A non-contact AO channel current monitoring pre-diagnosis method and system

Abstract

This invention relates to the field of electronic digital data processing technology, and discloses a non-contact AO channel current monitoring and pre-diagnosis method and system. The method includes: connecting a non-contact open-type current sensor to the AO module output cable without interrupting the circuit, thereby physically isolating the current in real time and generating a corresponding voltage signal; sampling the current signal using a high-speed ADC sampling circuit to obtain current waveform data, and extracting multi-dimensional waveform features from it; using a machine learning diagnostic model pre-trained based on historical data to analyze the extracted waveform features and obtain the status diagnosis result of the AO module output channel; the machine learning diagnostic model is a hybrid architecture combining convolutional neural networks and long short-term memory networks; and triggering corresponding level warnings or alarms based on the level of the status diagnosis result. This method solves the problems of delayed fault detection and difficulty in locating the root cause, achieving the goals of real-time monitoring and early warning, accurate fault tracing, and improved system reliability.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and in particular to a non-contact AO channel current monitoring and pre-diagnosis method and system. Background Technology

[0002] In the industrial sector, control systems and automation equipment play a crucial role. However, the AO (Automatic Animation) modules and control loops within these control systems often operate in harsh environments. AO modules are frequently used for controlling the opening of critical valves. Internal component failures, external load variations, short circuits, and other factors can cause abnormal channel states, leading to signal errors or even system shutdowns. While current modules incorporate numerous detection and control technologies, such as fault detection, monitoring, and feedback, they still suffer from the inability to detect signal changes and error states in a timely manner, impacting the safety and normal operation of the control system. Current input / output (I / O) modules also exhibit this problem. When an I / O channel malfunctions, abnormal signals cannot be detected promptly, requiring the signal to reach a significant fault diagnosis threshold before detection. Even worse, the erroneous signal may have already been transmitted and received by the system through the faulty channel's communication before the system can detect it, potentially causing a system shutdown. Therefore, effective management and control of real-time, efficient detection, diagnosis, and feedback for I / O modules are paramount.

[0003] For example, Chinese patent CN120105291A discloses a method for labeling voltage and current anomaly data based on CNN and LSTM, providing the following technical solution: Step S1: Constructing a voltage and current anomaly dataset and preprocessing the voltage and current anomaly data; Step S2: Building a voltage and current anomaly data recognition model based on a CNN model and an LSTM model, including a CNN-LSTM hybrid model with a first one-dimensional convolutional layer, a first one-dimensional pooling layer, a first dropout layer, a second one-dimensional convolutional layer, a second one-dimensional pooling layer, a second dropout layer, a first LSTM layer, a third dropout layer, a second LSTM layer, and an output layer; Step S3: Optimizing the CNN-LSTM hybrid model built in Step S2 based on transfer learning to obtain an optimized CNN-LSTM hybrid model; Step S4: Training and validating the optimized CNN-LSTM hybrid model using the voltage and current anomaly dataset to obtain a validated model; Step S5: Using the validated model to identify voltage and current anomaly data. This invention can automatically learn and identify anomaly patterns, improving the accuracy and efficiency of voltage and current anomaly data screening. However, the aforementioned voltage and current anomaly data labeling method based on CNN and LSTM is mainly aimed at the three-phase voltage and current of AC power systems. It adopts contact measurement and general anomaly labeling, but lacks targeted functions such as non-contact monitoring, hierarchical pre-diagnosis, multi-sensor fusion and fault prediction for industrial AO control loops. Summary of the Invention

[0004] This invention solves the problems of delayed fault detection, difficulty in locating the root cause, and inability to perform predictive maintenance in the prior art. It proposes a non-contact AO channel current monitoring and pre-diagnosis method and system, which achieves the goals of real-time monitoring and early warning, accurate fault tracing, and improved system reliability.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A non-contact AO channel current monitoring pre-diagnosis method includes: A non-contact open current sensor is connected to the output cable of the AO module without interrupting the circuit, and the current is sensed in real time and the corresponding voltage signal is generated in a physical isolation manner. The current signal is sampled by a high-speed ADC sampling circuit to obtain current waveform data, from which multi-dimensional waveform features are extracted. The machine learning diagnostic model, which is pre-trained based on historical data, analyzes the extracted waveform features to obtain the state diagnostic results of the AO module output channel. The machine learning diagnostic model is a hybrid architecture combining convolutional neural networks and long short-term memory networks. Based on the level of the status diagnosis result, a corresponding level of early warning or alarm is triggered.

