Cabinet fault prediction and diagnosis method based on machine learning

Abstract

The invention discloses a machine learning-based cabinet fault prediction and diagnosis method, and relates to the technical field of cabinet fault diagnosis, and the method comprises the following steps: S100, data collection; s200, preprocessing the data; s300, constructing and training a fault classification and prediction model based on the labeled historical fault data; s400, inputting the real-time data into a fault classification and prediction model, analyzing the real-time data through the fault classification and prediction model, and predicting a fault type and an occurrence time interval; s500, matching rules in a knowledge base according to the prediction result, and generating a fault diagnosis suggestion; and S600, outputting a result through an interactive interface or a remote platform, and triggering an alarm mechanism based on the prediction probability. And S700, correcting the model based on prediction and diagnosis results and actual results. Through multi-sensor fusion and intelligent algorithm optimization, real-time monitoring and fault prediction of the operation state of the cabinet are realized, the accidental downtime is reduced, and meanwhile, the diagnosis efficiency is improved by combining fuzzy reasoning of the knowledge base and dynamic prediction of the fault time.

Description

Technical Field

[0001] The present invention relates to the technical field of cabinet fault diagnosis, and in particular to a cabinet fault prediction and diagnosis method based on machine learning. Background Art

[0002] As a key carrier of electronic equipment, the stable operation of cabinets is crucial to ensuring the proper functioning of electronic systems. Traditional cabinet fault detection relies primarily on regular manual inspections and simple threshold judgments. This approach is not only inefficient but also difficult to detect potential faults in advance. Repairs are often performed only after a fault has occurred, leading to system downtime and financial losses. With the advancement of data acquisition technology and machine learning algorithms, it has become possible to use real-time data and machine learning models to predict and diagnose cabinet faults. However, current methods still have shortcomings in data processing, model selection, and fault diagnosis accuracy.

[0003] Therefore, it is necessary to invent a cabinet fault prediction and diagnosis method based on machine learning to solve the above problems. Summary of the Invention

[0004] The present invention aims to provide a cabinet fault prediction and diagnosis method based on machine learning. By collecting cabinet operation data in real time and using machine learning models to predict and diagnose faults, the reliability of cabinet operation and maintenance efficiency can be improved.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a cabinet fault prediction and diagnosis method based on machine learning, comprising the following steps:

[0006] S100, real-time collection of cabinet operation data;

[0007] S200, cleaning, normalizing and feature extracting the collected data;

[0008] S300, using historical fault data to train machine learning models;

[0009] S400, inputting real-time data into the model and analyzing the real-time data through the model;

[0010] S500, provides fault diagnosis suggestions based on expert knowledge base;

[0011] S600: Output prediction and diagnosis results, and trigger an alarm mechanism based on the prediction probability;

[0012] S700: Modify the model based on the prediction and diagnosis results and the actual results.

[0013] Preferably, step S300 includes the following steps:

[0014] S310. Marking data according to historical fault records;

[0015] S320, select the random forest model;

[0016] S330, establishing multiple different decision trees to analyze historical faults, and using the labeled historical fault data to train the established models and optimize model parameters;

[0017] S340. Use test data to evaluate model performance and ensure prediction accuracy.

[0018] Preferably, in step S330, the model is constructed in the following manner:

[0019] S331. Set the number of input features to determine the decision structure of a node on the decision tree;

[0020] S332. Randomly extract a portion of data from the training samples to form a training set, and use the unselected samples to make predictions and evaluate the error.

[0021] S333. For each node, randomly select features, calculate the best splitting method, and select the attribute with the largest information gain after splitting for splitting;

[0022] S334. Repeat the above steps to build another decision tree until the number of decision tree items reaches a predetermined value, thus completing the model construction.

[0023] Preferably, in step S330, the following calculation method is used to determine the fault weight in the prediction sequence:

[0024]

[0025] Where P is the probability of failure; N = 5, corresponding to temperature (T), humidity (H), vibration (V), current (I), voltage (U); W i The dynamic weight of the i-th parameter; f(x i ) is the fault contribution function of the i-th parameter; ε is the environmental interference correction term.

[0026] Preferably, the historical fault characteristic quantities and target quantities are extracted from the historical fault types and the temperature, humidity, vibration, current, voltage and other parameters of the cabinet when the corresponding fault occurs.

[0027] Preferably, when collecting real-time operating data, the collected data includes temperature, humidity, vibration, current, and voltage, and signal amplification, filtering, and A / D conversion technology are combined to improve data accuracy.

