Device fault prediction method and apparatus, and readable storage medium and electronic device

Abstract

Provided in the present disclosure are a device fault prediction method and apparatus, and a readable storage medium and an electronic device. After feature signals of a device are extracted, by selecting different feature extraction methods on the basis of feature types in the feature signals, and extracting different features in the feature signals on the basis of the different feature extraction methods, feature extraction methods adapted to fault features can be used to extract different fault features to realize secondary processing of the fault features, thereby extracting training samples with highly concentrated features; and then, by inputting the training samples into a machine learning classification algorithm for data mining, a fault prediction model with an accurate input and output relationship is established, so that the aim of accurately predicting a device fault is achieved by means of the established fault prediction model in combination with a device feature signal to be predicted.

Description

Equipment failure prediction method, device, readable storage medium and electronic device

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This disclosure claims priority to Chinese patent application number 202311282404X filed with the Chinese Patent Office on October 7, 2023, entitled “A method, device, readable storage medium and electronic device for predicting equipment failure,” the entire contents of which are incorporated by reference into this disclosure. Technical Field

[0003] The present disclosure relates to the technical field of equipment failure prediction, and in particular to an equipment failure prediction method, apparatus, readable storage medium, and electronic device. Background Art

[0004] With the rapid development of the artificial intelligence industry, the demand for semiconductors is also increasing. Semiconductor production requires the coordination of various devices. This includes the suction equipment used to extract chips from the jig. During the production process, various types of failures are inevitable due to human or environmental factors. Typical failure types include nozzle wear, nozzle blockage, loose suction rod, and vacuum generator damage. These failures can disrupt the process and even disrupt the entire production chain. Therefore, the early prediction and classification of failures are crucial for industrial safety.

[0005] Currently, existing fault prediction technologies are usually unable to achieve accurate fault identification and prediction effects.

[0006] Summary of the Invention

[0007] The purpose of the embodiments of the present disclosure is to provide a device fault prediction method, apparatus, readable storage medium and electronic device to improve the accuracy of device fault prediction.

[0008] The present disclosure provides a method for predicting equipment failures, including the following steps:

[0009] Extracting a preset number of time nodes and extracting characteristic signals corresponding to the devices at the time nodes;

[0010] Selecting at least one feature extraction method according to the feature type in the feature signal;

[0011] Extracting features from the feature signal using at least one of the feature extraction methods to obtain a training sample;

[0012] Inputting the training samples into a machine learning classification model for training to obtain a fault prediction model;

[0013] Acquire a characteristic signal of the device to be predicted, input the characteristic signal of the device to be predicted into the fault prediction model, and obtain a fault prediction result.

[0014] The present disclosure also provides a device for predicting equipment failures, including:

[0015] An extraction module is used to extract a preset number of time nodes and extract characteristic signals corresponding to the time node devices;

[0016] A selection module, configured to select at least one feature extraction method according to the feature type in the feature signal;

[0017] an extraction module, configured to extract features from the feature signal by using at least one of the feature extraction methods to obtain a training sample;

[0018] A training module, inputting the training samples into a machine learning classification model for training to obtain a fault prediction model;

[0019] The prediction module is used to obtain the characteristic signal of the device to be predicted, input the characteristic signal of the device to be predicted into the fault prediction model, and obtain a fault prediction result.

[0020] The embodiment of the present disclosure further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the above-mentioned method for predicting equipment failure are implemented.

[0021] An embodiment of the present disclosure also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, each step in the above-mentioned method for predicting equipment failure is implemented.

[0022] The beneficial effects of the present disclosure are: after extracting the characteristic signal of the equipment, different feature extraction methods are selected according to the feature type in the characteristic signal, and different features in the characteristic signal are extracted based on different feature extraction methods, that is, different fault features can be extracted using feature extraction methods suitable for the fault features, thereby realizing secondary processing of the fault features, thereby extracting training samples with a high feature concentration, and then by inputting the training samples into the machine learning classification algorithm for data mining, a fault prediction model with a precise input and output relationship is established, thereby achieving the purpose of accurately predicting equipment faults by combining the established fault prediction model with the characteristic signal of the equipment to be predicted. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] FIG1 is a flowchart of a method for predicting equipment failures according to an embodiment of the present disclosure;

[0024] FIG2 is a flowchart of another step of a method for predicting equipment failure in an embodiment of the present disclosure;

[0025] FIG3 is a deviation diagram of characteristic values ​​corresponding to different K values ​​in a method for predicting equipment failure in an embodiment of the present disclosure;

[0026] FIG4 is a schematic diagram of a frequency band hierarchical structure in a method for predicting equipment failures according to an embodiment of the present disclosure;

[0027] FIG5 is a diagram showing the energy proportions of each frequency band of the bottom node in a device fault prediction method according to an embodiment of the present disclosure;

[0028] FIG6 is a schematic diagram of a recursive tree structure in a method for predicting equipment failures according to an embodiment of the present disclosure;

[0029] FIG7 is a diagram showing a linear function calculation result of a fault analysis in a method for predicting equipment faults according to an embodiment of the present disclosure;

[0030] FIG8 is a schematic structural diagram of an apparatus for predicting equipment failure according to an embodiment of the present disclosure;

[0031] FIG9 is a schematic structural diagram of an electronic device according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0032] To illustrate the technical content, achieved objectives and effects of the present disclosure in detail, the following description is given in conjunction with the embodiments and accompanying drawings.