[0006] It achieves non-destructive, non-contact current monitoring, avoiding modifications to the original control system, ensuring electrical safety and isolation. High-speed ADC sampling improves data acquisition accuracy and response speed. Machine learning hybrid models combined with historical data pre-training improve the accuracy of abnormal waveform identification and status diagnosis. The hierarchical early warning mechanism can trigger corresponding alarms in a timely manner according to the severity of the fault, ensuring the safe and stable operation of the industrial control system.

[0007] Preferably, the hybrid architecture combining the convolutional neural network and the long short-term memory network includes: a one-dimensional convolutional layer, a bidirectional long short-term memory network layer, an attention mechanism layer, and a fully connected classification layer connected in sequence. The one-dimensional convolutional layer extracts local features of the current waveform, the bidirectional long short-term memory network layer captures the temporal dependencies of the waveform, the attention mechanism layer weights and highlights features of abnormal time periods, and the fully connected classification layer outputs the waveform state category and abnormality type.

[0008] The hybrid neural network architecture can comprehensively and efficiently process time-series current data. One-dimensional convolutional layers effectively extract local waveform details such as spikes and steps. Bidirectional long short-term memory network layers capture the time dependencies between different time periods to identify trend anomalies. Attention mechanism layers focus on key abnormal periods to improve diagnostic accuracy. Fully connected classification layers output specific states and anomaly types, enabling the model to have powerful feature learning and classification capabilities.

[0009] Preferably, the triggering of the corresponding level of warning or alarm includes: when the waveform characteristics deviate from the normal template but do not exceed the safety threshold, it is recorded as a state of attention and no reporting is triggered; when the waveform abnormality continues and the characteristics exceed the deviation threshold, a medium-level warning is triggered and alarm information is uploaded; when the waveform matches a known fault mode or exceeds the safety threshold, an advanced alarm is immediately triggered, and audible and visual prompts and remote notifications are supported.

[0010] The tiered early warning and alarm mechanism enables refined fault management, pays attention to internal status records to avoid unnecessary interference, thereby reducing false alarms while ensuring rapid handling of emergencies, and improving system security and maintenance efficiency.

[0011] Preferably, the training process of the machine learning diagnostic model includes: collecting historical normal waveform data and abnormal waveform data, labeling the abnormal waveform data with anomaly levels, using the labeled data to train the hybrid model, using a combination of cross-entropy loss and time-series smoothing regularization term as the loss function, and deploying the trained model to the main control chip after quantization, and updating the model through online incremental learning.

[0012] Training based on labeled data enables the model to learn real fault modes. The combination of cross-entropy loss and temporal smoothing regularization improves the model's generalization ability and reduces false alarm fluctuations. After model quantization, it is adapted to embedded main control chip resources. Online incremental learning supports continuous optimization, ensuring the accuracy and adaptability of the diagnostic model in actual operation.

[0013] Preferably, the machine learning diagnostic model further includes: when the operator confirms that an alarm is a false alarm or a missed alarm, the corresponding waveform data is manually marked, the system triggers incremental learning based on the marked data, updates the machine learning diagnostic model, and the model is retrained periodically using newly collected data to adapt to equipment aging or changes in operating conditions.

[0014] Driven by incremental learning through manual feedback labels, the system can dynamically correct false alarms and false negatives, and periodically retrain the model to adapt to equipment aging and changes in operating conditions, thereby continuously improving diagnostic accuracy and enhancing the system's self-learning ability and long-term robustness.

[0015] Preferably, the machine learning diagnostic model further includes: synchronizing and aligning the current waveforms of multiple AO channels corresponding to the same actuator with timestamps, calculating the difference features, correlation features and consistency features between the waveforms of multiple channels, and determining whether the fault is a common fault based on cross-channel features and preset fault rules, and locating the source of the fault as the AO module or actuator.

[0016] Multi-channel data synchronization and feature fusion analysis can distinguish between common faults and individual faults. By combining difference, correlation and consistency features with preset rules, it can accurately locate the root cause of the fault, help to quickly isolate problems in the AO module or actuator, and improve the accuracy and efficiency of fault diagnosis.

[0017] Preferably, the machine learning diagnostic model further includes: calculating a channel health index based on historical waveform data, wherein the health index is obtained by weighting multiple degradation features among waveform fidelity, noise level, response speed, and stability; performing time-series prediction on the health index; extrapolating the time required for it to decrease to the failure threshold to obtain the predicted remaining service life; and generating preventive maintenance recommendations based on the prediction results.

[0018] The health index quantitatively assesses the degradation status of the channel, reflects the overall performance based on multi-feature weighted calculation, and enables predictive maintenance by extrapolating the remaining service life through time-series prediction. The early generation of maintenance suggestions can avoid sudden failures, extend equipment life, and reduce maintenance costs.

