, initialized to be labeled; for instantaneous signals exceeding the preset range, the judgment is made by comparing the parameter change trends of three adjacent acquisition cycles. If it is determined to be an interference signal, it is directly rejected; if it is determined to be a precursor to a knowledge fault, it is marked as an abnormality to be verified and retained. When packaging the data of the line to be verified, the judgment result is attached and transmitted to the PLC main controller.

3. The PLC electrical control cabinet control method according to claim 2, characterized in that: Step S2 specifically includes the following steps: S21. Optimize the noise reduction algorithm for different signal types: Vibration signal and torque signal adopt 5-layer wavelet noise reduction + empirical mode decomposition, 10 intrinsic mode functions at the boundary, and reconstruct the signal after removing the noisy intrinsic mode functions. Current and voltage signals adopt Kalman filtering + sliding window averaging, with a window size of 5 acquisition cycles. The standardization process is interval standardization + deviation correction. First, the parameters are mapped to the interval [1, 1] according to the standardization value = (actual value, rated value) / (limit value, rated value). Then, the standardization result is corrected according to the equipment aging coefficient calculated from the running time and historical fault data. S22. For electrical parameters: calculate the average, peak, valley, variance, root mean square, and crest factor over 10 acquisition cycles. For current parameters, additionally calculate the harmonic content and the proportion of the 210th harmonic. For insulation resistance parameters, calculate the attenuation rate over 1 minute. For physical state parameters: calculate the average rate of change, cumulative temperature rise, and temperature difference between different parts of the equipment over 10 seconds. For vibration parameters, calculate the peak acceleration and the root mean square multiplied by time. Output a 32-dimensional basic statistical feature vector. S23. Perform sliding window trend analysis on basic statistical features. Set the window size to 3 statistical periods, calculate the slope of feature changes, such as the root mean square slope of current, the slope of temperature change rate, and trend consistency. Compare whether the three consecutive windows change in the same direction. Mark the point with the absolute value of the slope > 0.5 / period as the "trend break point". Spatial correlation features: calculate the correlation coefficient and coordination deviation of parameters between different devices, output an 18-dimensional spatiotemporal correlation feature vector, and merge it with the basic statistical feature vector to form a 50-dimensional feature set. S24. Based on the equipment operation mechanism, extract fault precursor features from the 50-dimensional feature set. For motor equipment: extract the bearing characteristic frequency harmonics as the proportion of the 23rd harmonic in the 100-500Hz frequency band, the stator current imbalance as the ratio of the three-phase current deviation to the average value, and the coupling coefficient of torque fluctuation and speed. For circuit equipment: extract the cable leakage current growth rate as the growth rate every 10 seconds, the insulation resistance temperature coefficient as the resistance change rate for every 1℃ increase in temperature, and the power factor mutation frequency. For actuator equipment: extract the lag coefficient of position deviation and voltage fluctuation as the ratio of position deviation response time to voltage change time, and the valve opening adjustment dead zone as the minimum difference between the command change and the actual opening change. Output a 22-dimensional fault hazard feature vector. S25. Using random forest, the correlation between 72-dimensional features and control objectives, including control accuracy, energy consumption, equipment lifespan, and failure risk, is calculated. Features with a correlation greater than 0.6 are retained. The mutual information entropy between the retained features is calculated, and redundant features are eliminated. For features with a mutual information entropy greater than 0.8, the one with the higher correlation is retained. Finally, 3040-dimensional key deep features are retained.

4. The PLC electrical control cabinet control method according to claim 3, characterized in that: Step S3 specifically includes the following steps: S31. The basic model library includes: PID control model, fuzzy control model, predictive control model, spatiotemporal correlation compensation model, and fault hazard suppression model. When loading, the initial parameters of the model with the highest matching degree with the current operating condition are called first. S32. Using K-means, the deep features of historical operating data are clustered into 5 typical operating conditions: Light load steady state (load rate <30%, parameter fluctuation <5%), Medium load transition state (load rate 30%-70%, parameter fluctuation 5%-15%), Heavy load steady state (load rate >70%, parameter fluctuation <5%), Abrupt load state (load change rate >20% / s, torque abrupt change rate >30% / s, current harmonic abrupt change (second harmonic proportion >5% / cycle), and Abnormal approach state (key parameters approaching alarm threshold). An SVM-trained operating condition classifier is used to classify the real-time deep feature vectors, outputting the operating condition type and confidence level. When the confidence level is <80%, a secondary identification is initiated: 10 additional cycles of data are collected and recalculated. S33. Adjust the model based on the fault hazard characteristics in the working condition type and depth characteristics: Light load steady state: adopt a simplified PID model with low energy consumption optimization; Medium load transition state: adopt a fuzzy PID spatiotemporal correlation compensation composite model; Heavy load steady state: adopt a PID model with feedforward with fault hazard suppression; Abrupt load state: adopt a predictive control model with torque compensation; Abnormal approach state: adopt a fault hazard suppression model with conservative control. S34. Calculate the change in key depth features after the model output. If the control deviation is greater than the threshold or the depth feature deviation threshold for three consecutive cycles, reload the basic model and adapt the parameters.

