Low-cost intelligent monitoring method for medium-low value non-standard equipment
By reusing data collected by the PLC system and combining it with unsupervised and supervised learning models, the problem of anomaly detection in low- and medium-value non-standard equipment was solved. This enabled low-cost, rapid deployment, and efficient anomaly detection, reducing hardware costs and deployment cycles, and improving detection accuracy and production continuity.
Patent Information
- Application Number
- CN202511583358.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-31
- Publication Date
- 2026-02-03
AI Technical Summary
Low- and medium-value non-standard equipment suffers from high hardware costs, poor adaptability to non-standard scenarios, and long deployment cycles in anomaly detection, resulting in a high rate of missed anomaly detections/false alarms.
By reusing the PLC system to collect motor current and load data, and using a combination of unsupervised and supervised learning models, data feature extraction and fault diagnosis are performed, along with dynamic resource management, enabling real-time anomaly detection without the addition of new hardware.
It enables low-cost, rapid deployment and efficient anomaly detection, reducing hardware costs, shortening deployment cycles, and improving the accuracy of anomaly detection and production continuity.
Smart Images
Figure CN121455047A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent monitoring technology for industrial equipment, specifically involving a low-cost intelligent monitoring method and supporting system for low- to medium-value, non-standardized industrial equipment, enabling real-time anomaly detection and predictive maintenance without additional hardware investment. Background Technology
[0002] Existing low- and medium-value equipment (such as 3C packaging machines and general servo motor production lines) often lacks monitoring devices or has only a few monitoring devices installed, making anomaly detection difficult and resulting in a high rate of missed or false alarms.
[0003] Existing industrial equipment monitoring solutions have the following core flaws: 1. Cost constraint contradiction: Due to the low value of a single unit of low- and medium-value equipment, traditional monitoring devices (with added vibration / pressure sensors) make the overall system more complex after adding monitoring devices, and the hardware and implementation costs account for too high a proportion, resulting in very low willingness of enterprises to deploy them.
[0004] 2. Poor adaptability to non-standard scenarios: Non-standard equipment has diverse and unknown failure modes, and supervised learning models that rely on fault labeling data have poor generalization ability; while existing unsupervised models have not solved the problems of "insufficient feature discrimination and difficulty in dynamically adjusting thresholds", resulting in a high rate of false alarms / false positives.
[0005] 3. Long deployment cycle: Hardware debugging and complex model training take several months, making it difficult to quickly respond to the needs of intelligent transformation of production lines. Summary of the Invention
[0006] The technical problem this invention aims to solve is: how to overcome the monitoring barriers of low- and medium-value, non-standard equipment, and achieve: ① Zero hardware addition: Reuse the existing PLC system to collect motor current, load and cylinder flow rate data, completely avoiding the investment of additional sensors; ② Adaptation to non-standard scenarios: Construct a "normal working condition baseline" through unsupervised learning, combine dynamic feature engineering to deal with unknown faults, and support accurate supplementation by supervised models; ③ Dynamic resource scheduling: Select data acquisition and processing strategies based on the real-time performance of the industrial control computer to ensure system stability and effectiveness; ④ Rapid deployment: Complete normal data collection and model deployment within 1-2 weeks, supporting continuous optimization through online learning.
[0007] To solve the above problems, the present invention is achieved through the following technical solution: A low-cost intelligent monitoring method for low- and medium-value non-standard equipment includes the following steps: Step 1: Data Acquisition: The PLC system acquires the phase current and average load rate of the motors in the industrial equipment; Step 2, Periodic Feature Engineering: Step 2.1: Servo motors are mostly used in industrial equipment. The motor operation cycle is identified based on the state machine model. The servo motor operation cycle is divided into two types: one is a complex cycle: phase 0 is a complete stop, phase 1 executes an action, phase 2 is a stage stop, and phase 3 is a return action; the other is a simple cycle: phase 0 is a complete stop, and phase 1 executes an action. Step 2.2: Calculate the positioning state transition point through vectorization and verify the legality of the transition; based on the servo motor's operating cycle pattern and the legality verification results, filter out insufficient sample stages and invalid cycles; Step 2.3: Identify and output the stage boundaries and basic statistics for each servo motor operating cycle. The stage boundaries are the start / end times, and the basic statistics include the sample index, duration, and total number of samples. Step 2.4: Multi-dimensional feature extraction: Extract multi-dimensional features for two or four stages of each servo motor's operating cycle; Step 3: Intelligent monitoring using a hybrid model; Step 3.1: Based on the extracted multi-dimensional features, use supervised and / or unsupervised models for identification; Step 3.2: Based on the identification results, perform fault warning and diagnosis: establish a normal state characteristic model and compare the operating data in real time; Step 4: Dynamic resource management, uploading monitoring results to the cloud database in real time; Step 5: Output the results.
