Vacuum dry pump fault prediction method and device based on CMS data cost-sensitive learning

Abstract

The vacuum dry pump fault prediction method and device based on CMS data cost-sensitive learning provided by the present invention relate to the technical field of equipment health monitoring. The present invention obtains CMS monitoring data when the vacuum dry pump is running; inputs the CMS monitoring data into a pre-trained vacuum dry pump fault prediction model based on CMS data cost-sensitive learning, and outputs a prediction result; wherein the pre-trained vacuum dry pump fault prediction model obtains filtered data by pre-processing the acquired vacuum dry pump CMS historical monitoring data; then samples based on a sliding window, and combines an oversampling strategy to obtain sampled data of each fixed window length; and performs normalization processing with the window as a unit; finally, based on the Catboost algorithm, the normalized sampled data and the adjusted cost-sensitive loss function are used to train the fault prediction model until the model meets the evaluation criteria, and a trained model is obtained. The present invention can overcome the problem of data imbalance, making the model prediction more accurate and reliable.

Description

Technical Field

[0001] The present invention relates to the technical field of machine equipment health monitoring, and in particular to a vacuum dry pump fault prediction method and device based on CMS data cost-sensitive learning. Background Art

[0002] With the improvement of industrial automation and production efficiency requirements, the traditional maintenance model based on regular maintenance obviously cannot meet the actual needs. In order to reduce equipment downtime, reduce maintenance costs and improve production efficiency, more and more companies are beginning to pay attention to predictive maintenance of equipment. Predictive maintenance can identify potential problems in advance before equipment fails through real-time monitoring and data analysis of equipment operating status, thereby achieving timely intervention and repair.

[0003] In this context, fault prediction of vacuum dry pumps has become an important research field. Most existing prediction technologies rely on sensor data of equipment, such as vibration, temperature and pressure, to perform fault diagnosis and early warning. However, a major challenge facing these technologies is how to accurately identify fault signals in massive amounts of data and make correct predictions for various possible failure modes.

[0004] In recent years, data-driven methods, especially the introduction of machine learning technology, have provided a new way to address this challenge. By deeply analyzing the historical data of the equipment, the machine learning model can learn the characteristics of the equipment in normal and faulty states, thereby improving the accuracy of fault prediction. However, these models often face the problem of data imbalance in practical applications, that is, the normal data is far more than the fault data. The scarcity of fault samples makes it easy for the model to overfit the normal state, resulting in deviations in the prediction results.

[0005] In view of this, the applicant filed this application after studying the existing technology. Summary of the invention

[0006] The present invention aims to provide a vacuum dry pump fault prediction method and device based on CMS data cost-sensitive learning, so as to solve the problem that in the existing method, due to data imbalance, the model is prone to overfitting the normal state, resulting in deviation in the prediction result.

[0007] In order to solve the above technical problems, the present invention is implemented through the following technical solutions:

[0008] A vacuum dry pump fault prediction method based on CMS data cost-sensitive learning, comprising:

[0009] S1, obtain CMS monitoring data when the vacuum dry pump is running;

[0010] S2, inputting the CMS monitoring data into a pre-trained vacuum dry pump fault prediction model based on CMS data cost-sensitive learning, and outputting a prediction result; wherein the prediction result includes the warning level and predicted downtime of the vacuum dry pump, and the warning level is used to indicate the time until the downtime moment;

[0011] The pre-trained vacuum dry pump fault prediction model based on CMS data cost-sensitive learning is trained by the following steps:

[0012] A1, based on the historical CMS monitoring data of the vacuum dry pump, preprocessing is performed to obtain preprocessed filtered data;

[0013] A2, according to the filtered data, preset the sliding window length and sliding step length, perform data sampling based on the sliding window, and combine the fault sample oversampling strategy to obtain the sampling data of each fixed window length; wherein the fault sample oversampling strategy is to dynamically adjust the sliding step length of the next window according to whether the starting position data in the current window is a fault sample, so as to increase the number of fault samples;

[0014] A3, taking the window as the unit, normalizes the sampled data in each window to convert the sampled data from two dimensions to one dimension and capture the local features of the data;

[0015] A4, based on the Catboost algorithm, the normalized sampling data and the adjusted cost-sensitive loss function are used to train the fault prediction model until the model meets the evaluation criteria, and a trained vacuum dry pump fault prediction model based on CMS data cost-sensitive learning is obtained.

[0016] Preferably, the CMS monitoring data is the time series data of status information collected by the CMS every second when the vacuum dry pump is running, including running time, lower pump power, upper pump power, lower pump speed and lower pump temperature.

