Big data equipment fault prediction method based on machine learning

Abstract

The invention discloses a big data equipment fault prediction method based on machine learning, and the method comprises the steps: S1, collecting data in the operation process of equipment, and carrying out the preprocessing of the collected data; s2, multi-dimensional features in the preprocessed equipment data are extracted, and statistical characteristics are calculated; s3, processing the equipment data by adopting a time sequence analysis method, and generating a time sequence feature matrix; s4, processing the input time sequence feature matrix by using the improved Transform model, and constructing a fault prediction model; s5, monitoring the equipment in real time by using the trained fault prediction model; s6, fault early warning information is sent to operation and maintenance personnel according to a prediction result; and S7, continuously optimizing and updating the fault prediction model according to real-time monitoring and early warning feedback. According to the method, an efficient and scientific optimization scheme can be provided in big data equipment fault prediction, and remarkable technical values and economic benefits are brought to practical application.

Description

Technical Field

[0001] The present invention relates to the field of industrial equipment health management, and in particular to a big data equipment fault prediction method based on machine learning. Background Art

[0002] With the continuous advancement of industrial automation and intelligence, modern equipment is increasingly being used in production processes. However, equipment failure remains a major issue affecting production efficiency and operational stability. Traditional equipment fault detection methods mostly rely on experience-based preventive maintenance or regular inspections, which have many limitations. First, traditional methods rely on manual experience and cannot detect potential equipment failures in a timely manner. Second, preventive maintenance often fails to provide personalized diagnosis based on the actual operating status of the equipment, resulting in a large amount of unnecessary maintenance and inspections, which wastes manpower and material resources. Furthermore, traditional methods take a long time to predict failures, making real-time monitoring and rapid response impossible. As a result, once equipment failure occurs, it often leads to serious consequences such as production stagnation and equipment damage.

[0003] With the continuous development of big data and artificial intelligence technologies, machine learning-based equipment fault prediction methods have gradually become a research hotspot. By leveraging real-time equipment operating data and historical fault records, machine learning methods can achieve early warning of faults and provide efficient fault detection and prediction. These methods typically include steps such as data preprocessing, feature extraction, model training, and prediction. Deep learning and ensemble learning algorithms are widely used to build fault prediction models. However, existing machine learning-based equipment fault prediction methods still have some problems and limitations.

[0004] First, existing equipment fault prediction methods have certain defects in feature extraction and data fusion. Many methods rely only on a single data source of the equipment, such as sensor data or historical fault records, and ignore the potential of multi-dimensional data fusion. In actual equipment operation, environmental conditions, the status of equipment components, and the relationship between different data sources have a significant impact on fault prediction. Therefore, fault prediction based only on a single data source cannot fully reflect the health status of the equipment, which may lead to missed detections and misjudgments.

[0005] Secondly, traditional fault prediction models are relatively simple in terms of fault warning and anomaly detection, typically determining faults based on fixed thresholds. While this approach is simple to use, for complex industrial equipment, the operating status of the equipment often changes dynamically, and a single fixed threshold cannot effectively adapt to the changes in different equipment and operating environments. This results in low sensitivity and accuracy in fault warnings, making it difficult to detect potential equipment failures in a timely manner.

[0006] Therefore, the existing equipment fault prediction methods based on machine learning are in urgent need of improvement, especially in data fusion, feature extraction, dynamic adjustment of thresholds, improvement of fault warning accuracy and real-time monitoring. In order to overcome the shortcomings of the existing technology, a equipment fault prediction method based on deep learning and multi-dimensional data fusion is proposed. This method combines equipment historical data, real-time sensor data and environmental data to form a more comprehensive feature matrix to improve the accuracy of the prediction results. At the same time, by dynamically adjusting the fault warning threshold and combining the anomaly detection model, early warning and precise positioning of equipment failures can be achieved, thereby providing operation and maintenance personnel with more accurate fault information and improving fault handling efficiency. Through this improvement, the health status of the equipment can be continuously monitored at all stages of equipment operation, and the prediction model can be optimized in real time, thereby reducing the frequency of equipment failures and maintenance costs, and ensuring efficient and stable production operation. Summary of the Invention

[0007] One purpose of the present invention is to propose a big data equipment fault prediction method based on machine learning. The present invention can provide an efficient and scientific optimization solution in big data equipment fault prediction, bringing significant technical value and economic benefits to practical applications.