[0008] During data collection, a layered collection strategy is adopted, which combines continuous collection and polling collection: Continuous data collection of key components: Fault tree analysis is used to screen the core paths of the equipment and to continuously monitor them to ensure zero missed detection of core faults; the core paths of the equipment include, but are not limited to: the main drive motor and its associated mechanisms. Non-critical component polling data collection: Collect short time window data in a polling manner at configurable time intervals.
[0009] It also includes step 2.5: binding the extracted multi-dimensional features to the stages in the cycle according to the servo motor's operating cycle.
[0010] For non-standard equipment and low-to-medium value equipment, an unsupervised model is used; for standardized equipment, a supervised model is used.
[0011] Step 3.2 specifically involves: 1) If the peak current continues to exceed the upper limit during the execution phase of phase 1, it indicates that the mechanical load is too large; 2) If the load rate fluctuation range suddenly increases during the stop in phase 2, it indicates wear of the braking components; 3) If the peak timing difference of multiple motors increases during the execution of actions in phase 1, it indicates that the coordinated action has lost synchronization; 4) Energy consumption optimization and performance improvement: Analyze the energy consumption ratio at each stage, reduce ineffective energy consumption by optimizing motion parameters; adjust servo parameters based on motor load rate distribution to avoid long-term high-load operation; 5) Life prediction and maintenance decision-making: Long-term tracking of characteristic trends, combined with motor life model to predict remaining life; triggering preventive maintenance through abnormal frequency statistics.
[0012] Step 4 specifically involves: Real-time resource monitoring: The psutil library is used to monitor the memory and CPU usage of the industrial control computer in real time, and processing is paused when the threshold is exceeded. Dynamic data chunking: Automatically calculates the size of data chunks based on available memory, and optimizes memory usage by combining pandas streaming processing; State persistence: The last processing end time is saved via pickle, supporting incremental data processing and avoiding duplicate calculations of historical data.
[0013] Compared with the prior art, the present invention has the following beneficial effects: 1. Dramatically reduced costs; 2. Hardware cost: Reusing existing PLC data channels; 3. Data cost: Unsupervised models only require normal data, reducing annotation costs by more than 80%; 4. Implementation period: Deployment can be completed in 1-2 weeks, which is 70% shorter than traditional solutions.
[0014] 5. Breakthrough in scene adaptability; 6. Non-standard equipment: Unsupervised models do not require fault data and adapt to unknown faults through dynamic baselines; 7. Standardized equipment: The monitoring model accurately classifies faults and guides maintenance decisions; 8. Resource constraints: Dynamic scheduling strategies are adapted to the performance of industrial control computers to ensure long-term stable operation; 9. A leap in operational efficiency; 10. Shifting from reactive maintenance to predictive maintenance reduces downtime by 60%-80%; 11. Supports tiered anomaly response (early warning / shutdown) to balance production continuity and losses. Attached Figure Description
[0015] Figure 1 This is an overall flowchart of the present invention; Figure 2Flowchart for feature extraction in the periodic stage; Figure 3 This is a diagram showing the system architecture and module interaction relationships. Detailed Implementation
[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0017] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0018] like Figure 1 As shown, the low-cost intelligent monitoring method for low- and medium-value non-standard equipment includes the following steps: Step 1: Data Acquisition: The PLC system acquires the motor phase current and average load rate of the industrial equipment. The sampling frequency is fixed at 100Hz to satisfy the Nyquist sampling theorem and capture high-frequency operating condition changes.
[0019] The effective value of phase current is a direct representation of motor energy consumption and load intensity. Its dynamic characteristics have a clear mapping relationship with mechanical faults, and its numerical changes can effectively reflect the sudden changes in resistance and abnormal impacts of the transmission mechanism.
[0020] Average load rate: Reflects the ratio of the actual load of the servo motor to the rated load. It is a sensitive indicator for monitoring abnormal mechanical resistance. By observing the load fluctuation characteristics, faults such as wear of transmission components and abnormal tension can be identified.