[0017] Preferably, the filtering data includes filtering data of downtime pumps and filtering data of non-downtime pumps; the downtime pump is a vacuum dry pump that has downtime before a specified time, and the data filtering operation is specifically as follows:

[0018] Delete the abnormal data, and then use the valid data of the preset time period before the vacant moment to fill it; if the vacant moment is the starting moment of the data, then discard this moment and use the valid data of the next moment as the starting moment; wherein, the abnormal data is that within the preset unit time length, the maximum values ​​of the upper pump power and the lower pump power of the vacuum dry pump exceed the threshold value;

[0019] Adjust the warning label of downtime pumps whose remaining life time is less than the threshold.

[0020] Preferably, when combining the fault sample oversampling strategy, for the data in each window, if the data at the starting position in the current window is a fault sample, the sliding step size of the next window is adjusted to the second sliding step size, and continue to judge whether the data at the starting position of the next window is a fault sample. If not, restore to the original sliding step size until all windows are traversed. The expression is as follows:

[0021] ;

[0022] in, is the starting position of the next window, is the starting position of the current window, for Warning labels at all times, is the step size function, defined as:

[0023] ;

[0024] in, is the sliding step length, is the second sliding step length, > ;

[0025] represents a normal sample; represents a fault sample;

[0026] Each sample data in the dataset contains multiple features and a warning label, expressed as:

[0027] ;

[0028] in, is the feature vector at the tth time point, express The corresponding warning label, N is the total number of samples.

[0029] Preferably, the specific steps of normalizing the sampled data in units of windows are as follows:

[0030] For each feature j in each window, calculate its data in the window The mean and standard deviation , and the calculation formulas are:

[0031] ;

[0032] ;

[0033] is a w×c matrix, expressed as:

[0034] ;

[0035] in, , n is the total number of windows of sampled data, w is the window length, and c represents the number of features; is the value of the jth feature at the tth time point; is the value of the cth feature in the ith window;

[0036] Then, each eigenvalue in the window Normalized to , the standardized formula is:

[0037] ;

[0038] in, is the characteristic value, is the standardized value.

[0039] Preferably, during the model training process, the difference between the predicted value and the true label is optimized by using an adjusted cost-sensitive loss function, and the expression of the adjusted cost-sensitive loss function is:

[0040] ;

[0041] in, is the adjusted cost sensitivity function, is the true label of the i-th sample, is the predicted probability after cost-sensitive adjustment, The formula is;

[0042] ;

[0043] ;

[0044] ;

[0045] in, is the weighted average cost; is the logarithmic adjustment parameter of the cost ratio; is the probability of predicting the i-th sample as a positive class;

[0046] C 01 is the cost of misclassifying a negative sample as a positive sample; C 10 The cost of misclassifying a positive sample as a negative sample;

[0047] According to the imbalance degree of the sample The misclassification cost parameter is set, and the degree of sample imbalance The calculation formula is:

[0048] ;

[0049] in, is the number of positive samples, is the number of negative samples; generally, set Value is Value times.

[0050] Preferably, based on the Catboost algorithm, the adjusted cost-sensitive function is used as the loss function, and the loss function of the model is minimized by gradient boosting, thereby optimizing the performance of the vacuum dry pump fault prediction model, as follows:

[0051] First, the cost-sensitive objective function of the Catboost algorithm is constructed, and its expression is:

[0052] ;

[0053] in, represents the tth decision tree generated during the training process, Represents the predicted value of the tth decision tree;

[0054] Then, the gradient loss function is used to calculate the residual between the predicted value and the actual value of the current model, that is, the rate of change of the model output. The expression of the gradient loss function adjusted by the cost-sensitive function is:

[0055] ;

[0056] in, Represents the gradient of the model at the tth decision tree; , , is the Sigmoid activation function; represents the first-order partial derivative;

[0057] Next, the second-order gradient Hessian function is used to measure the curvature of the loss function to correct the error of the fitting model. The expression of the second-order gradient Hessian function adjusted by the cost-sensitive function is:

[0058] ;

[0059] in, represents the second-order gradient of the t-th tree; represents the second-order partial derivative;

[0060] Finally, update the leaf node value of the model, the leaf node value The update formula of is affected by cost sensitivity:

[0061] ;

[0062] in Represents the samples contained in the jth leaf node of the decision tree;

[0063] The model update rule based on cost-sensitive adjustment is obtained as follows:

[0064] ;

[0065] in, is the learning rate, which is used to control the step size of the model update in each iteration; Represents the t-th tree for the training sample The predicted output.