[0008] A big data device fault prediction method based on machine learning according to an embodiment of the present invention includes the following steps:

[0009] S1. Collect sensor data, operation and maintenance logs, and historical fault records during equipment operation, and pre-process the collected equipment data;

[0010] S2. Extract multidimensional features from the preprocessed device data, calculate statistical characteristics, and obtain a unified feature matrix;

[0011] S3. Use the time series analysis method to process the unified feature matrix, add timestamp information and apply position coding to generate a time series feature matrix;

[0012] S4. Use the improved Transformer model based on the hybrid expert model and graph convolutional network to process the input time series feature matrix, extract long-term dependencies and local features, optimize model parameters through multiple rounds of training, and build a fault prediction model;

[0013] S5. Use the trained fault prediction model to monitor the equipment in real time, predict the probability of equipment failure and generate prediction results;

[0014] S6. Based on the prediction results and the set fault warning threshold, the alarm mechanism is triggered and fault warning information is sent to the operation and maintenance personnel;

[0015] S7. Continuously optimize and update the fault prediction model based on real-time monitoring and early warning feedback.

[0016] Optionally, the S1 includes the following steps:

[0017] S11. Collect operating data from the device's sensors, including temperature, current, voltage, vibration, and pressure characteristics, and obtain the device's operation and maintenance logs and historical fault records;

[0018] S12. Clean the collected equipment operation data to remove missing values, outliers, and duplicate data. Missing values are filled using linear interpolation, and outliers are identified and removed using an outlier detection method based on a box plot.

[0019] S13. Normalize the cleaned data:

[0020]

[0021] Where x is the original data, x' is the normalized data, min(x) is the minimum value in the data, and max(x) is the maximum value in the data.

[0022] Optionally, the S2 includes the following steps:

[0023] S21. Extracting multiple characteristic variables from the equipment operation data after cleaning, including temperature, current, voltage, vibration, and pressure;

[0024] S22. Calculate statistical characteristics for each feature variable, including mean, variance, maximum, minimum, skewness, and kurtosis;

[0025] The calculation formula for skewness is:

[0026]

[0027] Among them, x i is the i-th data point, μ is the mean of the data, σ is the standard deviation of the data, and N is the number of features;

[0028] The calculation formula for kurtosis is:

[0029]

[0030] S23. Perform frequency domain analysis on the characteristics of each sensor data, convert the time domain data using Fourier transform to obtain frequency domain characteristics, and calculate the main frequency components and amplitudes in the frequency domain:

[0031]

[0032] Where X(f) is the complex value in the frequency domain, x(t) is the time domain data, f is the frequency, and T is the length of the time series;

[0033] S24, merging all extracted features to form a unified feature matrix;

[0034] S25. Perform feature selection on the unified feature matrix. Use a chi-square test-based method to select features that are closely related to equipment failures and reduce the impact of redundant features.

[0035]

[0036] Among them, O i is the observed value, E i is the expected value, χ 2 is the chi-square statistic, and N is the number of features.

[0037] Optionally, S3 includes the following steps:

[0038] S31, dividing the unified feature matrix into multiple time series samples, the time steps of the time series samples are of fixed length, and each time series sample contains multidimensional data of multiple sensors;

[0039] S32. Add timestamp information to each time step of each time series sample. The timestamp information indicates the relative time when the data is collected. The timestamp information is a discrete time index, expressed as t1, t2, ..., t n , where t i is the timestamp of the i-th time step, and n is the length of the time series;

[0040] S33. Apply position encoding to each time series sample to supplement the time order information of the time series. The position encoding is calculated based on the position of each time step:

[0041]

[0042] Where PE(t,2i) is the position encoding of the t-th time step in the 2i dimension, PE(t,2i+1) is the position encoding of the t-th time step in the 2i+1 dimension, d is the feature dimension, t is the time step, and i is the feature dimension index;

[0043] S34. Combine the sensor data of each time step with the corresponding timestamp and position code to form the input vector of each time step:

[0044] x(t)=[x1(t),x2(t),...,x m (t),PE(t,2i),PE(t,2i+1)];

[0045] Among them, x(t) is the input vector of the tth time step, x1(t),x2(t),...,x m (t) is the individual features of the sensor data, PE(t,2i) and PE(t,2i+1) are the position codes;

[0046] S35. Combine the input vectors of all time steps into a time series feature matrix X = [x1, x2, ..., x n ], where n is the number of time steps in the time series, x i is the input vector of the i-th time step.