[0021] Better yet, a layered data collection strategy should be adopted during data acquisition, combining continuous and polling methods: Continuous monitoring of key components: Fault tree analysis (FTA) is used to screen the core paths of the equipment and continuously monitor them to ensure zero missed detection of core faults; the core paths of the equipment include, but are not limited to: the main drive motor and its associated mechanisms; Non-critical component polling data collection: Data is collected in short time windows (e.g., 10 seconds) at configurable time intervals (e.g., 10 minutes), balancing comprehensive monitoring with industrial control computer resource consumption. In this way, the present invention uses time-segmented scheduling (e.g., cron tasks) to poll all motors / cylinders at intervals (e.g., 10 minutes), collecting data for short time windows (e.g., 10 seconds) each time. This is suitable for scenarios with many and extremely dispersed devices (e.g., discrete machine clusters), balancing comprehensive coverage with industrial control computer performance (memory / CPU), and the cost is dynamically adjusted according to the polling interval (the longer the interval, the lower the cost).
[0022] Step 2, Periodic Feature Engineering: Step 2.1: Unlike wind power systems, industrial equipment mostly uses servo motors. Servo motors start and stop continuously according to operating conditions. Therefore, the motor operation cycle is identified based on the state machine model. The servo motor operation cycle is divided into two types: one is a complex cycle: phase 0 is a complete stop, phase 1 executes an action, phase 2 is a stage stop, and phase 3 is a return action (which can be abbreviated as: 0→1→2→3); the other is a simple cycle: phase 0 is a complete stop, and phase 1 executes an action (which can be abbreviated as: 0→1).
[0023] Step 2.2: Locate the state transition point through vectorized calculation (np.where) and verify the legality of the transition. According to the definition in this patent application, the sequence of the running cycle is legal, and the rest is illegal. For example, 0→1 is legal, and 1→0 is illegal. Based on the running cycle pattern of the servo motor and the legality verification results, filter out the insufficient sample stage and invalid cycle. Step 2.3: Identify and output the stage boundaries and basic statistics for each servo motor operating cycle. The stage boundaries are the start / end times, and the basic statistics include the sample index, duration, and total number of samples. Step 2.4: Multi-dimensional Feature Extraction: For each servo motor's operating cycle, 69 features are extracted across the four phases (phase 0: overall stop; phase 1: action execution; phase 2: phase stop; phase 3: return action), covering: Basic statistics: mean, maximum and minimum values, peak-to-peak value (reflecting load / current intensity); Volatility characteristics: variance, standard deviation, root mean square (RMS, reflecting the degree of volatility); Distribution characteristics: skewness (distribution symmetry), kurtosis (the steepness of the peak value, to identify anomalous peaks).
[0024] For a servo motor with four stages in a cycle, each stage has two parameters: motor phase current and average load rate. Each parameter has at least eight features in a stage (3 basic statistical features + 3 fluctuation features + 2 distribution features), which is 8×8=64 features. In addition, there is the duration of each stage, four stages with four times, and one cycle with one total time, which adds up to a total of 69 features. One cycle has 69 features.
[0025] For a servo motor with two stages in one cycle, there are 35 features.
[0026] Even better, step 2.5 includes: binding the extracted multi-dimensional features to phases within the servo motor's operating cycle. For example, binding the focus current peak to phase 1 (the action phase) and binding load stability to phase 2 (the stopping phase). For instance, regarding the aforementioned fluctuation characteristics, smaller load fluctuations indicate a more stable load. Step 3: Hybrid model intelligent monitoring (AI algorithm); Step 3.1: Based on the extracted multi-dimensional features, use supervised and / or unsupervised models for identification; ① Unsupervised model as the main method (adaptation to non-standard equipment): Using algorithms such as Isolation Forest and One-Class SVM, a "normal behavior baseline" is constructed only through normal operating data (such as data during equipment trial operation and stable production periods) to identify anomalies that deviate from the baseline.
[0027] This unsupervised model is applicable to non-standard equipment and low-to-medium value equipment. Non-standard equipment: Failure modes are unknown or diverse (such as customized packaging machines), and can be deployed without fault data; Low- to medium-value equipment: Fault simulation costs are high (e.g., artificially created jams may damage the equipment), while normal data is easy to obtain (trial operation period, stable production period); Technical advantages: Fast deployment: No annotation required, normal data collection and model training can be completed in 1-2 weeks; Better support for online learning: Regularly incorporate new normal data to update the model and dynamically track gradual changes in equipment parameters (such as increased current fluctuations caused by bearing aging).
[0028] ② Supervised model as a supplement (filling in for standard equipment): For standard equipment with fixed failure modes, use labeled failure data (such as jamming, overload) to train SVM and random forest models to achieve accurate failure classification (such as distinguishing between "jamming" and "abnormal load") and guide maintenance decisions (random forest, support vector machine, gradient boosting tree).