[0033] A method for predicting equipment failures, comprising the steps of:

[0034] Extracting a preset number of time nodes and extracting characteristic signals corresponding to the time node devices;

[0035] Extracting different features from the feature signal using different feature extraction methods to obtain training samples;

[0036] Inputting the training samples into a machine learning classification model for training to obtain feature values ​​and a mapping relationship between the feature values ​​and faults;

[0037] Fitting the characteristic values ​​and the mapping relationship between the characteristic values ​​and the faults according to a regression algorithm to obtain a fault prediction model;

[0038] Acquire a characteristic signal of the device to be predicted, input the characteristic signal of the device to be predicted into the fault prediction model, and obtain a fault prediction result.

[0039] From the above description, it can be seen that the beneficial effect of the present disclosure is that after extracting the characteristic signal of the equipment, different feature extraction methods are selected according to the feature type in the characteristic signal, and different features in the characteristic signal are extracted based on different feature extraction methods, that is, different fault features can be extracted using feature extraction methods suitable for the fault features, thereby realizing secondary processing of the fault features, thereby extracting training samples with a high feature concentration, and then by inputting the training samples into the machine learning classification algorithm for data mining, a fault prediction model with a precise input and output relationship is established, thereby achieving the purpose of accurately predicting equipment faults by combining the established fault prediction model with the characteristic signal of the equipment to be predicted.

[0040] Optionally, extracting features from the feature signal by using at least one feature extraction method to obtain a training sample includes:

[0041] Extracting time domain features from the characteristic signal by principal component analysis to obtain a time domain training set, and extracting frequency domain features from the characteristic signal by wavelet packet transform to obtain a frequency domain training set;

[0042] The training samples are obtained according to the time domain training set and the frequency domain training set.

[0043] From the above description, it can be seen that the principal component analysis method is used for time domain features to reduce the dimensionality of a large amount of time domain feature data into a suitable dimension size, which reduces the computational difficulty and ensures the validity of the data; and the wavelet packet transformation method is used for frequency features, which takes advantage of its local frequency feature analysis to achieve the simultaneous decomposition of low-frequency and high-frequency signals, and can perform better time-frequency localization analysis, so that training samples with a high feature concentration can be extracted based on the principal component analysis method and the wavelet packet transformation method.

[0044] From the above description, it can be seen that by adopting a method of extracting a preset number of features from the time domain training set and the frequency domain training set with replacement, the total number of features in the time domain training set and the frequency domain training set after extraction remains unchanged, so that multiple extractions can be performed and model training can be performed based on the multiple extracted training samples, and then the best prediction model can be selected from the multiple training results to improve the prediction effect of equipment failure.

[0045] Optionally, extracting the time domain features from the feature signal by using a principal component analysis method to obtain a time domain training set includes:

[0046] Combining the original data in the time domain features into an original matrix with preset row and column sizes;

[0047] Performing zero-mean transformation on the original matrix to obtain a zero-mean transformation matrix;

[0048] Calculating the covariance matrix corresponding to the zero-mean matrix;

[0049] Calculating the eigenvalues ​​of the covariance matrix and the eigenvectors corresponding to the eigenvalues;

[0050] According to the size of the eigenvalue, a preset number of eigenvectors are obtained to form a eigenmatrix in rows;

[0051] Performing dimensionality reduction on the original matrix according to the characteristic matrix to obtain a dimensionality reduction matrix;

[0052] The dimension-reduced matrix is ​​used as the time-domain training set.

[0053] From the above description, it can be seen that after the original data in the time domain features are organized into an original matrix of preset row and column sizes and zero-meaned, the covariance matrix corresponding to the zero-meaned matrix, the eigenvalues ​​of the covariance matrix, and the eigenvectors corresponding to the eigenvalues ​​are calculated in sequence, and then the preset eigenvectors are obtained according to the size of the eigenvalues ​​to form a feature matrix by rows, thereby realizing the screening of features and redundant processing, ensuring that the main content of the data will not be lost, while reducing the amount of original data and improving the efficiency of feature extraction.

[0054] Optionally, extracting frequency domain features from the characteristic signal by using a wavelet packet transform method to obtain a frequency domain training set includes:

[0055] Obtain data sampling frequency, data type and number of wavelet packet transform processing layers;

[0056] Processing the data sampling frequency according to the number of wavelet packet change method processing layers to obtain a preset number of sub-frequency segments;

[0057] Extracting energy features corresponding to each of the sub-frequency segments, and obtaining energy proportions corresponding to each of the sub-frequency segments;

[0058] The sub-frequency segment combination with the highest energy proportion is selected according to the data type as the frequency domain training set.

[0059] From the above description, it can be seen that the sampling frequency is divided into multiple sub-frequency segments by processing the data sampling frequency with a preset number of layers of wavelet packet transform, and the energy characteristics of each sub-frequency segment are extracted respectively to obtain the energy proportion of each sub-frequency segment. Then, the sub-frequency segment with the highest energy proportion is selected according to the data type to form a frequency domain training set, so that the sub-frequency segment most relevant to the fault can be selected, and training samples with a high feature concentration can be obtained, thereby improving the model training effect and improving the prediction accuracy.

[0060] Optionally, obtaining the training sample according to the time domain training set and the frequency domain training set includes:

[0061] A preset number of features are extracted from the time domain training set and the frequency domain training set respectively with replacement, and the features are combined to obtain training samples.