[0047] Optionally, the S4 includes the following steps:

[0048] S41, the time series feature matrix X=[x1,x2,...,x n ] as input to the improved Transformer model;

[0049] S42. Apply multi-head attention in the self-attention mechanism to the input feature matrix and capture the dependencies between different time steps by calculating the similarity between the query vector, key vector, and value vector:

[0050]

[0051] Among them, Q is the query matrix, K is the key matrix, V is the value matrix, d k is the dimension of the key vector, softmax(·) is the softmax normalization function;

[0052] S43. Introducing a sparse mixed expert model into the multi-head attention mechanism improves the computational efficiency and expressiveness of the model by dynamically selecting part of the expert network for calculation:

[0053]

[0054] Among them, w i is the weight of the expert network, Expert i is the calculation output of the i-th expert network, h is the number of experts, and the most relevant experts are selected for processing;

[0055] S44. The spatial dependency relationship between devices is modeled through a graph convolutional network. The input of the graph convolutional network is the adjacency matrix A between devices and the sensor data matrix X. The graph convolution operation is performed using the following formula:

[0056] H ′ =σ(AHW);

[0057] Where A is the adjacency matrix between devices, H is the input feature matrix, W is the weight matrix of the graph convolutional network, and σ(·) is the activation function;

[0058] S45. Combine the output results of the multi-head self-attention and graph convolutional networks, and fuse the time series features and the spatial dependency information between devices through weighted summation:

[0059] H final =α·AttentionOutput+β·GCNOutput;

[0060] Among them, α and β are weighted coefficients, AttentionOutput represents the output result of multi-head self-attention, GCNOutput represents the output result of graph convolution network, H final is the final feature representation;

[0061] S46, the final feature representation H final Processing is performed through multiple feedforward neural network layers, which include two fully connected layers. The output formula of the first fully connected layer is:

[0062] FFN(x)=max(0,xW1+b1)W2+b2;

[0063] Where W1 and W2 are weight matrices, x is the input feature matrix, b1 and b2 are bias terms, and max(0, ) is the ReLU activation function.

[0064] S47. Process the features output by the feedforward network through the decoder to generate equipment fault prediction results:

[0065]

[0066] Among them, H final is the final feature representation, W out is the weight matrix of the output layer, b out is the bias term of the output layer, σ(·) is the Sigmoid activation function, is the predicted probability of failure;

[0067] S48. The loss function used in the model training process is:

[0068]

[0069] Where N is the number of samples, y i is the actual label, is the prediction result, L is the loss value;

[0070] S49. Use the back propagation algorithm and gradient descent method to minimize the loss function L, and build a fault prediction model through multiple rounds of training.

[0071] Optionally, the S5 includes the following steps:

[0072] S51. During the operation of the equipment, obtain real-time equipment data and convert it into a feature matrix X real-time =[x1,x2,...,x m ], where x i is the input data of the i-th time step, and m is the length of the real-time data;

[0073] S52, real-time data X real-time Perform preprocessing and obtain the processed feature matrix X processed , the processed real-time data X processed Input into the trained fault prediction model to obtain the prediction results

[0074] S53. Prediction results of model output Perform threshold determination, if the predicted value If the set threshold is exceeded, the device is judged to be at risk of failure:

[0075]

[0076] Among them, θ is the set fault probability threshold, if If the device is faulty, it is predicted that the device is faulty; otherwise, the device is considered normal.

[0077] S54. Based on the prediction results of the equipment And the judgment result Fault_Detected generates a real-time monitoring report.