[0029] This supervisory model is adapted to standard equipment and business value: Standardized equipment: Fault modes are fixed (such as jamming or overload of standard model motors), and it is easy to collect and label data; Business value: After the equipment is finalized, precise fault classification guides maintenance decisions (e.g., jamming requires clearing foreign objects, overload requires checking the load), improving maintenance efficiency.
[0030] After the new machine is produced, 1. First, use an unsupervised model to perform anomaly identification (the unsupervised model only needs normal data and only judges whether it is normal or abnormal). 2. As the machine's operating time increases, the amount of data (including normal and abnormal data) increases. A supervised model is then used (the supervised model requires a large amount of abnormal and normal data. Its effect is that it can not only identify normal and abnormal data, but also determine which type of abnormal situation it is).
[0031] Step 3.2: Based on the identification results, perform fault early warning and diagnosis: establish a normal state characteristic model and compare the operating data in real time. 1) If the peak current continues to exceed the upper limit during the execution phase of phase 1, it indicates that the mechanical load is too large (such as insufficient lubrication of the guide rail). 2) If the load rate fluctuation range suddenly increases during the stop in phase 2, it indicates wear of the braking components; 3) If the peak timing difference of multiple motors increases during the execution of the action in phase 1, it indicates that the coordinated action has lost synchronization (affecting packaging accuracy).
[0032] 4) Energy consumption optimization and performance improvement: Analyze the energy consumption ratio of each stage (e.g., phase 1 energy consumption accounts for 70% of the total cycle energy consumption), and reduce ineffective energy consumption by optimizing motion parameters (e.g., reducing the acceleration slope of phase 5); based on the motor load rate distribution, adjust servo parameters (e.g., torque limiting) to avoid long-term high-load operation. 5) Life prediction and maintenance decision-making: Long-term tracking of characteristic trends (such as the increasing curve of the average current in phase 0 with the running time), combined with the motor life model to predict the remaining life, where the motor life model is a known model; triggering preventive maintenance (such as replacing bearings) by statistical analysis of abnormal frequency (such as the number of abnormal times of a certain motor in a month > 5 times).
[0033] Step 4: Dynamic resource management, uploading monitoring results to the cloud database in real time; Real-time resource monitoring: The psutil library monitors the industrial control computer's memory usage (≤80%) and CPU usage (≤80%) in real time. Processing is paused when the threshold is exceeded to prevent system crashes.
[0034] Dynamic data chunking: Automatically calculates the data chunk size based on available memory (e.g., 100MB of memory corresponds to 100,000 rows of chunks), and optimizes memory usage by combining pandas streaming processing (chunksize).
[0035] State persistence: The last processing end time is saved via pickle, supporting incremental data processing and avoiding duplicate calculations of historical data.
[0036] Step 5: Output the results.
[0037] The supporting system of this invention includes the following modules to achieve efficient collaboration: Configuration management module: centrally manages database parameters (server, driver, table name), sampling strategies (polling interval, time window), resource thresholds (memory, CPU limit), and feature and state rules (valid state, transition logic), supporting rapid adaptation across scenarios.
[0038] Data acquisition module: This module interfaces with an SQL Server database to acquire existing data and schedule tiered data collection strategies, outputting standardized time-series data. Alternatively, other databases, such as MySQL, can also be used.
[0039] Feature Engineering Module: Completes period segmentation, multi-dimensional feature extraction and data type optimization (such as downcast compression of numeric types, low cardinality string to category conversion), reducing memory usage by 30%-50%.
[0040] Model module: Includes unsupervised and supervised models, which will judge based on the input features and output abnormal results (including time interval, stage, and feature deviation description).
[0041] Resource Management Module: Monitors system resources, dynamically adjusts block size and garbage collection, and ensures stable operation.
[0042] Results output module: Supports saving feature and anomaly data in Parquet (high-efficiency compression), Excel, and CSV formats, and links to the original data index for easy traceability and analysis.
[0043] Based on the above data system, intelligent management of the entire equipment lifecycle can be supported: product quality is closely linked to the execution standards of each process, and real-time monitoring of whether the operation data of each process meets the standards is the core tool for accurately controlling production quality and promoting continuous quality improvement.