[0028] Preferably, in step S200:

[0029] The data cleaning includes noise filtering and outlier detection. Noise filtering uses wavelet denoising or low-pass filtering algorithm, and outlier detection is based on isolation forest or threshold judgment method. Normalization processing selects minimum-maximum normalization or Z-score normalization according to the data type.

[0030] Preferably, in step S700, the actual fault parameters and fault types are compared with the data in the historical database to supplement the historical database.

[0031] Preferably, when supplementing the historical database, if the actual fault type does not coincide with the historical fault type in the historical database, a new fault type is created in the database, and the fault parameters and fault type are placed in the database; if the actual fault type coincides with the historical fault type in the historical database, the fault parameters are completed to the corresponding historical fault type.

[0032] Technical effects and advantages of the present invention:

[0033] The present invention realizes real-time monitoring of cabinet operation status and fault prediction through multi-sensor fusion and intelligent algorithm optimization, reducing unexpected downtime. At the same time, it combines knowledge base fuzzy reasoning to dynamically predict fault time and improve diagnostic efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 Schematic diagram of the diagnostic method of the present invention.

[0035] Figure 2 Schematic diagram of the model training process of the present invention.

[0036] Figure 3 Schematic diagram of the model construction process of the present invention. DETAILED DESCRIPTION

[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0038] First embodiment

[0039] To solve the problems that traditional cabinet fault detection methods mainly rely on regular maintenance and manual inspection, which are inefficient and cannot predict faults in real time; and existing fault diagnosis systems are mostly based on rules or simple threshold judgments, which are difficult to handle complex nonlinear fault modes and lack in-depth analysis of cabinet operation data and intelligent fault prediction capabilities.

[0040] The present invention provides Figures 1 to 3 A cabinet fault prediction and diagnosis method based on machine learning is shown, comprising the following steps:

[0041] S100, real-time collection of cabinet operation data;

[0042] S200, cleaning, normalizing and feature extracting the collected data;

[0043] S300, using historical fault data to train machine learning models;

[0044] S400, inputting real-time data into the model and analyzing the real-time data through the model;

[0045] S500, provides fault diagnosis suggestions based on expert knowledge base;

[0046] S600: Output prediction and diagnosis results, and trigger an alarm mechanism based on the prediction probability;

[0047] S700: Modify the model based on the prediction and diagnosis results and the actual results.

[0048] In this embodiment, in step S100: data acquisition is performed by collecting cabinet temperature, humidity, vibration, current, and voltage data through multiple patch sensor modules, and the communication module uses the Internet of Things protocol to transmit data.

[0049] It should be noted that cabinet data acquisition is a data acquisition module developed based on electronic technologies such as surface-mount temperature-sensing electronic components, surface-mount humidity-sensing electronic components, surface-mount vibration electronic components, surface-mount photosensitive electronic components, single-chip microcomputers, and supporting power amplifier processing circuits.

[0050] The collected data is transmitted to the data processing unit through a wired or wireless communication module.

[0051] After completing the cabinet data collection, it is necessary to use IoT transmission technology to transmit the data to the smart control cabinet health monitoring system. The use of IoT transmission technology can ensure that the data flow from each sensor to the central monitoring system is continuous, stable and secure.

[0052] In this embodiment, in step S200: data cleaning includes noise filtering and outlier detection. Noise filtering uses wavelet denoising or low-pass filtering algorithm, and outlier detection is based on isolation forest or threshold judgment method. Normalization processing selects minimum-maximum normalization or Z-score normalization according to the data type. By cleaning the collected data, noise and outliers in the data are removed to ensure the accuracy of the subsequent model.

[0053] It should be noted that outliers can be handled by deletion, replacement, or interpolation. The appropriate method should be selected based on the specific situation. For example, if there are a small number of outliers, they can be directly deleted; if there are a large number, they may need to be replaced with the mean or median, or interpolation may be used to estimate reasonable values.

[0054] When normalizing data, you need to convert the data to a specific range (such as [0, 1] or [-1, 1]) to improve model performance. Common methods include min-max normalization and Z-score normalization. The following uses Python's numpy library as an example to introduce these two methods:

[0055] Min-max normalization:

[0056] Map the data to the [0,1] interval. The formula is:

[0057] X_{norm}=\frac{X-X_{min}}{X_{max}-X_{min}}

[0058] Z-score standardization

[0059] Make the data have a mean of 0 and a standard deviation of 1. The formula is:

[0060] Z=\frac{X-\mu}{\sigma}

[0061] Where \mu is the mean and \sigma is the standard deviation.