[0078] The beneficial effects of the present invention are:

[0079] (1) The equipment fault prediction method of the present invention can significantly improve the accuracy and timeliness of equipment fault detection. Traditional methods generally rely on manual experience or fixed rules, often with the risk of missed and misjudgment. Our deep learning-based model can utilize a large amount of historical equipment data and real-time sensor data to extract features from multiple dimensions and perform comprehensive fault prediction. This method not only breaks through the limitations of a single data source, but also uses advanced technologies such as deep neural networks to deeply explore the implicit relationships between data, thereby improving the accuracy and reliability of fault prediction.

[0080] (2) This invention uses a dynamic threshold setting method. Compared with traditional fixed threshold prediction methods, it can more flexibly adapt to changes in equipment under different operating conditions. Through historical data analysis and failure rate calculation, this invention can dynamically adjust the fault warning threshold, thereby providing timely warnings even when there are minor changes in equipment status. This dynamic adjustment mechanism not only improves the sensitivity of fault warnings, but also effectively reduces the false alarm rate, providing more accurate prediction information for operation and maintenance personnel.

[0081] (3) The present invention further enhances the robustness of equipment fault prediction by combining multi-dimensional data fusion and anomaly detection models. By combining real-time data, historical fault data, and the relationships between equipment components, early detection of potential equipment faults can be achieved. The introduction of anomaly detection models can detect potential abnormal behaviors during equipment operation, which is of great significance for predicting complex faults and preventing systemic faults. With the comprehensive fault prediction information and anomaly detection results, operation and maintenance personnel can respond more quickly and take necessary maintenance measures to avoid further deterioration of equipment faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0083] Figure 1 This is a flowchart of a big data equipment fault prediction method based on machine learning proposed by the present invention;

[0084] Figure 2 This is a structural diagram of the fault prediction model in the big data equipment fault prediction method based on machine learning proposed in the present invention. DETAILED DESCRIPTION

[0085] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0086] refer to Figure 1-Figure 2 , a big data equipment fault prediction method based on machine learning, comprising the following steps:

[0087] S1. Collect sensor data, operation and maintenance logs, and historical fault records during equipment operation, and pre-process the collected equipment data;

[0088] S2. Extract multidimensional features from the preprocessed device data, calculate statistical characteristics, and obtain a unified feature matrix;

[0089] S3. Use the time series analysis method to process the unified feature matrix, add timestamp information and apply position coding to generate a time series feature matrix;

[0090] S4. Use the improved Transformer model based on the hybrid expert model and graph convolutional network to process the input time series feature matrix, extract long-term dependencies and local features, optimize model parameters through multiple rounds of training, and build a fault prediction model;

[0091] S5. Use the trained fault prediction model to monitor the equipment in real time, predict the probability of equipment failure and generate prediction results;

[0092] S6. Based on the prediction results and the set fault warning threshold, the alarm mechanism is triggered and fault warning information is sent to the operation and maintenance personnel;

[0093] S7. Continuously optimize and update the fault prediction model based on real-time monitoring and early warning feedback.

[0094] In this embodiment, S1 includes the following steps:

[0095] S11. Collect operating data from the device's sensors, including temperature, current, voltage, vibration, and pressure characteristics, and obtain the device's operation and maintenance logs and historical fault records;

[0096] S12. Clean the collected equipment operation data to remove missing values, outliers, and duplicate data. Missing values are filled using linear interpolation, and outliers are identified and removed using an outlier detection method based on a box plot.

[0097] S13. Normalize the cleaned data:

[0098]

[0099] Where x is the original data, x' is the normalized data, min(x) is the minimum value in the data, and max(x) is the maximum value in the data.

[0100] In this embodiment, S2 includes the following steps:

[0101] S21. Extracting multiple characteristic variables from the equipment operation data after cleaning, including temperature, current, voltage, vibration, and pressure;

[0102] S22. Calculate statistical characteristics for each feature variable, including mean, variance, maximum, minimum, skewness, and kurtosis;

[0103] The calculation formula for skewness is:

[0104]

[0105] Among them, x i is the i-th data point, μ is the mean of the data, σ is the standard deviation of the data, and N is the number of features;

[0106] The calculation formula for kurtosis is:

[0107]

[0108] S23. Perform frequency domain analysis on the characteristics of each sensor data, convert the time domain data using Fourier transform to obtain frequency domain characteristics, and calculate the main frequency components and amplitudes in the frequency domain:

[0109]

[0110] Where X(f) is the complex value in the frequency domain, x(t) is the time domain data, f is the frequency, and T is the length of the time series;

[0111] S24, merging all extracted features to form a unified feature matrix;

[0112] S25. Perform feature selection on the unified feature matrix. Use a chi-square test-based method to select features that are closely related to equipment failures and reduce the impact of redundant features.