[0066] Preferably, the specific operation of judging whether the model meets the evaluation criteria is:

[0067] For each sliding window, calculate the sum of the predicted values ​​within the window , the calculation formula is:

[0068] ;

[0069] Among them, p(t ) is the prediction result at time point t in the i-th window. If p(t ) = 1, indicating that a failure is predicted in the future; p(t ) = 0, indicating that no failure is predicted in the future;

[0070] When the Q(t) of the sliding window exceeds the set threshold for the first time When, that is:

[0071] ;

[0072] The starting time point of the current window is regarded as the first predicted downtime, that is, ;

[0073] The actual downtime of the sample that crashed in the test set is recorded as ;when The prediction is considered correct when the following conditions exist and are met:

[0074] ;

[0075] Where, d is the allowed time difference threshold;

[0076] The number of correct predictions for:

[0077] ;

[0078] Where N is the total number of samples; is an indicator function, which is 1 when the condition is met and 0 otherwise;

[0079] is the actual downtime of the i-th sample;

[0080] is the predicted downtime of the i-th sample;

[0081] If the predicted time Than the actual failure time Earlier than d, that is:

[0082] , it is judged as a morning newspaper;

[0083] The number of morning papers for:

[0084] ;

[0085] False positives refer to the time when the model predicts a failure to occur , but no failure actually occurred, that is, the actual downtime of the sample does not exist, then the number of false positives for:

[0086] ;

[0087] in, Indicates that no actual failure has occurred;

[0088] False negatives refer to the failure that actually occurred but was not predicted by the model. If it does not exist, the number of missed reports for:

[0089] ;

[0090] Then, the model prediction accuracy The calculation formula is:

[0091] ;

[0092] in, Indicates the number of all predicted downtime events;

[0093] Then, the model predicts the false alarm rate The calculation formula is:

[0094] ;

[0095] Then, the model predicts the false negative rate The calculation formula is:

[0096] ;

[0097] When the false positive rate of the model , false negative rate and accuracy The model is evaluated as qualified when it meets the preset requirements.

[0098] The present invention also provides a vacuum dry pump fault prediction device based on CMS data cost-sensitive learning, comprising:

[0099] An acquisition unit, used to acquire CMS monitoring data when the vacuum dry pump is running;

[0100] A model application unit, used to input the CMS monitoring data into a pre-trained vacuum dry pump fault prediction model based on CMS data cost-sensitive learning, and output a prediction result; wherein the prediction result includes an early warning level and a predicted downtime of the vacuum dry pump, and the early warning level is used to indicate the time until the downtime moment;

[0101] The pre-trained vacuum dry pump fault prediction model based on CMS data cost-sensitive learning is trained by the following steps:

[0102] A1, based on the historical CMS monitoring data of the vacuum dry pump, preprocessing is performed to obtain preprocessed filtered data;

[0103] A2, according to the filtered data, preset the sliding window length and sliding step length, perform data sampling based on the sliding window, and combine the fault sample oversampling strategy to obtain the sampling data of each fixed window length; wherein the fault sample oversampling strategy is to dynamically adjust the sliding step length of the next window according to whether the starting position data in the current window is a fault sample, so as to increase the number of fault samples;

[0104] A3, taking the window as the unit, normalizes the sampled data in each window to convert the sampled data from two dimensions to one dimension and capture the local features of the data;

[0105] A4, based on the Catboost algorithm, the normalized sampling data and the adjusted cost-sensitive loss function are used to train the fault prediction model until the model meets the evaluation criteria, and a trained vacuum dry pump fault prediction model based on CMS data cost-sensitive learning is obtained.

[0106] Preferably, the model application unit comprises:

[0107] The training data processing module is used to obtain preprocessed filtered data through preprocessing based on the acquired historical CMS monitoring data of the vacuum dry pump;

[0108] A sliding window sampling module is used to preset a sliding window length and a sliding step size according to the filtered data, perform data sampling based on the sliding window, and combine the fault sample oversampling strategy to obtain sampling data of each fixed window length; wherein the fault sample oversampling strategy is to dynamically adjust the sliding step size of the next window according to whether the starting position data in the current window is a fault sample, so as to increase the number of fault samples;

[0109] The normalization module is used to normalize the sampled data in each window in units of windows, so as to convert the sampled data from two dimensions to one dimension and capture the local features of the data;

[0110] The model training module is used to train the fault prediction model based on the Catboost algorithm, using the normalized sampling data and the adjusted cost-sensitive loss function until the model meets the evaluation criteria, thereby obtaining a trained vacuum dry pump fault prediction model based on CMS data cost-sensitive learning.

[0111] The present invention also provides a vacuum dry pump fault prediction device based on CMS data cost-sensitive learning, comprising a processor and a memory, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement a vacuum dry pump fault prediction method based on CMS data cost-sensitive learning as described above.