[0019] Preferably, the extraction of multi-dimensional waveform features includes: extracting time-domain and frequency-domain features of the current waveform; the time-domain features include statistical features, dynamic response features, and morphological and distribution features, the statistical features include mean and variance, the dynamic response features include rise time and fall time and overshoot amplitude, the morphological and distribution features include peak factor, pulse index, and waveform entropy; the frequency-domain features include fundamental and harmonic amplitudes, total harmonic distortion rate, and spectral centroid extracted by Fourier transform.

[0020] Preferably, the system includes: a non-contact open-type current sensor, directly mounted on the output cable of the AO control loop, for non-contact measurement of the loop current; a high-speed ADC sampling circuit, with its input terminal electrically connected to the signal output terminal of the non-contact open-type current sensor, for analog-to-digital conversion of the current signal; a main control chip, with its input terminal connected to the output terminal of the high-speed ADC sampling circuit, for receiving and analyzing the sampled data, and generating diagnostic results and corresponding alarm information; and a host computer, which communicates with the main control chip via a communication interface to receive alarm information and waveform data.

[0021] The system integrates non-contact sensors, high-speed ADCs, main control chips, and host computers. Non-contact measurement ensures safe isolation, high-speed ADCs guarantee data accuracy, the main control chip enables real-time analysis, and the host computer provides human-machine interaction and data management, achieving efficient and reliable current monitoring and fault pre-diagnosis.

[0022] Preferably, the non-contact open-ended current sensor adopts an openable magnetic ring structure. Its outer shell is made of multiple layers of electromagnetic shielding material, with the inner layer being a high permeability alloy, the outer layer being a conductive coating and grounded, and the middle being filled with wave-absorbing material. The internal structure uses a Hall effect element or a magnetoresistive sensor. The non-contact open-ended current sensor is also equipped with multiple signal output interfaces, and its opening is equipped with a self-locking mechanism.

[0023] The multi-layer electromagnetic shielding design effectively suppresses external interference, Hall or magnetoresistive sensors ensure high-precision non-contact measurement, multiple signal output interfaces enhance system compatibility, and the self-locking mechanism ensures magnetic circuit closure and measurement stability, thereby improving the sensor's practicality, reliability, and anti-interference capability.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows.

[0025] 1. This invention monitors the output channel current in real time using a non-contact method, avoiding physical intervention in the original control loop and ensuring the continuity and safety of system operation. The sensor is directly connected to the cable without cutting the circuit, making installation simple and supporting flexible switching of multiple ranges. This greatly reduces deployment difficulty and modification risks, providing a plug-and-play reliable monitoring method for industrial sites.

[0026] 2. This invention employs machine learning algorithms for intelligent analysis and feature extraction of current waveforms. The system can accurately identify normal and abnormal operating conditions, enabling tiered early warning and alarm functions. Through historical data learning and adaptive model updates, the system continuously optimizes diagnostic accuracy, effectively distinguishing between intermittent interference and genuine faults, significantly improving the timeliness and reliability of pre-diagnosis.

[0027] 3. This invention enables rapid fluctuation detection and multi-channel synchronous monitoring, helping to quickly locate the source and sequence of faults, significantly shortening troubleshooting time. During long-term operation, it accumulates equipment health data and supports lifespan prediction, thereby guiding preventative maintenance, reducing unplanned downtime, improving production efficiency, and lowering overall maintenance costs. Attached Figure Description

[0028] Figure 1 This is an overall flowchart of a non-contact AO channel current monitoring and pre-diagnosis method according to the present invention.

[0029] Figure 2 This is a system block diagram of a non-contact AO channel current monitoring and pre-diagnosis system according to the present invention. Detailed Implementation

[0030] See Figures 1-2 As shown, a non-contact AO channel current monitoring and pre-diagnosis method includes: A non-contact open current sensor is connected to the output cable of the AO module without interrupting the circuit, and the current is sensed in real time and the corresponding voltage signal is generated in a physical isolation manner. The current signal is sampled by a high-speed ADC sampling circuit to obtain current waveform data, from which multi-dimensional waveform features are extracted. The machine learning diagnostic model, which is pre-trained based on historical data, analyzes the extracted waveform features to obtain the state diagnostic results of the AO module output channel. The machine learning diagnostic model is a hybrid architecture combining convolutional neural networks and long short-term memory networks. Based on the level of the status diagnosis result, a corresponding level of early warning or alarm is triggered.