[0113]

[0114] Among them, O i is the observed value, E i is the expected value, χ 2 is the chi-square statistic, and N is the number of features.

[0115] In this embodiment, S3 includes the following steps:

[0116] S31, dividing the unified feature matrix into multiple time series samples, the time steps of the time series samples are of fixed length, and each time series sample contains multidimensional data of multiple sensors;

[0117] S32. Add timestamp information to each time step of each time series sample. The timestamp information indicates the relative time when the data is collected. The timestamp information is a discrete time index, expressed as t1, t2, ..., t n , where t i is the timestamp of the i-th time step, and n is the length of the time series;

[0118] S33. Apply position encoding to each time series sample to supplement the time order information of the time series. The position encoding is calculated based on the position of each time step:

[0119]

[0120] Where PE(t,2i) is the position encoding of the t-th time step in the 2i dimension, PE(t,2i+1) is the position encoding of the t-th time step in the 2i+1 dimension, d is the feature dimension, t is the time step, and i is the feature dimension index;

[0121] S34. Combine the sensor data of each time step with the corresponding timestamp and position code to form the input vector of each time step:

[0122] x(t)=[x1(t),x2(t),...,x m (t),PE(t,2i),PE(t,2i+1)];

[0123] Among them, x(t) is the input vector of the tth time step, x1(t),x2(t),...,x m (t) is the individual features of the sensor data, PE(t,2i) and PE(t,2i+1) are the position codes;

[0124] S35. Combine the input vectors of all time steps into a time series feature matrix X = [x1, x2, ..., x n ], where n is the number of time steps in the time series, x i is the input vector of the i-th time step.

[0125] In this embodiment, S4 includes the following steps:

[0126] S41, the time series feature matrix X=[x1,x2,...,x n ] as input to the improved Transformer model;

[0127] S42. Apply multi-head attention in the self-attention mechanism to the input feature matrix and capture the dependencies between different time steps by calculating the similarity between the query vector, key vector, and value vector:

[0128]

[0129] Among them, Q is the query matrix, K is the key matrix, V is the value matrix, d k is the dimension of the key vector, softmax(·) is the softmax normalization function;

[0130] S43. Introducing a sparse mixed expert model into the multi-head attention mechanism improves the computational efficiency and expressiveness of the model by dynamically selecting part of the expert network for calculation:

[0131]

[0132] Among them, w i is the weight of the expert network, Expert i is the calculation output of the i-th expert network, h is the number of experts, and the most relevant experts are selected for processing;

[0133] S44. The spatial dependency relationship between devices is modeled through a graph convolutional network. The input of the graph convolutional network is the adjacency matrix A between devices and the sensor data matrix X. The graph convolution operation is performed using the following formula:

[0134] H ′ =σ(AHW);

[0135] Where A is the adjacency matrix between devices, H is the input feature matrix, W is the weight matrix of the graph convolutional network, and σ(·) is the activation function;

[0136] S45. Combine the output results of the multi-head self-attention and graph convolutional networks, and fuse the time series features and the spatial dependency information between devices through weighted summation:

[0137] H final =α·AttentionOutput+β·GCNOutput;

[0138] Among them, α and β are weighted coefficients, AttentionOutput represents the output result of multi-head self-attention, GCNOutput represents the output result of graph convolution network, H final is the final feature representation;

[0139] S46, the final feature representation H final Processing is performed through multiple feedforward neural network layers, which include two fully connected layers. The output formula of the first fully connected layer is:

[0140] FFN(x)=max(0,xW1+b1)W2+b2;

[0141] Where W1 and W2 are weight matrices, x is the input feature matrix, b1 and b2 are bias terms, and max(0, ) is the ReLU activation function.