5. A PLC electrical control cabinet control method according to claim 4, characterized in that: In step S4, the priority optimization matrix includes: normal operating conditions: priority from high to low is control accuracy, energy consumption, and response speed; high temperature operating conditions: priority from high to low is equipment protection, control accuracy, and energy consumption; fault hazard operating conditions: priority from high to low is fault suppression, equipment protection, and accuracy control; multi-device collaborative operating conditions: priority from high to low is spatiotemporal synchronization, control accuracy, and energy consumption.

6. The PLC electrical control cabinet control method according to claim 5, characterized in that: In step S4, if the adjustment amount exceeds the safety threshold and the corresponding potential fault characteristics are within the normal range, then only the amplitude is limited and recorded. If the adjustment exceeds the safety threshold and the corresponding fault hazard characteristics are not within the normal range, a hazard warning will be triggered while limiting the amplitude, and pushed to the monitoring terminal for early intervention.

7. A PLC electrical control cabinet control method according to claim 6, characterized in that: Step S5 specifically includes the following steps: The S51 and PLC controllers synchronize the latest fault feature library from the cloud server every hour, and combine it with the fault records of the past 24 hours in the local operation history data to supplement and match deep fault features to build deep feature fault templates. Each fault type corresponds to 35 core deep feature thresholds. S52. Calculate the cosine similarity between the real-time 22-dimensional fault hazard features and the fault template, and mark those with a cosine similarity greater than the threshold as suspected faults. S53. Verification based on equipment mechanism rules: if the verification passes, it is determined to be a confirmed fault; otherwise, it is marked as pending observation. The diagnosis cycle is synchronized with the data acquisition cycle. S54. Develop a self-healing strategy based on the fault type and depth of the fault characteristics, and verify the fault recovery in two stages: monitor whether the basic parameters have returned to the normal range; monitor whether the fault hazard characteristics have returned to the safe range. If both stages are met, the self-healing is considered successful; otherwise, initiate secondary processing, such as remote operation and maintenance intervention.

8. The PLC electrical control cabinet control method according to claim 1, characterized in that: Step S6 specifically includes the following steps: S61. Based on deep features, evaluation indicators are formed, including: control accuracy, stability, energy consumption, failure rate, spatiotemporal synchronization, and failure hazard rate. Evaluation is performed at fixed intervals to generate evaluation reports. S62. Based on the evaluation metrics, the model is iterated using the particle swarm optimization algorithm. Each iteration ensures that all evaluation metrics meet the standards and at least two metrics improve to the threshold. S63. Each working condition is further subdivided into 34 deep feature patterns. When invoking a strategy, the working condition is matched first, and then the corresponding optimization strategy is invoked based on the real-time deep feature pattern. S64. The cloud platform pushes a new feature extraction algorithm package, which adopts score update and breakpoint resume technology. After the update is completed, it is verified through offline simulation and tested with historical data. The algorithm is then enabled after the verification is successful.

9. A PLC electrical control cabinet control system, characterized in that: This system is used to implement a PLC electrical control cabinet control method according to any one of claims 1-8, comprising: a multi-dimensional parameter fusion acquisition module, a parameter preprocessing and feature extraction module, a dynamic identification module, an adaptive adjustment module, a real-time fault diagnosis and self-healing module, and an optimization iteration module; Multi-dimensional parameter fusion acquisition module: used to acquire multi-dimensional data through multiple types of sensors, set hierarchical acquisition cycles according to parameter type, classify and store multi-source data in layers, and trace and determine the source of abnormal data; The parameter preprocessing and feature extraction module is used to extract deep features of data by employing different noise reduction algorithms for different signal types through multi-step processing. Dynamic identification module: used to build dynamic control models based on deep feature data, identify working conditions based on key deep features, dynamically adapt and optimize the model by combining working conditions and fault hazard features, and at the same time strengthen model verification and increase deep feature deviation verification. Adaptive adjustment module: used to dynamically adjust the control target priority matrix based on deep features, incorporate deep feature refinement calculation, adopt a distributed model prediction and spatiotemporal synchronous verification collaborative upgrade strategy, combine the adjustment amount increase with early warning, and associate fault hazard characteristics; Real-time fault diagnosis and self-healing module: It is used to construct deep feature fault templates by matching deep features with the fault feature library in the cloud. It adopts a dual diagnosis algorithm of deep feature matching and mechanism rule judgment, and formulates a graded self-healing control strategy by combining fault type and deep feature level to extend the fault recovery verification cycle. The verification is carried out in two stages. The optimization iteration module is used to perform quantitative evaluation of the system control effect at regular intervals. It adopts a deep feature-guided particle swarm optimization algorithm to upgrade the model parameter optimization algorithm, stores the optimization strategy library according to the working condition type and deep feature mode, expands the remote upgrade function, and supports online upgrade of the deep feature extraction algorithm.

Images