[0031] like Figure 1 In one embodiment shown, Figure 1 This is an overall flowchart of a non-contact AO channel current monitoring and pre-diagnosis method according to the present invention. The present invention discloses a non-contact AO channel current monitoring and pre-diagnosis method and system for real-time status monitoring and fault early warning of output channels in industrial automation control. The present invention achieves intelligent monitoring and hierarchical alarm of the AO module output circuit through a series of coordinated steps, specifically including the following steps: First, a non-contact open-type current sensor is used to acquire the current signal from the output cable of the AO module. This sensor is directly connected to the outside of the cable without interrupting the original electrical circuit, and senses current changes in real time through physical isolation, converting them into corresponding voltage signals.

[0032] Subsequently, the analog signal output by the sensor is sampled with high precision and high frequency using a high-speed ADC sampling circuit to obtain continuous current waveform data. Based on this data, the system further extracts multi-dimensional waveform features, including multiple indicators in the time and frequency domains, such as mean, variance, rise time, overshoot amplitude, total harmonic distortion, and spectral centroid.

[0033] In the signal analysis and diagnosis stage, the system employs a pre-trained machine learning diagnostic model to intelligently analyze the extracted waveform features. This model is a hybrid architecture combining convolutional neural networks and long short-term memory networks, specifically comprising a one-dimensional convolutional layer, a bidirectional long short-term memory network layer, an attention mechanism layer, and a fully connected classification layer. The one-dimensional convolutional layer extracts local waveform features, the bidirectional long short-term memory network layer captures temporal dependencies, the attention mechanism layer weights and highlights abnormal periods, and the fully connected classification layer outputs waveform state categories and anomaly type codes.

[0034] Based on the different confidence levels and degrees of abnormality of the diagnostic results, the system implements a graded response mechanism: when the waveform characteristics slightly deviate from the normal template but do not exceed the limit, only internal recording is made; if the abnormality continues and the characteristics deviate significantly, a medium-level warning is triggered and the information is reported; when the waveform matches a known fault mode or exceeds the safety threshold, an advanced alarm is immediately triggered, supporting on-site audible and visual prompts and remote notifications.

[0035] The machine learning diagnostic model is trained based on historical normal and abnormal waveform data, with abnormal data labeled with different levels. Cross-entropy loss combined with a temporal smoothing regularization term is used as the loss function during training. After training, the model is quantized and deployed to the main control chip, supporting online incremental learning and periodic retraining to continuously adapt to equipment aging and changes in operating conditions.

[0036] The system also has cross-channel collaborative analysis capabilities: it performs timestamp synchronization and data alignment for multiple AO channels serving the same actuator, calculates the differences, correlations and consistency characteristics between channels, and determines whether it is a common fault by combining preset rules, and helps to locate the source of the fault from the AO module or external actuator.

[0037] In addition, the system introduces health assessment and lifespan prediction functions: the channel health index is calculated by weighting multiple degradation characteristics, and time-series prediction is made based on its historical trend. The time required to decline to the failure threshold is extrapolated, thereby estimating the remaining service life and generating preventive maintenance recommendations.

[0038] This system consists of a sensing layer (current monitoring module), an edge analysis layer (MCU), an application layer (host computer), and a supporting data transmission and synchronization mechanism.

[0039] Core components: 1. Non-contact high-precision open-type current sensor The system uses a Hall effect element to sense the ring-shaped magnetic field around the AO output cable to measure the current flowing through the conductor. A key design feature is the openable / closable magnetic ring structure, which eliminates the need to cut the cable during installation, achieving complete physical and electrical isolation.

[0040] In one embodiment, a dual Hall differential array is used, in which two or more Hall elements are symmetrically arranged inside the sensor. Differential measurement is used to cancel common-mode electromagnetic interference and improve the signal-to-noise ratio.

[0041] It features active temperature compensation and a built-in high-precision temperature sensor that monitors ambient temperature and sensor heat generation in real time. The output signal is dynamically calibrated using pre-stored temperature-error curves or polynomial fitting models. For example, it guarantees a measurement accuracy better than ±0.5% across the entire temperature range (-40°C to 85°C).

[0042] It features multi-range adaptive measurement, supporting mainstream industrial signal ranges such as 0-20mA, 4-20mA, and 0-10V, which can be switched via DIP switches or commands from the host computer software.

[0043] High electromagnetic compatibility: The shell adopts a multi-layer shielding design (high magnetic permeability alloy inner layer, conductive coating outer layer, and intermediate absorbing material), with a shielding effectiveness of ≥60dB, ensuring stable operation in harsh industrial electromagnetic environments.

[0044] 2. High-speed ADC (Analog-to-Digital Converter) sampling circuit It adopts a high-speed ADC chip with a sampling rate of not less than 10kSPS, and is equipped with a programmable anti-aliasing filter at the front end. The cutoff frequency is adaptively adjusted according to the sampling rate to prevent signal aliasing.