[0044] The beneficial effects of this invention are as follows: 1. With more monitoring parameters and more extracted feature dimensions, it can better reflect the characteristics of the original data, the status of the organization, and better detect and identify abnormal states; 2. The cycle division is more refined, with stages 0, 1, 2, and 3 representing the operational stages of the mechanism. This is more precise and accurate, and it can detect more anomalies and is more sensitive. 3. Anomaly detection is improved to include both supervised and unsupervised models, which are better able to discover the non-linear relationship between normal and abnormal data features, making it more accurate. At the same time, the model can continuously learn and improve its performance.
[0045] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several changes and improvements without departing from the overall concept of the present invention, and these should also be considered within the scope of protection of the present invention.
Claims
1. A low-cost intelligent monitoring method for low- and medium-value non-standard equipment, characterized in that: Includes the following steps: Step 1: Data Acquisition: The PLC system acquires the phase current and average load rate of the motors in the industrial equipment; Step 2, Periodic Feature Engineering: Step 2.1: Servo motors are mostly used in industrial equipment. The motor operation cycle is identified based on the state machine model. The servo motor operation cycle is divided into two types: one is a complex cycle: phase 0 is a complete stop, phase 1 executes an action, phase 2 is a stage stop, and phase 3 is a return action; the other is a simple cycle: phase 0 is a complete stop, and phase 1 executes an action. Step 2.2: Calculate the positioning state transition point through vectorization and verify the legality of the transition; based on the servo motor's operating cycle pattern and the legality verification results, filter out insufficient sample stages and invalid cycles; Step 2.3: Identify and output the stage boundaries and basic statistics for each servo motor operating cycle. The stage boundaries are the start / end times, and the basic statistics include the sample index, duration, and total number of samples. Step 2.4: Multi-dimensional feature extraction: Extract multi-dimensional features for two or four stages of each servo motor's operating cycle; Step 3: Intelligent monitoring using a hybrid model; Step 3.1: Based on the extracted multi-dimensional features, use supervised and / or unsupervised models for identification; Step 3.2: Based on the identification results, perform fault warning and diagnosis: establish a normal state characteristic model and compare the operating data in real time; Step 4: Dynamic resource management, uploading monitoring results to the cloud database in real time; Step 5: Output the results.
2. The low-cost intelligent monitoring method for low- and medium-value non-standard equipment according to claim 1, characterized in that: During data collection, a layered collection strategy is adopted, which combines continuous collection and polling collection: Continuous data collection of key components: Fault tree analysis is used to screen the core paths of the equipment and to continuously monitor them to ensure zero missed detection of core faults; the core paths of the equipment include, but are not limited to: the main drive motor and its associated mechanisms. Non-critical component polling data collection: Collect short time window data in a polling manner at configurable time intervals.
3. The low-cost intelligent monitoring method for low- and medium-value non-standard equipment according to claim 1, characterized in that: It also includes step 2.5: binding the extracted multi-dimensional features to the stages in the cycle according to the servo motor's operating cycle.
4. The low-cost intelligent monitoring method for low- and medium-value non-standard equipment according to claim 1, characterized in that: For non-standard equipment and low-to-medium value equipment, an unsupervised model is used; for standardized equipment, a supervised model is used.
5. The low-cost intelligent monitoring method for low- and medium-value non-standard equipment according to claim 1, characterized in that: Step 3.2 specifically involves: 1) If the peak current continues to exceed the upper limit during the execution phase of phase 1, it indicates that the mechanical load is too large; 2) If the load rate fluctuation range suddenly increases during the stop in phase 2, it indicates wear of the braking components; 3) If the peak timing difference of multiple motors increases during the execution of actions in phase 1, it indicates that the coordinated action has lost synchronization; 4) Energy consumption optimization and performance improvement: Analyze the energy consumption ratio at each stage, reduce ineffective energy consumption by optimizing motion parameters; adjust servo parameters based on motor load rate distribution to avoid long-term high-load operation; 5) Life prediction and maintenance decision-making: Long-term tracking of characteristic trends, combined with motor life model to predict remaining life; triggering preventive maintenance through abnormal frequency statistics.
6. The low-cost intelligent monitoring method for low- and medium-value non-standard equipment according to claim 1, characterized in that: Step 4 specifically involves: Real-time resource monitoring: The psutil library is used to monitor the memory and CPU usage of the industrial control computer in real time, and processing is paused when the threshold is exceeded. Dynamic data chunking: Automatically calculates the size of data chunks based on available memory, and optimizes memory usage by combining pandas streaming processing; State persistence: The last processing end time is saved via pickle, supporting incremental data processing and avoiding duplicate calculations of historical data.