[0062] Optionally, inputting the training samples into a machine learning classification model for training to obtain a fault prediction model includes:

[0063] Dividing the training samples into a training set and a validation set;

[0064] Inputting the training set into the machine learning classification model for training to obtain a regression model for predicting faults;

[0065] The regression model is verified according to the verification set to obtain the fault prediction model.

[0066] Optionally, the machine learning classification model includes a two-classification model and a sixteen-classification model.

[0067] Optionally, the verifying the regression model according to the verification set to obtain the fault prediction model includes: inputting the verification set into a pre-constructed verification binary classification model and a verification sixteen classification model for training to obtain a verification regression model;

[0068] Validating the regression model according to the validation regression model;

[0069] If the regression model passes the verification, the regression model is determined as the fault prediction model.

[0070] Optionally, the method further includes:

[0071] If the regression model fails to pass the verification, the regression model is iteratively trained until the regression model passes the verification. Optionally, the inputting the training samples into the machine learning classification model for training to obtain the fault prediction model includes:

[0072] Performing feature classification on the training samples using a decision tree algorithm, a support vector machine algorithm, or a random forest algorithm to obtain feature values ​​and a mapping relationship between the feature values ​​and the faults;

[0073] The characteristic values ​​and the mapping relationship between the characteristic values ​​and the faults are fitted according to a regression algorithm to obtain a fault prediction model.

[0074] From the above description, it can be seen that the decision tree algorithm, support vector machine algorithm or random forest algorithm is used to perform feature classification on the training samples, providing classification models with different effects such as the shortest running time and the highest prediction accuracy, thereby meeting different needs in different application scenarios and improving the practicality of the method.

[0075] Optionally, fitting the characteristic values ​​and the mapping relationship between the characteristic values ​​and the faults according to a regression algorithm to obtain a fault prediction model includes:

[0076] Establishing a fault prediction function, taking the characteristic value as the function independent variable;

[0077] Calculating the mapping relationship between the characteristic value and the fault by the least square method or the gradient descent method to obtain the independent variable coefficient;

[0078] Regression fitting is performed based on the independent variables and the independent variable coefficients to obtain the fault prediction model.

[0079] From the above description, it can be seen that after obtaining the characteristic values ​​and the mapping relationship between various faults of the equipment and the characteristic values, the mapping relationship between the characteristic values ​​and the faults is calculated by the least squares method or the gradient descent method to obtain the independent variable coefficient, so as to fit the functional relationship between the faults and the characteristics, and achieve the effect of accurately predicting the equipment failure based on the input equipment physical quantity data.

[0080] Optionally, after fitting the characteristic values ​​and the mapping relationship between the characteristic values ​​and the faults according to the regression algorithm to obtain a fault prediction model, the method further includes:

[0081] Testing the prediction accuracy of each of the fault prediction models;

[0082] The step of inputting the characteristic signal of the device to be predicted into the fault prediction model comprises:

[0083] The device characteristic signal to be predicted is input into the fault prediction model with the highest prediction accuracy.

[0084] Another embodiment of the present disclosure provides a device for predicting equipment failure, comprising:

[0085] An extraction module is used to extract a preset number of time nodes and extract characteristic signals corresponding to the time node devices;

[0086] A selection module, configured to select at least one feature extraction method according to the feature type in the feature signal;

[0087] an extraction module, configured to extract features from the feature signal by using at least one of the feature extraction methods to obtain a training sample;

[0088] A training module, inputting the training samples into a machine learning classification model for training to obtain a fault prediction model;

[0089] The prediction module is used to obtain the characteristic signal of the device to be predicted, input the characteristic signal of the device to be predicted into the fault prediction model, and obtain a fault prediction result.

[0090] Another embodiment of the present disclosure provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the computer program implements the steps of the above-mentioned method for predicting equipment failure.

[0091] Another embodiment of the present disclosure provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, each step in the above-mentioned device fault prediction method is implemented.

[0092] The device fault prediction method, apparatus, computer-readable storage medium, and electronic device disclosed herein can be applied to fault prediction of various types of equipment, such as fault prediction of chip extraction equipment. Specific implementations are described below.

[0093] FIG1 is a flowchart of the steps of a device failure prediction method in an embodiment of the present disclosure. Please refer to FIG1 and FIG2 . A device failure prediction method includes the following steps.

[0094] S1. Extract a preset number of time nodes and extract characteristic signals corresponding to the devices at the time nodes.

[0095] Optionally, several time nodes can be randomly selected during the equipment operation and after the equipment is repaired, and characteristic signals of the loading and unloading machines and identification of whether there is a fault and the type of fault corresponding to the equipment in the corresponding time period can be extracted.

[0096] Optionally, the extraction ratio of the feature signals can be specifically adjusted and configured according to the machine learning accuracy requirements and the operation time requirements, for example, the number of feature signals can be 10% of the sampling time node.

[0097] S2. Select at least one feature extraction method according to the feature type in the feature signal.

[0098] Optionally, the characteristic signal may include but is not limited to parameters such as equipment temperature, pressure, vacuum flow of the vacuum generator, air consumption, pressure at the suction port, adsorption response time, current and voltage of the motor, and the like.

[0099] In an optional embodiment, for example, for characteristic signals having characteristics such as the mean, variance, skewness, peak value, and peak-to-peak value of the amplitudes of four sets of data including the suction port pressure, air consumption, friction at the suction rod, and vacuum flow, a time domain feature extraction method may be used; for example, for characteristic signals having characteristics such as the power in different frequency ranges of the four sets of data including the suction port pressure, air consumption, friction at the suction rod, and vacuum flow, such as the power in the low frequency range of 10-20 Hz, the power in the medium frequency range of 40-60 Hz, the power in the high frequency range of >100 Hz, and the frequency of the spectral peak state peak, a frequency domain feature extraction method may be used;

[0100] S3. Extract features from the feature signal using at least one of the feature extraction methods to obtain a training sample.