[0045] The circuit employs a double buffering mechanism, setting up two physical storage areas (such as DDR RAM or SRAM), one for real-time sampling and analysis, and the other for historical data archiving, to achieve continuous recording without data loss.

[0046] 3. Microcontrollers with machine learning capabilities The MCU integrates a neural network acceleration unit or a high-performance DSP core, supporting the running of lightweight machine learning models (models quantized using TensorFlow Lite). It is responsible for real-time data preprocessing, feature extraction, model inference, and decision output.

[0047] It supports online learning and model updates. After the on-site engineer confirms or rejects the alarm, the MCU will store the relevant waveform data and annotations in the internal cache, and start incremental learning during the off-peak period (night) to fine-tune the model parameters and achieve adaptive optimization that becomes more accurate with use.

[0048] 4. Machine Learning Pre-diagnostic Algorithms Model architecture: Employs a hybrid CNN-LSTM network. CNN layer (one-dimensional convolution): responsible for extracting local features of waveform signals, such as spikes and abrupt changes; LSTM layer (Long Short-Term Memory): Responsible for capturing the time dependence of waveforms and identifying trend changes; Attention mechanism: Focusing on abnormal periods to improve the model's sensitivity to key features; Output layer: Performs multi-classification, and the output includes four levels: "normal / basic anomaly / intermediate anomaly / high-level anomaly"; In this embodiment, the kernel width of the one-dimensional convolutional layer is 5, the stride is 1, the number of channels is 32, and the hidden units of the bidirectional LSTM layer are 64.

[0049] Training Implementation: Data source: Large-scale historical data (>100,000 normal samples, >10,000 abnormal samples), including real faults and simulated injected fault data.

[0050] A two-stage training method is adopted—first, large-scale pre-training is performed on the server side, and then fine-tuning is performed on the embedded device using field data to ensure the model's universality and field adaptability. After pruning and quantization, the model is deployed to the MCU, ensuring that the single inference time is <10ms, which fully meets the real-time requirements.

[0051] 5. Host computer software system As a human-machine interface, it enables data visualization, alarm management, parameter configuration, model updates, and report generation. It supports multi-channel data fusion display and integrates a fault root cause reasoning knowledge base (based on a rule engine) to assist in fault location analysis. It provides standard OPC UA or MQTT interfaces, enabling seamless integration with factory MES, SCADA, or digital twin platforms.

[0052] Tiered pre-diagnosis workflow The entire system's workflow is a continuous, adaptive loop, divided into multiple stages such as initialization, monitoring, analysis, and early warning.

[0053] 1. System Initialization and Self-Check Process Power-on self-test: This includes sensor zero-point and range self-calibration, ADC reference voltage detection, memory integrity verification, and execution of the built-in self-calibration sequence (such as the MCU outputting a known analog test current to verify whether the entire signal chain of the sensor and ADC is normal).

[0054] Communication handshake and configuration synchronization: Establish a connection with the host computer and synchronize information such as device parameters, alarm thresholds, and machine learning model versions.

[0055] 2. Real-time monitoring and data preprocessing High-speed sampling: Continuous acquisition at a sampling rate of 10kHz or higher; Double buffer rotation: ensures continuous recording, and can retain raw waveform data for up to 30 seconds or longer; Real-time data cleaning: An algorithm based on median filtering and standard deviation threshold is used to automatically remove outliers (such as transient strong interference) before the data enters the analysis process.

[0056] 3. Waveform Feature Extraction and State Recognition The system extracts time-domain and frequency-domain features in parallel: Time-domain features: mean, variance, peak factor, waveform entropy, rise / fall time; Frequency-domain features: spectrum, total harmonic distortion, and high-frequency noise energy extracted via fast Fourier transform.

[0057] These features serve as input to the machine learning model and also enter a multi-state recognition state machine (states include: normal, observation, warning, alarm, fault confirmation, and recovery). State transitions are determined by a combination of feature deviation, duration, and historical case matching.

[0058] 4. Tiered early warning and alarm output Primary warning (observation status): The waveform characteristics deviate slightly. The system records the information internally but does not actively report it to avoid information overload.

[0059] Intermediate warning: When the waveform is abnormal and the characteristics deviate significantly from the normal range, the system sends an alarm message to the host computer, records the detailed waveform and characteristics, and highlights them on the software interface.

[0060] Advanced Alarm: The waveform closely matches a historically confirmed severe fault mode or exceeds a safety threshold. In this case, in addition to the host computer alarm, the system can also activate on-site audible and visual alarms and remotely notify key personnel via SMS, email, etc.