[0142] S47. Process the features output by the feedforward network through the decoder to generate equipment fault prediction results:

[0143]

[0144] Among them, H final is the final feature representation, W out is the weight matrix of the output layer, b outis the bias term of the output layer, σ(·) is the Sigmoid activation function, is the predicted probability of failure;

[0145] S48. The loss function used in the model training process is:

[0146]

[0147] Where N is the number of samples, y i is the actual label, is the prediction result, L is the loss value;

[0148] S49. Use the back propagation algorithm and gradient descent method to minimize the loss function L, and build a fault prediction model through multiple rounds of training.

[0149] In this embodiment, S5 includes the following steps:

[0150] S51. During the operation of the equipment, obtain real-time equipment data and convert it into a feature matrix X real-time =[x1,x2,...,x m ], where x i is the input data of the i-th time step, and m is the length of the real-time data;

[0151] S52, real-time data X real-time Perform preprocessing and obtain the processed feature matrix X processed , the processed real-time data X processed Input into the trained fault prediction model to obtain the prediction results

[0152] S53. Prediction results of model output Perform threshold determination, if the predicted value If the set threshold is exceeded, the device is judged to be at risk of failure:

[0153]

[0154] Among them, θ is the set fault probability threshold, if If the device is faulty, it is predicted that the device is faulty; otherwise, the device is considered normal.

[0155] S54. Based on the prediction results of the equipment And the judgment result Fault_Detected generates a real-time monitoring report.

[0156] Example:

[0157] This embodiment is applied to the industrial production line of a large manufacturing enterprise. The production line includes multiple high-precision equipment, such as CNC machine tools, automated assembly robots, and industrial conveyor belts. These devices are responsible for completing product cutting, assembly, welding, and other process operations, and are core links in the entire production process. Due to the long operating hours and heavy workloads of these equipment, equipment failures often cause production line downtime, resulting in production delays and significant repair costs. To address this problem, the enterprise decided to introduce a machine learning-based equipment failure prediction system, employing the fault prediction method based on deep learning and data fusion proposed in this invention.

[0158] In this production line, equipment is usually equipped with multiple sensors, including temperature sensors, vibration sensors, current sensors, etc. These sensors monitor the operating status of the equipment in real time and transmit data to the central control system. In the past, companies mainly relied on experience and regular maintenance to maintain equipment. However, many faults gradually appear during the operation of the equipment. Traditional regular maintenance methods cannot detect these hidden faults in time, resulting in equipment failures often being in a serious state when they occur.

[0159] The equipment failure prediction method proposed in this invention integrates historical failure data, real-time monitoring data and equipment operation status data to construct a deep learning-based fault prediction model. In actual application, the implementer first collects various sensor data of the equipment in real time, and combines it with historical equipment failure data to perform data preprocessing and feature extraction. Through feature selection and feature engineering technology, the implementer extracts features closely related to equipment failures from multiple dimensions, including equipment temperature, vibration, power consumption, current and other parameters, and conducts a comprehensive analysis in combination with the equipment's operating load and environmental factors.

[0160] After processing, the data is input into a fault prediction model based on a deep neural network for training. The implementers use a model architecture that combines long short-term memory networks and graph convolutional neural networks to capture the timing characteristics of the equipment's operating status and the mutual influence relationship between various devices. By learning the failure modes in historical data, the model can accurately predict the time and type of possible equipment failure.

[0161] Once model training is complete and reaches the preset accuracy, the system enters the real-time monitoring phase. Real-time monitoring data from the equipment is fed into the model, and predictions are generated in real time, sending fault warnings to maintenance personnel. These predictions include not only the probability of a fault occurring, but also the specific fault type and probability range, helping them determine the severity and priority of the fault. If the prediction exceeds the preset fault warning threshold, the system automatically triggers an alarm, and maintenance personnel receive an immediate warning, enabling them to take timely action to prevent the fault.

[0162] By comparing with the traditional method, the specific data are as follows Table 1:

[0163] Table 1 Comparison of key performances of the present invention and traditional methods in the fault detection process

[0164]

[0165] Throughout the embodiment, implementers can monitor multi-dimensional data of a device in real time and obtain fault detection results through the method of the present invention.