[0062] It's important to note that in practical applications, the normalization method should be selected based on the data characteristics and model requirements. If the data has clear boundaries, min-max normalization is appropriate, such as with temperature, humidity, and brightness data. If the data distribution has no clear boundaries, z-score normalization is more commonly used, such as with vibration data.

[0063] In this embodiment, in step S300:

[0064] The machine learning model is a random forest, support vector machine, or neural network. Model parameter optimization is achieved through grid search or random search, and data imbalance is addressed by adjusting class weights or using sampling techniques. Historical fault data is annotated using time window alignment technology to associate fault events with sensor data. Annotation dimensions include fault type, severity level, and time tag.

[0065] It should be noted that when extracting key features, the statistical feature extraction method is used to extract the mean and variance of the data. The mean represents the average level of the data, and the variance represents the degree of dispersion of the data. When analyzing the original data (temperature, humidity, vibration, brightness), the mean and variance can be used to quickly understand the overall level and fluctuation of the data and determine whether the sensor is working properly.

[0066] In this embodiment, step S300 includes the following steps:

[0067] S310. Marking data according to historical fault records;

[0068] S320, select the random forest model;

[0069] S330, establishing multiple different decision trees to analyze historical faults, and using the labeled historical fault data to train the established models and optimize model parameters;

[0070] S340. Use test data to evaluate model performance and ensure prediction accuracy.

[0071] In S301, the preprocessed historical data records system or device failures. This data is labeled to indicate whether a failure has occurred and the type of failure. Useful information is extracted from the failure records, and the failure is associated with the corresponding data time period. The time window problem in the labeling is also addressed. Through data labeling, historical failure records are converted into labeled data that can be used by machine learning models.

[0072] In S302, a suitable machine learning algorithm is selected, such as a random forest, support vector machine (SVM), neural network, etc. After comparing various machine learning algorithms, the present invention selects random forest as the learning algorithm. The characteristics of random forest are that it can take advantage of centralized learning for medium-sized data; it can provide robust prediction results for regression tasks and can handle noise and outliers in the data; and its model performance is high accuracy, fast training speed, and good interpretability. Therefore, the random forest algorithm is more suitable for this application.

[0073] In this embodiment, the model of step S330 is constructed in the following manner:

[0074] S331. Set the number of input features to determine the decision structure of a node on the decision tree;

[0075] S332. Randomly extract a portion of data from the training samples to form a training set, and use the unselected samples to make predictions and evaluate the error.

[0076] S333. For each node, randomly select features, calculate the best splitting method, and select the attribute with the largest information gain after splitting for splitting;

[0077] S334. Repeat the above steps to build another decision tree until the number of decision tree items reaches a predetermined value, thus completing the model construction.

[0078] In this embodiment, the following calculation method is used to determine the fault weight in the prediction sequence:

[0079]

[0080] Where P is the probability of failure; N = 5, corresponding to temperature (T), humidity (H), vibration (V), current (I), voltage (U); W i The dynamic weight of the i-th parameter; f(x i ) is the fault contribution function of the i-th parameter; ε is the environmental interference correction term.

[0081] Dynamic weight W i The calculation method is:

[0082]

[0083] Among them, H(x i )=-∑p(x i )lnp(x i ) is the parameter information entropy. The smaller the entropy value, the greater the parameter variation and the higher the weight. For example, temperature mutation has more warning value than stable voltage. n is the fault type; j is the jth evaluation index (such as temperature, humidity, vibration, etc.). Assuming there are m indicators in total, the value range of j is 1 to m, p(x i ) is the normalized proportion of the i-th sample value under the j-th indicator, that is, the contribution probability of the sample in the indicator.

[0084] If temperature and vibration are used as evaluation indicators (j = 1 and j = 2), by calculating the p(x i ) and entropy values, the weights can be adjusted dynamically so that abnormal temperatures (high variation) receive higher weights, thereby improving prediction accuracy.

[0085] The single-parameter contribution function is calculated using normalization and logistic regression transformation, as follows:

[0086]

[0087] Among them, θ i is the parameter threshold, k i is the slope coefficient.

[0088] The correlation correction factor is based on the Bayesian conditional probability optimization synergy effect, which is as follows:

[0089] ε = α·∏ m∈S P(F丨x m )

[0090] Among them, S is the parameter combination with strong correlation, and α is an empirical coefficient determined by analyzing historical fault data.