[0061] IV. In-depth Implementation of Advanced Functions 1. Multi-sensor data fusion diagnostics Technical Implementation: For situations where the same actuator is controlled by multiple AO channels (such as redundant systems), the system uses high-precision clock synchronization (such as PTP, ±1μs) in hardware to ensure that the data has a unified timestamp.

[0062] Data alignment and commonality analysis: Establish logical "channel groups" and calculate the real-time differences and correlation coefficients between channels within the group. If all channels synchronously exhibit similar anomalies, it is determined to be a common source of fault (such as fluctuations in the common power supply). Then, cross-validation is performed by combining the actuator status feedback from the AI ​​(analog input) module to accurately locate whether the fault lies in the AO module, the line, or the actuator itself.

[0063] Output: Fault root cause probability analysis, such as "Probability of fault located in AO module: 85%", with a detailed cross-validation report.

[0064] 2. Adaptive threshold adjustment mechanism Technical Implementation: To overcome the limitations of fixed thresholds, the system constructs an environment and state vector S(t), specifically [T_env, T_sensor, U_operating_hours, Load_Type, ...]. Here, T_env is the ambient temperature, T_sensor is the sensor temperature, U_operating_hours is the cumulative operating hours, Load_Type is the load type encoding, and Avg_Current_Recent is the recent average current.

[0065] Dynamic calculation: Threshold_actual(t) equals Threshold_base multiplied by K_temp, then by K_aging, and finally by K_load, where each influencing factor (temperature, aging, load) is calculated based on measured data. K_temp, K_aging, and K_load are the influencing factors for temperature, aging, and load, respectively.

[0066] Reinforcement learning optimization: The threshold adjustment is modeled as a reinforcement learning problem: the state is S(t), the action is a small increase / decrease in the key threshold, and the reward is calculated based on the correct alarms, false alarms, and missed alarms over a period of time, specifically: Reward equals R_correct minus (λ1 multiplied by R_false_alarm) minus (λ2 multiplied by R_missed_detection). Where R_correct is the reward for a correct alarm, R_false_alarm is the penalty for a false alarm, R_missed_detection is the penalty for a missed alarm (discovered through manual confirmation or system self-check), and λ1 and λ2 are weighting coefficients.

[0067] By using algorithms such as Proximal Policy Optimization (PPO), a policy network is periodically trained on the host computer, enabling the system to automatically learn the optimal threshold policy under different operating conditions and dynamically balance sensitivity and specificity.

[0068] 3. Failure Prediction and Lifespan Assessment Health Index Calculation: A comprehensive health index CHI(t) (value 0-100) is defined, obtained by weighting the scores of multiple performance degradation features (such as waveform distortion, increased noise, and slowed response). Specifically: CHI(t) = 100 minus Σ [w_i multiplied by F_i(t)], where F_i(t) is the score of the i-th performance degradation feature, and w_i is its weight. Degradation features include: The waveform fidelity decreases, which is measured by comparing the dynamic time warping (DTW) distance between the real-time waveform and the factory / health period standard waveform; As noise levels rise, calculate the effective value of current noise in a specific high-frequency band (e.g., above 1 kHz) and observe its trend over time. The response speed is slower; the rise time of the statistical current to the step control command is compared with the historical baseline. To assess stability degradation, calculate the rate of change of the long-term current fluctuation variance.

[0069] Each feature score F_i(t) is obtained by linear or nonlinear mapping through the position of its current value relative to the warning threshold and the failure threshold.

[0070] Remaining useful life prediction: Time series analysis (ARIMA, LSTM prediction network) is performed on historical data of CHI(t) to predict its future trend. Combined with an engineering-defined failure threshold (e.g., CHI=30), the predicted remaining useful life is calculated. An ensemble prediction algorithm is used to combine the outputs of multiple models to improve the confidence level of the prediction.

[0071] Output: Not only does it provide the current health status, but it also outputs the predicted remaining lifespan (e.g., "It is expected that CHI will drop to the warning line in 6 months"), and generates specific preventive maintenance work orders based on this, such as "It is recommended to prepare spare parts and arrange downtime for maintenance within 3 months", thus realizing the transformation from passive maintenance to predictive maintenance.

[0072] like Figure 2 In one embodiment shown, Figure 2This is a system block diagram of a non-contact AO channel current monitoring and pre-diagnosis system according to the present invention.

[0073] In another embodiment, the technical solution of the AO channel current monitoring and diagnostic module provided by the present invention includes the following parts: 1. High-precision open-ended current sensor: The open-ended current sensor can be directly fitted onto the AO control loop, completely isolated from it. The current generated in the AO control loop can be detected by a high-precision data acquisition unit based on the sensor's principle, while avoiding physical contact and modification of the original system.