[0166] By incorporating deep learning models and combining the strengths of long-short-term memory networks and graph convolutional networks, this method accurately captures the timing characteristics of equipment operation and the interactions between devices, enabling real-time prediction of the probability and type of equipment failures. Compared to traditional fault prediction methods based on manual inspections and fixed-cycle maintenance, this method can proactively identify potential equipment failures by monitoring multi-dimensional data such as temperature, vibration, and current in real time, providing more accurate and timely early warning information.

[0167] This invention dynamically adjusts fault warning thresholds, optimizing them based on the device's actual operating status and historical data. This makes fault prediction more flexible and adaptable, effectively avoiding misjudgments or missed detections caused by fixed threshold settings, particularly in high-load or complex operating environments. Through an online update mechanism, the system dynamically adjusts the model based on real-time device data, ensuring continued accuracy and adaptability in predictions.

[0168] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A big data equipment fault prediction method based on machine learning, characterized in that: The steps include: S1. Collect data during the operation of the equipment and pre-process the collected equipment data; S2. Extract multidimensional features from the preprocessed device data, calculate statistical characteristics, and obtain a unified feature matrix; S3. Use the time series analysis method to process the unified feature matrix, add timestamp information and apply position coding to generate a time series feature matrix; S4. Use the improved Transformer model based on the hybrid expert model and graph convolutional network to process the input time series feature matrix, extract long-term dependencies and local features, optimize model parameters through multiple rounds of training, and build a fault prediction model; S5. Use the trained fault prediction model to monitor the equipment in real time, predict the probability of equipment failure and generate prediction results; S6. Based on the prediction results and the set fault warning threshold, the alarm mechanism is triggered and fault warning information is sent to the operation and maintenance personnel; S7. Continuously optimize and update the fault prediction model based on real-time monitoring and early warning feedback.

2. The method for predicting equipment failure based on big data and machine learning according to claim 1, characterized in that: Said S1 comprises the following steps: S11. Collect operating data from the device's sensors, including temperature, current, voltage, vibration, and pressure characteristics, and obtain the device's operation and maintenance logs and historical fault records; S12. Clean the collected equipment operation data to remove missing values, outliers, and duplicate data. Missing values are filled using linear interpolation, and outliers are identified and removed using an outlier detection method based on a box plot. S13. Normalize the cleaned data: Where x is the original data, x' is the normalized data, min(x) is the minimum value in the data, and max(x) is the maximum value in the data.

3. The big data equipment fault prediction method based on machine learning according to claim 1 is characterized in that: The S2 comprises the following steps: S21. Extracting multiple characteristic variables from the equipment operation data after cleaning, including temperature, current, voltage, vibration, and pressure; S22. Calculate statistical characteristics for each feature variable, including mean, variance, maximum value, minimum value, skewness, and kurtosis; The calculation formula for skewness is: Among them, x i is the i-th data point, μ is the mean of the data, σ is the standard deviation of the data, and N is the number of features; The calculation formula for kurtosis is: S23. Perform frequency domain analysis on the characteristics of each sensor data, convert the time domain data using Fourier transform to obtain frequency domain characteristics, and calculate the main frequency components and amplitudes in the frequency domain: Where X(f) is the complex value in the frequency domain, x(t) is the time domain data, f is the frequency, and T is the length of the time series; S24, merging all extracted features to form a unified feature matrix; S25. Perform feature selection on the unified feature matrix, using a chi-square test-based method to select features that are closely related to equipment failures and reduce the impact of redundant features: Among them, O i is the observed value, E i is the expected value, χ 2 is the chi-square statistic, and N is the number of features.