[0101] In an optional embodiment, if it is necessary to extract time domain features and frequency domain features at the same time, the time domain features in the characteristic signal can be extracted by principal component analysis to obtain a time domain training set, thereby obtaining four groups of data including suction port pressure, air consumption, friction at the suction rod and vacuum flow, and each group of data corresponds to 14 characteristics, totaling 56 groups of characteristics; and the frequency domain features in the characteristic signal can be extracted by wavelet packet transformation method to obtain a frequency domain training set, thereby obtaining four groups of data including suction port pressure, air consumption, friction at the suction rod and vacuum flow, and each group of data corresponds to 4 characteristics, totaling 16 groups of characteristics.

[0102] S4. Input the training samples into a machine learning classification model for training to obtain a fault prediction model.

[0103] Optionally, the training samples can be divided into a training set and a validation set. The training set can be used as the learning set for machine learning and input into the model for training, while the validation set can be used to verify the model after training. In one example, if 2000 sets of data are obtained as training samples, 1600 sets of data can be used as the training set and 400 sets of data can be used as the validation set.

[0104] Optionally, the training set can be input into a machine learning classification model for training to obtain a regression model for predicting faults, and then the regression model is verified based on the validation set to obtain a fault prediction model.

[0105] Optionally, if fault detection is required for the suction equipment, the fault prediction model may be a suction equipment fault classification model; wherein the suction equipment generally includes four types of faults: nozzle wear, nozzle blockage, loose suction rod, and vacuum generator damage.

[0106] Alternatively, if fault detection is performed on other equipment, the fault type can be determined by investigating key parts or parts prone to faults of the equipment.

[0107] In one example, taking the fault type of the absorption device as an example, the fault is segmented by threshold value to determine the fault judgment condition as follows:

[0108] If the pressure at the suction port of the nozzle is less than 0.08Mpa, it is considered that the nozzle is worn and needs to be replaced;

[0109] If the suction port pressure is 0 and the air consumption is 0, the nozzle is considered blocked and needs maintenance.

[0110] If the friction force at the suction rod is less than 0.02N, the suction rod is considered loose;

[0111] If the vacuum flow rate is less than 1e^-4m^3 / s, the vacuum generator is considered damaged and needs to be replaced or repaired. The thresholds of various parameters can be set according to the specific manifestations of different equipment failures.

[0112] In this example, the characteristic signal can be identified as a fault according to the above fault judgment conditions. First, there are two cases, namely, the presence of a fault and the absence of a fault, which is a binary classification model. In this model, the label is set to 0 to indicate no fault, and the label is set to 1 to indicate a fault. Through cross-validation of the identification and features, it can be determined whether the system has defects.

[0113] Secondly, in order to complete the judgment of the specific fault type, it is necessary to set flags for different situations; the four faults occurring alone are set as 2, 3, 4, and 5 respectively; the two faults occurring simultaneously are set as 6, 7, 8, 9, 10, and 11; the three faults occurring simultaneously are set as 12, 13, 14, and 15; and the four faults occurring simultaneously are set as 16. The sixteen-classification model is obtained, as shown in the following table.

[0114] Table 1. Fault signal identification

[0115] Optionally, the training set in the training sample can be input into the constructed binary classification model and the sixteen-classification model for training to obtain a regression model for predicting faults. At the same time, the validation set in the training sample can be input into the previously constructed validation binary classification model and the validation sixteen-classification model for training to obtain a validation regression model for verifying the above regression model. The regression model is verified by the validation regression model, and the regression model that passes the verification is determined as the final fault prediction model obtained by training. S5. Obtain the characteristic signal of the device to be predicted, input the characteristic signal of the device to be predicted into the fault prediction model, and obtain a fault prediction result.

[0116] Optionally, for time domain features, when the number of training sets is large, for example, when 1,600 groups of data are input as training sets, the dimension of the data learning set is large, resulting in reduced accuracy and efficiency of machine learning. Therefore, feature selection is required to eliminate or process redundant data, and use principal component analysis to process time domain features.

[0117] In some embodiments, the time domain features may be processed through the following steps.

[0118] T1. The original data in the time domain features are organized into an original matrix with preset row and column sizes.

[0119] In some embodiments, the original data may be organized into an n-row and m-column original matrix X, wherein n represents the number of time-domain features, and m represents the number of attributes corresponding to the time-domain features.

[0120] T2. Perform zero-mean conversion on the original matrix to obtain a zero-mean conversion matrix.

[0121] Optionally, each row of the original matrix X may be zero-meaned, that is, the mean of the row may be subtracted; wherein each row of the original matrix X represents an attribute field of the time domain feature.

[0122] T3. Calculate the covariance matrix corresponding to the zero-mean matrix.

[0123] Alternatively, the covariance matrix C can be calculated according to the following formula:

[0124] Where C is a p*p dimensional symmetric matrix;

[0125] T4. Calculate the eigenvalues ​​of the covariance matrix and the eigenvectors corresponding to the eigenvalues.

[0126] In an example, the eigenvalues ​​of the covariance matrix may be obtained as λ1, λ2, ..., λp, and the eigenvectors corresponding to the eigenvalues ​​λ1, λ2, ..., λp.