[0074] 2. High-speed ADC sampling circuit: The use of a high-speed ADC greatly improves the system's response speed and sampling accuracy, enabling the capture of rapid fluctuations at the millisecond level to ensure data accuracy. After the ADC samples the data, the conversion circuit converts the analog signal into a digital signal for storage and transmission.

[0075] 3. MCU with Machine Learning Algorithm: The AO channel current monitoring and diagnostic module of this invention employs an MCU with machine learning algorithm functionality. By cleaning and comparing historical data of the sampled AO control loop current data, it selectively extracts features and builds a model to identify the normal state of the sampled data. It then uses methods such as clustering to group waveforms, marks and records abnormal waveform data, and uploads it to the host computer in real time. Simultaneously, the machine learning algorithm can be updated and optimized according to actual needs, enabling timely monitoring and analysis of the AO channel control loop, ensuring the timely detection and repair of abnormal data in the automation system.

[0076] 4. Machine Learning Pre-Diagnostic Algorithm: Based on the AO data change time and corresponding sampled values, the slope of the AO sampled values ​​over time, as well as the amplitude and duration of historical waveforms such as waveform overshoot, undervoltage, and jitter, can be extracted. Combining the signal feedback characteristics of the resistors, capacitors, operational amplifiers, and other components in the control loop, an algorithm model highly correlated with the characteristics of this channel and the equipment load impedance is generated. This algorithm model is then applied to judge the current and time data waveforms sampled in real time. When abnormal waveform characteristics are identified, a medium-level pre-diagnostic alarm is triggered. When an abnormal waveform confirmed by the user's history is identified, a high-level fault alarm is triggered directly.

[0077] 5. Host Computer: The host computer serves as the interface for modules and applications, completing data reading, analysis, processing, storage, and presentation. This enables monitoring of the AO control loop at multiple levels, including overall system operation, detection, transmission, and data analysis, allowing for timely detection and handling of anomalies, and ultimately evaluating indicators such as the operational quality of the automation system.

[0078] This invention, by introducing core technologies such as back-checking circuits and machine learning, achieves the important objectives of real-time detection, sampling, diagnosis, and feedback of fault information in input / output modules, and has the following advantages: 1. This invention provides comprehensive, accurate and timely detection by non-contactly monitoring and diagnosing the current of the output channel in real time, ensuring the safe and stable operation of the industrial control system, improving factory production efficiency, and reducing equipment maintenance and management costs.

[0079] 2. The open-type current sensor of this invention can be directly fitted onto the AO control loop, completely isolated from the AO control loop, avoiding the problem of modifying the original system as in traditional methods. The range of the open-type sensor can be adjusted according to the magnitude of the current being measured on site, by means of a DIP switch or by replacing it with a current sensor of the corresponding range.

[0080] 3. The high-speed ADC sampling circuit of the present invention adopts a high-speed ADC and a sample buffer, which greatly improves the system's response speed and sampling accuracy. It can capture rapid fluctuations at the millisecond level, ensuring the accuracy and reliability of the data.

[0081] 4. The MCU with machine learning algorithm function of the present invention combines historical data cleaning and comparative analysis with the establishment of specific models to quantitatively evaluate the performance of AO control loop, and can automatically detect abnormal waveforms, mark and record them, automatically judge and alarm for abnormalities, providing a reliable guarantee for the safety of automated control system.

[0082] 5. Host Computer: The host computer serves as the interface for modules and applications, completing data reading, analysis, processing, storage, and presentation. This enables monitoring of the AO control loop at multiple levels, including overall system operation, detection, transmission, and data analysis. It allows for timely detection and handling of anomalies, marking abnormal data as input for machine learning, and ultimately evaluating indicators such as the operational quality of the automation system.

[0083] All data collection and extraction in this invention are carried out under compliant and legal conditions.

Claims

1. A non-contact AO channel current monitoring and pre-diagnosis method, characterized in that, include: A non-contact open current sensor is connected to the output cable of the AO module without interrupting the circuit, and the current is sensed in real time and the corresponding voltage signal is generated in a physical isolation manner. The current signal is sampled by a high-speed ADC sampling circuit to obtain current waveform data, from which multi-dimensional waveform features are extracted. The machine learning diagnostic model, which is pre-trained based on historical data, analyzes the extracted waveform features to obtain the state diagnostic results of the AO module output channel. The machine learning diagnostic model is a hybrid architecture combining convolutional neural networks and long short-term memory networks. Based on the level of the status diagnosis result, a corresponding level of early warning or alarm is triggered.