4. The method for predicting equipment failure based on big data and machine learning according to claim 1, characterized in that: The S3 includes the following steps: S31, dividing the unified feature matrix into multiple time series samples, the time steps of the time series samples are of fixed length, and each time series sample contains multidimensional data of multiple sensors; S32. Add timestamp information to each time step of each time series sample. The timestamp information indicates the relative time when the data is collected. The timestamp information is a discrete time index, expressed as t1, t2, ..., t n , where t i is the timestamp of the i-th time step, and n is the length of the time series; S33. Apply position encoding to each time series sample to supplement the time order information of the time series. The position encoding is calculated based on the position of each time step: Where PE(t,2i) is the position encoding of the t-th time step in the 2i dimension, PE(t,2i+1) is the position encoding of the t-th time step in the 2i+1 dimension, d is the feature dimension, t is the time step, and i is the feature dimension index; S34. Combine the sensor data of each time step with the corresponding timestamp and position code to form the input vector of each time step: x(t)=[x1(t),x2(t),...,x m (t),PE(t,2i),PE(t,2i+1)]; Among them, x(t) is the input vector of the tth time step, x1(t),x2(t),...,x m (t) is the individual features of the sensor data, PE(t,2i) and PE(t,2i+1) are the position codes; S35. Combine the input vectors of all time steps into a time series feature matrix X = x1, x2, ..., x n ], where n is the number of time steps in the time series, x i is the input vector for the i-th time step.

5. The method for predicting equipment failure based on big data and machine learning according to claim 4, characterized in that: The S4 comprises the following steps: S41, the time series feature matrix X=[x1,x2,...,x n ] as input to the improved Transformer model; S42. Apply the multi-head attention in the self-attention mechanism to the input feature matrix and capture the dependencies between different time steps by calculating the similarity between the query vector, key vector, and value vector: Among them, Q is the query matrix, K is the key matrix, V is the value matrix, d k is the dimension of the key vector, softmax(·) is the softmax normalization function; S43. Introducing a sparse mixed expert model into the multi-head attention mechanism improves the computational efficiency and expressiveness of the model by dynamically selecting part of the expert network for calculation: Among them, w i is the weight of the expert network, Expert i is the calculation output of the i-th expert network, h is the number of experts, and the most relevant experts are selected for processing; S44. The spatial dependency relationship between devices is modeled through a graph convolutional network. The input of the graph convolutional network is the adjacency matrix A between devices and the sensor data matrix X. The graph convolution operation is performed using the following formula: H ′ =σ(AHW); Where A is the adjacency matrix between devices, H is the input feature matrix, W is the weight matrix of the graph convolutional network, and σ(·) is the activation function; S45. Combine the output results of the multi-head self-attention and graph convolutional networks, and fuse the time series features and the spatial dependency information between devices through weighted summation: H final =α·AttentionOutput+β·GCNOutput; Among them, α and β are weighted coefficients, AttentionOutput represents the output result of multi-head self-attention, GCNOutput represents the output result of graph convolution network, H final is the final feature representation; S46, the final feature representation H final Processing is performed through multiple feedforward neural network layers, which include two fully connected layers. The output formula of the first fully connected layer is: FFN(x)=max(0,xW1+b1)W2+b2; Where W1 and W2 are weight matrices, x is the input feature matrix, b1 and b2 are bias terms, and max(0, ) is the ReLU activation function. S47. Process the features output by the feedforward network through the decoder to generate equipment fault prediction results: Among them, H final is the final feature representation, W out is the weight matrix of the output layer, b out is the bias term of the output layer, σ(·) is the Sigmoid activation function, is the predicted probability of failure; S48. The loss function used in the model training process is: Where N is the number of samples, y i is the actual label, is the prediction result, L is the loss value; S49. Use the back propagation algorithm and gradient descent method to minimize the loss function L, and build a fault prediction model through multiple rounds of training.

6. The method for predicting equipment failure based on big data and machine learning according to claim 1, characterized in that: The S5 comprises the following steps: S51. During the operation of the equipment, obtain real-time equipment data and convert it into a feature matrix X real-time =[x1,x2,...,x m ], where x i is the input data of the i-th time step, and m is the length of the real-time data; S52, real-time data X real-time Perform preprocessing and obtain the processed feature matrix X processed , the processed real-time data X processed Input into the trained fault prediction model to obtain the prediction results S53. Prediction results of model output Perform threshold determination, if the predicted value If the set threshold is exceeded, the device is judged to be at risk of failure: Among them, θ is the set fault probability threshold, if If the device is faulty, it is predicted that the device is faulty; otherwise, the device is considered normal. S54. Based on the prediction results of the equipment and the judgment result Fault_Detected, and generates a real-time monitoring report.

Images