[0127] T5. According to the magnitude of the eigenvalues, a preset number of eigenvectors are obtained and arranged into a eigenmatrix in rows.

[0128] In an optional embodiment, the feature vectors may be arranged into a matrix in rows from large to small and from top to bottom according to their features, and the first k rows are taken to form the feature matrix P.

[0129] T6. Perform dimensionality reduction on the original matrix according to the characteristic matrix to obtain a reduced-dimensional matrix.

[0130] Optionally, if Y is set as the data after dimensionality reduction, then Y=PX is the data after dimensionality reduction to k dimensions.

[0131] Optionally, when reducing the dimensionality of the original matrix, if the reduced dimension is too low, it is easy to cause the main content of the data to be lost. Therefore, it is necessary to select a suitable k value to ensure that the main content of the data is not lost.

[0132] In an example, please refer to Figure 3, which shows the distribution of eigenvalues ​​obtained by a principal component analysis. It can be seen from the figure that when the eigenvalue is between 2 and 0.25, the eigenvalue is large but decays quickly. When the eigenvalue is less than 0.25, the eigenvalue is small but decays slowly. Therefore, the inflection point can be selected as the k value, that is, k = 3, to ensure that the main content of the data is not lost.

[0133] T7. Use the dimension-reduced matrix as the time-domain training set.

[0134] Optionally, the frequency domain features may be processed using a wavelet packet transform method.

[0135] In some embodiments, the frequency domain features may be processed through the following steps.

[0136] F1. Obtain data sampling frequency, data type, and number of wavelet packet transform processing layers.

[0137] In this embodiment, the data sampling frequency used is 256 Hz.

[0138] Optionally, the sampling frequency can be adjusted based on the accuracy and computational time requirements of the classification application. For example, based on the sampling theorem, the Nyquist sampling frequency is 128 Hz, and the number of wavelet packet transform processing layers is set to 3.

[0139] F2. Processing the data sampling frequency according to the number of wavelet packet transformation processing layers to obtain a preset number of sub-frequency segments;

[0140] In an example, please refer to Figure 4. The Nyquist sampling frequency is divided into two equal parts three times in sequence. The first halving divides the sampling frequency 128Hz into two equal parts to obtain a 0-64Hz frequency band and a 64-128Hz frequency band; the second halving divides the 0-64Hz frequency band and the 64-128Hz frequency band into two equal parts to obtain four groups of frequency bands respectively; the third halving divides the four groups of frequency bands into two equal parts to obtain eight groups of frequency bands respectively; at the same time, the frequency band of each node is named: (0, 0) represents the original signal, and each layer of nodes is named in sequence from low frequency to high frequency, starting from (1, 0), which is No. 1, (1, 1) is No. 2, and so on. In the last layer, (3, 0) is No. 7 and (3, 7) is No. 14.

[0141] In this example, the sub-frequency bands refer to eight groups of frequency bands obtained after the third bisection.

[0142] F3. Extract the energy features corresponding to each of the sub-frequency segments, and obtain the energy proportions corresponding to each of the sub-frequency segments.

[0143] Optionally, the bottom layer nodes (3, 0)-(3, 7) can be subjected to wavelet transformation, that is, low-frequency and high-frequency signals are processed simultaneously, energy features are extracted, and the proportion of frequency band energy of the third layer nodes (3, 0)-(3, 7) is obtained.

[0144] F4. Select the sub-frequency segment combination with the highest energy proportion according to the data type as the frequency domain training set.

[0145] In one example, please refer to Figure 5. It can be seen from Figure 5 that the node (3, 0), that is, the 0-16Hz frequency band, has the highest energy proportion. Therefore, the sub-band node (3, 0), that is, the 0-16Hz frequency band combination is selected as the frequency domain training set.

[0146] In an optional implementation, a preset number of features may be extracted from the time domain training set and the frequency domain training set respectively with replacement, and the features may be combined to obtain training samples.

[0147] In this embodiment, the training samples can be learned through an algorithm to obtain a fault prediction model.

[0148] In this embodiment, the training samples can be feature classified by using a decision tree algorithm, a support vector machine algorithm, and a random forest algorithm, respectively, to obtain characteristic values ​​and a mapping relationship between characteristic values ​​and faults; and the characteristic values ​​and the mapping relationship between characteristic values ​​and faults can be fitted according to a regression algorithm to obtain a fault prediction model.

[0149] It should be noted that the random forest algorithm and the decision tree algorithm are related in many aspects. For example, the decision tree algorithm is the basic building block of the random forest algorithm. The random forest is composed of multiple decision trees. Both can be used for classification and regression problems, and both involve feature selection.

[0150] However, there are certain differences between the random forest algorithm and the decision tree algorithm. First, in terms of structure and complexity, the decision tree algorithm is a single tree structure, which is easier to understand and explain, but prone to overfitting, while the random forest is an integrated model composed of multiple decision trees, which improves accuracy and robustness through the prediction results of multiple decision trees, and is usually more difficult to explain than a single decision tree; secondly, in terms of the training process, the decision tree generally generates a tree structure by recursively splitting the data set, where each split point is based on the best split of a certain feature, while the random forest extracts multiple subsets from the training set through the Bootstrap sampling method, and each subset generates a decision tree, and each tree only considers the feature subset when splitting at each node , thereby increasing diversity; thirdly, in terms of anti-overfitting ability, it is easier to overfit when making decisions because it tends to grow deeply to fully fit the training data, while random forest significantly reduces the risk of overfitting by combining the prediction results of multiple decision trees, and the model is more stable and accurate; in addition, in terms of the prediction process, the decision tree is the prediction result of a single tree, while the random forest is the prediction result of multiple trees. Among them, the voting method is generally used in classification problems, and the average is generally taken in regression problems; finally, in terms of computing resources, the decision tree is relatively computationally efficient and suitable for smaller data sets, while the random forest needs to train multiple trees, so the computing overhead and memory requirements are high, and it is suitable for larger data sets.