[0091] It should be noted that grid search traverses all possible parameter combinations, with a large computational amount but being comprehensive; random search randomly selects parameter combinations and may find better parameters in a shorter time. Therefore, different optimization methods are selected according to different characteristic parameters. In addition, due to the possible problem of class imbalance in fault data, that is, the number of normal samples is much larger than that of fault samples, class imbalance needs to be processed at this time. The class_weight parameter in random forest can be set to 'balanced', or the class weights can be adjusted manually, such as giving higher weights to the minority classes.

[0092] In S304, test data is used to evaluate the model performance to ensure prediction accuracy. In terms of model evaluation, comprehensive evaluation is carried out from aspects such as precision, recall, and F1 score. At the same time, the ROC-AUC curve is used to evaluate the overall performance of the model.

[0093] In this embodiment, in step S600, the output results are displayed in the form of visual charts, including the fault probability trend chart, the heat map of sensor data, and the maintenance strategy decision tree; some tasks from steps S100 to S600 are deployed on the local cabinet device through the edge computing architecture to achieve low-latency processing.

[0094] In this embodiment, in step S700, the actual fault parameters and fault types are compared with the data in the historical database to supplement the historical database; when supplementing the historical database, if the actual fault type does not coincide with the historical fault types in the historical database, a new fault type is created in the database, and the current fault parameters and fault type are put into the database; if the actual fault type coincides with the historical fault types in the historical database, the current fault parameters are completed for the corresponding historical fault type.

[0095] It should be noted that the method includes an alarm function, which automatically triggers an alarm when a serious fault is predicted.

[0096] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A cabinet fault prediction and diagnosis method based on machine learning, characterized in that: The following steps are involved: S100, real-time collection of cabinet operation data; S200, cleaning, normalizing and feature extracting the collected data; S300, using historical fault data to train machine learning models; S400, inputting real-time data into the model and analyzing the real-time data through the model; S500, provides fault diagnosis suggestions based on expert knowledge base; S600: Output prediction and diagnosis results, and trigger an alarm mechanism based on the prediction probability; S700: Modify the model based on the prediction and diagnosis results and the actual results.

2. The fault prediction and diagnosis method according to claim 1, characterized in that: Step S300 includes the following steps: S310. Marking data according to historical fault records; S320, select the random forest model; S330, establishing multiple different decision trees to analyze historical faults, and using the labeled historical fault data to train the established models and optimize model parameters; S340. Use test data to evaluate model performance and ensure prediction accuracy.

3. The fault prediction and diagnosis method according to claim 2, characterized in that: In step S330, the model is constructed in the following manner: S331. Set the number of input features to determine the decision structure of a node on the decision tree; S332. Randomly extract a portion of data from the training samples to form a training set, and use the unselected samples to make predictions and evaluate the error. S333. For each node, randomly select features, calculate the best splitting method, and select the attribute with the largest information gain after splitting for splitting; S334. Repeat the above steps to build another decision tree until the number of decision tree items reaches a predetermined value, thus completing the model construction.

4. The fault prediction and diagnosis method according to claim 3, characterized in that: In step S330, the following calculation method is used to determine the fault weight in the prediction sequence: Where P is the probability of failure; N=5, corresponding to temperature (T), humidity (H), vibration (V), current (I), and voltage (U); The dynamic weight of the i-th parameter; is the fault contribution function of the i-th parameter; is the environmental interference correction item.

5. The fault prediction and diagnosis method according to claim 4, characterized in that: The historical fault characteristic quantities and target quantities are extracted from the historical fault types and the temperature, humidity, vibration, current, voltage and other parameters of the cabinet when the corresponding fault occurred.

6. The fault prediction and diagnosis method according to claim 5, characterized in that: When collecting real-time operating data, the collected data includes temperature, humidity, vibration, current, and voltage, and the data accuracy is improved by combining signal amplification, filtering, and A / D conversion technology.

7. The fault prediction and diagnosis method according to claim 6, characterized in that: In step S200: The data cleaning includes noise filtering and outlier detection. Noise filtering uses wavelet denoising or low-pass filtering algorithm, and outlier detection is based on isolation forest or threshold judgment method. Normalization processing selects minimum-maximum normalization or Z-score normalization according to the data type.

8. The fault prediction and diagnosis method according to claim 7, characterized in that: In step S700, the actual fault parameters and fault types are compared with the data in the historical database, and the historical database is supplemented.

9. The fault prediction and diagnosis method according to claim 8, characterized in that: When supplementing the historical database, if the actual fault type does not coincide with the historical fault type in the historical database, a new fault type will be created in the database, and the fault parameters and fault type will be added to the database; If the actual fault type coincides with the historical fault type in the historical database, the fault parameters will be added to the corresponding historical fault type.

Images