2. The non-contact AO channel current monitoring and pre-diagnosis method according to claim 1, characterized in that, The hybrid architecture combining convolutional neural networks and long short-term memory networks includes: a one-dimensional convolutional layer, a bidirectional long short-term memory network layer, an attention mechanism layer, and a fully connected classification layer connected in sequence. The one-dimensional convolutional layer extracts local features of the current waveform, the bidirectional long short-term memory network layer captures the temporal dependencies of the waveform, the attention mechanism layer weights and highlights features of abnormal time periods, and the fully connected classification layer outputs the waveform state category and abnormality type.

3. The non-contact AO channel current monitoring and pre-diagnosis method according to claim 2, characterized in that, The triggering of the corresponding level of warning or alarm includes: When the waveform characteristics deviate from the normal template but do not exceed the safety threshold, it is recorded as an alert state and no reporting is triggered; when the waveform abnormality continues and the characteristics exceed the deviation threshold, a medium-level warning is triggered and alarm information is uploaded; when the waveform matches a known fault mode or exceeds the safety threshold, an advanced alarm is immediately triggered, and audible and visual prompts and remote notifications are supported.

4. The non-contact AO channel current monitoring and pre-diagnosis method according to claim 3, characterized in that, The training process of the machine learning diagnostic model includes: collecting historical normal waveform data and abnormal waveform data, labeling the abnormal waveform data with anomaly levels, using the labeled data to train the hybrid model, using a combination of cross-entropy loss and temporal smoothing regularization term as the loss function, and deploying the trained model to the main control chip after quantization, and updating the model through online incremental learning.

5. The non-contact AO channel current monitoring and pre-diagnosis method according to claim 4, characterized in that, The machine learning diagnostic model also includes: when an operator confirms that an alarm is a false alarm or a missed alarm, the corresponding waveform data is manually marked, the system triggers incremental learning based on the marked data, updates the machine learning diagnostic model, and the model is retrained periodically using newly collected data to adapt to equipment aging or changes in operating conditions.

6. A non-contact AO channel current monitoring and pre-diagnosis method according to claim 4 or 5, characterized in that, The machine learning diagnostic model further includes: synchronizing and aligning the current waveforms of multiple AO channels corresponding to the same actuator with timestamps, calculating the difference features, correlation features, and consistency features between the waveforms of multiple channels, and determining whether the fault is a common fault based on cross-channel features and preset fault rules, and locating the fault originating from the AO module or actuator.

7. A non-contact AO channel current monitoring and pre-diagnosis method according to claim 6, characterized in that, The machine learning diagnostic model further includes: calculating a channel health index based on historical waveform data. The health index is obtained by weighting multiple degradation features in waveform fidelity, noise level, response speed, and stability. The health index is then used for time-series prediction, extrapolating the time required for it to drop to the failure threshold to obtain the predicted remaining service life. Based on the prediction results, preventive maintenance recommendations are generated.

8. A non-contact AO channel current monitoring and pre-diagnosis method according to claim 1 or 7, characterized in that, The extraction of multi-dimensional waveform features includes: extracting time-domain and frequency-domain features of the current waveform; the time-domain features include statistical features, dynamic response features, and morphological and distribution features, the statistical features include mean and variance, the dynamic response features include rise time and fall time and overshoot amplitude, the morphological and distribution features include peak factor, impulse index, and waveform entropy; the frequency-domain features include the fundamental and harmonic amplitudes, total harmonic distortion rate, and spectral centroid extracted by Fourier transform.

9. A non-contact AO channel current monitoring and pre-diagnosis system, employing the non-contact AO channel current monitoring and pre-diagnosis method according to any one of claims 1-8, characterized in that, include: The non-contact open-type current sensor is directly fitted onto the output cable of the AO control circuit to measure the circuit current in a non-contact manner. The high-speed ADC sampling circuit has its input terminal electrically connected to the signal output terminal of the non-contact open current sensor to perform analog-to-digital conversion on the current signal. The main control chip has its input terminal connected to the output terminal of the high-speed ADC sampling circuit. It receives and analyzes the sampled data and generates diagnostic results and corresponding alarm information. The host computer communicates with the main control chip via a communication interface to receive alarm information and waveform data.

10. A non-contact AO channel current monitoring and pre-diagnosis system according to claim 9, characterized in that, The non-contact open-ended current sensor adopts an openable magnetic ring structure. Its outer shell is made of multiple layers of electromagnetic shielding material, with an inner layer of high magnetic permeability alloy, an outer layer of conductive coating and grounding, and a wave-absorbing material in the middle. The internal structure uses a Hall effect element or a magnetoresistive sensor. The non-contact open-ended current sensor is also equipped with multiple signal output interfaces, and its opening is equipped with a self-locking mechanism.

Images