[0151] Based on the above statements, the decision tree algorithm is more suitable for scenarios that require simple and easy-to-interpret models, but overfitting problems may occur. The random forest algorithm provides higher prediction accuracy and robustness by integrating multiple decision trees, but it requires higher computing resources and has poor model interpretability. Therefore, it is necessary to select a suitable algorithm based on specific application requirements and data characteristics.

[0152] In this embodiment, the random forest algorithm is used as an example to illustrate:

[0153] In an optional implementation, decision tree training may be performed using the training samples to obtain M decision tree sub-models.

[0154] The number of decision tree sub-models can be determined based on classification accuracy and computation time.

[0155] In one example, the decision tree sub-model in this embodiment may be five decision trees.

[0156] Optionally, a voting method may be used to determine the classification category of the decision tree sub-model with the most votes as the final category.

[0157] Next, the training principle of a single decision tree is introduced with reference to FIG6 .

[0158] Please refer to Figure 6. The training samples are input as feature subsets into the decision tree algorithm for feature classification to obtain the classified feature subsets. With various features as input and fault labels as output, the decision tree constructs the classification in the form of a top-down recursive tree structure.

[0159] In a recursive tree structure, each internal node (represented by a triangle in the diagram) represents a formal logical judgment, such as x1>=a, where x1 is an attribute and a is a value of x1. Leaf nodes (represented by circles in the diagram) represent classes. The edges of the tree are the branching results of the logical judgments, and the branches represent the connections that lead to these classes. The topmost node is the root node. A decision tree is a process that gradually divides data into subsets while gradually expanding the branches of the decision tree. It specifically includes the following steps:

[0160] Step 1: Determine the training sample as the root node;

[0161] Step 2: Traverse each feature in the feature subset according to different selection criteria and select the best threshold as the segmentation point.

[0162] In some embodiments, the above-mentioned fault determination conditions may be used as threshold segmentation points; different selection criteria may include sequential selection or random selection.

[0163] In an optional implementation, the difference in entropy before and after dividing the data set with a certain feature can be calculated. Since the uncertainty of the sample set can be represented by entropy, the greater the entropy, the greater the uncertainty of the sample. Therefore, the difference in set entropy before and after division can be used to measure the effectiveness of using the current feature to divide the training sample.

[0164] Step 3: Split the data according to the split point and create child nodes.

[0165] Step 4: Determine whether the child node has only one category and whether the number of records in the child node is less than a first preset value and the current number of iterations reaches a second preset value. If either of the conditions is met, stop the segmentation; if neither condition is met, determine the child node as a leaf node, and determine the output of the leaf node as the target category with the largest number of categories in the classification category.

[0166] Step 5: Complete the construction of the node.

[0167] Optionally, in order to extract fault status information from fault features, it is necessary to establish a regression model that can set each value in the input as a variable, and then calculate each coefficient of the variable to generate a relationship between the input and output of the model, so that after setting the input value, the corresponding prediction result can be output.

[0168] In this embodiment, after the training samples are classified by the classification model, the following steps may be further included.

[0169] A1. Establish a fault prediction function and use the characteristic value as the function independent variable.

[0170] Optionally, when using the least squares method or the gradient descent method for calculation, an input-output function needs to be established first.

[0171] In this embodiment, the least squares method uses all inputs of the training samples as the input matrix and all outputs as the output matrix, and then obtains the coefficient of each input through matrix calculation; while the gradient descent method further formulates a cost function of the difference between the actual value and the predicted value, and obtains the most suitable coefficient by minimizing the cost function.

[0172] A2. Calculate the mapping relationship between the characteristic value and the fault by the least square method or the gradient descent method to obtain the independent variable coefficient.

[0173] In this embodiment, the coefficients of the variables may be calculated using the least squares method or the gradient descent method.

[0174] A3. Perform regression fitting based on the independent variables and the independent variable coefficients to obtain the fault prediction model.

[0175] Optionally, the collected features may be used as input to a fault prediction model, and the fault prediction model may extract a fault identifier, that is, whether a fault exists and the specific type of the fault, as the output of the model.

[0176] In this embodiment, by collecting characteristic data of the loading and unloading machine to be predicted and inputting it into the fault prediction model, the corresponding fault prediction result can be output.

[0177] It is understandable that the above method can be used to obtain three fault prediction models: a fault prediction model based on a decision tree algorithm, a fault prediction model based on a support vector machine algorithm, and a fault prediction model based on a random forest algorithm. Based on this, the accuracy of the predictions of each fault prediction model can be tested. After obtaining the characteristic signal of the device to be predicted, the characteristic signal of the device to be predicted is input into the fault prediction model with the highest prediction accuracy for prediction. Optionally, to verify the accuracy of the predictions of the above fault models, this example provides a specific experimental case for illustration.

[0178] In this example, two feature extraction methods—principal component analysis (PCA) and wavelet packet transform (WPT)—and three classification methods—random forest, support vector machine (SVM), and decision tree (DT)—were used to establish comparative classification models. Corresponding models were developed for the binary fault prediction problem. Initially, the random forest, support vector machine, and decision tree classifiers were constructed using the original dataset (i.e., 72 features). PCA and WPT algorithms were then used to filter all 72 features. The classifiers were then tested using the optimal filtered feature set.

[0179] Tables 2 and 3 show that the accuracy of the classification model built using PCA+WPT is significantly improved compared to the original dataset. Classifiers applied to the feature selection dataset generally achieve higher accuracy than those directly applied to the original data. Furthermore, the results in Table 2 show that the random forest model can achieve an optimal correct classification rate of up to 98.7% for binary classification problems, indicating that it can accurately predict corresponding equipment faults.

[0180] Table 2. Accuracy of the binary classification model

[0181] Table 3. Runtime of binary classification model

[0182] For the fault classification problem (i.e., the sixteen-class classification problem), the correct classification rates of all models are shown in the table below. Referring to Table 4, both decision tree and random forest models are more effective at classifying single or multiple faults, with average accuracy rates of around 95%. However, the results show that SVM is not suitable for this multi-classification problem, achieving only around 64% classification accuracy. This may be due to noise, and pursuing a perfect fit can lead to overfitting, meaning that a perfect fit of the training data does not yield good predictive performance.

[0183] Table 4. Accuracy of the sixteen-class classification model

[0184] According to Tables 2, 3, and 4, it is clear that the PCA+WPT+Random Forest model is the most effective model. Refer to Figure 7 for the linear function calculation results for nozzle wear fault analysis. It can be clearly seen that almost all samples follow this function and the error is small. Furthermore, through cross-validation experiments, the fault prediction model can well predict whether a fault will occur and the specific type of fault. By calculating, R 2 (Goodness of fit, the closer it is to 1, the higher the fitting accuracy; ( in, Represents the residual sum of squares, Represents the total sum of squares, yi Represents the actual value, Characterize the predicted value, The average value representing the actual value is "0.966", indicating that the two models have high accuracy.

[0185] Optionally, an embodiment of the present disclosure further provides a device for predicting equipment failure. Referring to FIG. 8 , the device for predicting equipment failure includes:

[0186] An extraction module is used to extract a preset number of time nodes and extract characteristic signals corresponding to the time node devices;

[0187] A selection module, configured to select at least one feature extraction method according to the feature type in the feature signal;

[0188] an extraction module, configured to extract features from the feature signal by using at least one feature extraction method to obtain a training sample;

[0189] A training module, inputting the training samples into a machine learning classification model for training to obtain a fault prediction model;

[0190] The prediction module is used to obtain the characteristic signal of the device to be predicted, input the characteristic signal of the device to be predicted into the fault prediction model, and obtain a fault prediction result.

[0191] The embodiment of the present disclosure further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, each step of the above-mentioned method for predicting equipment failure is implemented.

[0192] An embodiment of the present disclosure also provides an electronic device, please refer to Figure 9, the electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, each step in the above-mentioned device fault prediction method can be implemented.

[0193] In summary, the present disclosure provides a method, device, readable storage medium and electronic device for predicting equipment faults. After extracting the characteristic signal of the equipment, different feature extraction methods are selected according to the feature type in the characteristic signal, and different features in the characteristic signal are extracted based on different feature extraction methods. That is, different fault features can be extracted using feature extraction methods suitable for the fault features, thereby realizing secondary processing of the fault features, thereby extracting training samples with a high feature concentration, and then inputting the training samples into a machine learning classification algorithm for data mining, thereby establishing a fault prediction model with a precise input and output relationship, thereby achieving the purpose of accurately predicting equipment faults by combining the established fault prediction model with the characteristic signal of the equipment to be predicted.

[0194] In the above embodiments provided in the present disclosure, it should be understood that the disclosed methods, devices, computer-readable storage media, and electronic devices can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the modules is merely a logical function division. There may be other division methods in actual implementation, such as multiple components or modules can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or components or modules, which can be electrical, mechanical or other forms.

[0195] The components described as separate parts may or may not be physically separate, and the components shown as components may or may not be physical modules, that is, they may be located in one place or distributed across multiple network modules. Some or all of these components may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0196] In addition, the functional modules in the various embodiments of the present disclosure may be integrated into a single processing module, or each component may exist physically separately, or two or more modules may be integrated into a single module. The aforementioned integrated modules may be implemented in the form of hardware or software functional modules.

[0197] If the integrated module is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0198] It should be noted that for the aforementioned method embodiments, for ease of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present disclosure is not limited by the order of the actions described, because according to the present disclosure, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present disclosure.

[0199] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0200] The above descriptions are merely embodiments of the present disclosure and are not intended to limit the patent scope of the present disclosure. Any equivalent transformations made using the contents of the present disclosure and the drawings, or directly or indirectly applied in related technical fields, are also included in the patent protection scope of the present disclosure. Industrial Applicability:

[0201] The embodiments of the present disclosure provide a method, apparatus, readable storage medium, and electronic device for predicting equipment failures, which can achieve the purpose of accurately predicting equipment failures by combining an established fault prediction model with characteristic signals of the equipment to be predicted.

Claims

1. A method for predicting equipment failure, characterized in that: Includes steps: Extracting a preset number of time nodes, and extracting characteristic signals corresponding to the devices at the time nodes; Selecting at least one feature extraction method according to the feature type in the feature signal; Extracting features from the feature signal by using at least one of the feature extraction methods to obtain a training sample; Inputting the training samples into a machine learning classification model for training to obtain a fault prediction model; Acquire a characteristic signal of a device to be predicted, input the characteristic signal of the device to be predicted into the fault prediction model, and obtain a fault prediction result.

2. A method for predicting equipment failure according to claim 1, characterized in that: The extracting features from the feature signal by at least one of the feature extraction methods to obtain a training sample comprises: Extracting time domain features from the characteristic signal by principal component analysis to obtain a time domain training set, and extracting frequency domain features from the characteristic signal by wavelet packet transformation to obtain a frequency domain training set; The training samples are obtained according to the time domain training set and the frequency domain training set.

3. A method for predicting equipment failure according to claim 2, characterized in that: The extracting time domain features from the feature signal by principal component analysis to obtain a time domain training set comprises: Combining the original data in the time domain features into an original matrix of preset row and column sizes; Performing zero mean conversion on the original matrix to obtain a zero mean conversion matrix; Calculate the covariance matrix corresponding to the zero-mean matrix; Calculate the eigenvalues ​​of the covariance matrix and the eigenvectors corresponding to the eigenvalues; According to the size of the eigenvalue, the eigenvectors of a preset number are obtained to form a eigenmatrix in rows; Performing dimension reduction on the original matrix according to the characteristic matrix to obtain a dimension reduction matrix; The dimension reduction matrix is ​​used as the time domain training set.

4. The method for predicting equipment failure according to claim 2, characterized in that: The extracting frequency domain features in the characteristic signal by wavelet packet transformation method to obtain a frequency domain training set comprises: Obtain data sampling frequency, data type and number of wavelet packet transform processing layers; Processing the data sampling frequency according to the number of wavelet packet change method processing layers to obtain a preset number of sub-frequency segments; Extracting energy features corresponding to each of the sub-frequency bands, and obtaining energy proportions corresponding to each of the sub-frequency bands; The sub-frequency segment combination with the highest energy proportion is selected according to the data type as the frequency domain training set.

5. A method for predicting equipment failure according to any one of claims 2 to 4, characterized in that: The obtaining the training samples according to the time domain training set and the frequency domain training set comprises: A preset number of features are extracted from the time domain training set and the frequency domain training set respectively with replacement, and the features are combined to obtain training samples.

6. A method for predicting equipment failure according to claims 1-5, characterized in that: The step of inputting the training samples into a machine learning classification model for training to obtain a fault prediction model comprises: Dividing the training samples into a training set and a validation set; Inputting the training set into the machine learning classification model for training to obtain a regression model for predicting faults; The regression model is verified according to the verification set to obtain the fault prediction model.

7. A method for predicting equipment failure according to claim 6, characterized in that: The machine learning classification model includes a two-classification model and a sixteen-classification model.

8. A method for predicting equipment failure according to claim 7, characterized in that: The verifying the regression model according to the verification set to obtain the fault prediction model includes: inputting the verification set into a verification binary classification model and a verification 16-classification model constructed in advance for training to obtain a verification regression model; Validating the regression model according to the validation regression model; If the regression model passes the verification, the regression model is determined as the fault prediction model.

9. The method according to claim 8, characterized in that The method further comprises: If the regression model fails to pass the verification, the regression model is iteratively trained until the regression model passes the verification.

10. A method for predicting equipment failure according to any one of claims 1 to 9, characterized in that: The inputting the training samples into the machine learning classification model for training to obtain the fault prediction model comprises: The training samples are respectively subjected to feature classification by using a decision tree algorithm, a support vector machine algorithm and a random forest algorithm to obtain feature values ​​and a mapping relationship between the feature values ​​and the faults; The characteristic values ​​and the mapping relationship between the characteristic values ​​and the faults are fitted according to a regression algorithm to obtain a fault prediction model.

11. A method for predicting equipment failure according to claim 10, characterized in that: The step of fitting the characteristic value and the mapping relationship between the characteristic value and the fault according to the regression algorithm to obtain the fault prediction model includes: Establishing a fault prediction function, taking the characteristic value as the function independent variable; Calculate the mapping relationship between the characteristic value and the fault by the least square method or the gradient descent method to obtain the independent variable coefficient; Regression fitting is performed according to the independent variables and the independent variable coefficients to obtain the fault prediction model.

12. A method for predicting equipment failure according to any one of claims 10-11, characterized in that: After fitting the characteristic value and the mapping relationship between the characteristic value and the fault according to the regression algorithm to obtain the fault prediction model, the method further includes: Testing the prediction accuracy of each of the fault prediction models; The step of inputting the device characteristic signal to be predicted into the fault prediction model comprises: The device characteristic signal to be predicted is input into the fault prediction model with the highest prediction accuracy.

13. A device for predicting equipment failure, characterized in that: include: An extraction module, used to extract a preset number of time nodes and extract characteristic signals corresponding to the time node devices; A selection module, used for selecting at least one feature extraction method according to the feature type in the feature signal; An extraction module, used to extract features from the feature signal by using at least one of the feature extraction methods to obtain a training sample; A training module, inputting the training samples into a machine learning classification model for training to obtain a fault prediction model; The prediction module is used to obtain the characteristic signal of the device to be predicted, input the characteristic signal of the device to be predicted into the fault prediction model, and obtain the fault prediction result.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, each step of a method for predicting equipment failure as described in any one of claims 1 to 12 is implemented.

15. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, each step of the device fault prediction method according to any one of claims 1 to 12 is implemented.

Images