[0112] The present invention also provides a computer-readable storage medium, on which computer-readable instructions are stored. When the computer-readable instructions are executed by a processor of a device where the computer-readable storage medium is located, a vacuum dry pump fault prediction method based on CMS data cost-sensitive learning as described above is implemented.

[0113] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0114] The present invention combines sliding window sampling, cost-sensitive learning and Catboost algorithm to solve the problem of overfitting of the model and biased prediction results when the data is unbalanced (i.e., the normal data is far more than the fault data). The present invention combines the fault sample oversampling strategy to dynamically shorten the sliding step of the window and increase the number of fault sample labels; through the normalization operation, the local characteristics of the time series characteristics are fully captured, which facilitates the machine learning model to perform effective feature learning and prediction; and through the adjusted cost-sensitive function, cost-sensitive learning is performed to make the model more sensitive to the minority class, i.e., fault data, when processing unbalanced data. Compared with traditional methods, the method of the present invention can overcome the data imbalance problem in traditional prediction methods and can provide more accurate and reliable fault warning predictions in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0115] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0116] Figure 1 A schematic diagram of a vacuum dry pump fault prediction method based on CMS data cost-sensitive learning provided in Example 1 of the present invention.

[0117] Figure 2 A training flow chart of a vacuum dry pump fault prediction model based on CMS data cost-sensitive learning provided in Example 1 of the present invention.

[0118] Figure 3 A schematic diagram of the prediction application of the vacuum dry pump fault prediction model provided in Example 1 of the present invention.

[0119] Figure 4 A schematic diagram of a vacuum dry pump fault prediction device based on CMS data cost-sensitive learning provided in Example 2 of the present invention.

[0120] Figure 5 A schematic diagram of each module in the model application unit provided in the second embodiment of the present invention.

[0121] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. DETAILED DESCRIPTION

[0122] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention claimed for protection, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0123] Embodiment 1

[0124] Embodiment 1 of the present invention provides a vacuum dry pump fault prediction method based on CMS data cost-sensitive learning, which can be implemented by a vacuum dry pump fault prediction device based on CMS data cost-sensitive learning, and in particular, executed by one or more processors in the vacuum dry pump fault prediction device.

[0125] In this embodiment, the vacuum dry pump fault prediction device may be an electronic device equipped with a processor, which has a computer program of the vacuum dry pump fault prediction method based on CMS data cost-sensitive learning and the computer program can be executed, such as a computer, a smart phone, a smart tablet, a workstation, etc.

[0126] In this embodiment, CMS (Central Monitoring System, central equipment status monitoring system), CMS monitoring data involves real-time monitoring of equipment status, data analysis and alarm functions.

[0127] like Figure 1 As shown, a vacuum dry pump fault prediction method based on CMS data cost-sensitive learning includes steps S1 to S2.

[0128] S1, obtain CMS monitoring data when the vacuum dry pump is running;

[0129] S2, inputting the CMS monitoring data into a pre-trained vacuum dry pump fault prediction model based on CMS data cost-sensitive learning, and outputting a prediction result; wherein the prediction result includes the warning level and predicted downtime of the vacuum dry pump, and the warning level is used to indicate the time until the downtime moment;

[0130] like Figure 2 As shown, the pre-trained vacuum dry pump fault prediction model based on CMS data cost-sensitive learning is trained by the following steps:

[0131] A1, based on the historical CMS monitoring data of the vacuum dry pump, the preprocessed filtered data is obtained through preprocessing.

[0132] Furthermore, the CMS monitoring data is time series data of status information collected from the CMS every second when the vacuum dry pump is running, including running time, lower pump power, upper pump power, lower pump speed and lower pump temperature.

[0133] Furthermore, the CMS monitoring data of the historical vacuum dry pumps are preprocessed to filter abnormal data and obtain filtered data of downtime pumps and non-downtime pumps.

[0134] The downtime pump is in the vacuum dry pump operation data collected at a specified time. According to the equipment status and time period, the vacuum dry pump that has downtime before the specified time is identified as a downtime pump. On the contrary, the vacuum dry pump that has not downtime before the limit is identified as a non-downtime pump.

[0135] The data filtering operations are as follows:

[0136] In this embodiment, the historical feature data is divided into data segments according to the unit time length T hours, and each data segment is filtered.

[0137] The filtering condition is: within a unit time, if the maximum value of the upper pump power exceeds 1.5 times the minimum value and the maximum value of the lower pump power also exceeds 1.5 times the minimum value, or the maximum value of the upper pump power is less than 1.5 times the minimum value and the maximum value of the lower pump power exceeds 1.25 times the minimum value, then the data segment is determined to be abnormal and deleted. After deleting the abnormal data, if there is a data gap, the gap is filled with the most recent valid data before the gap moment. If the gap occurs at the start time of the data, the moment is discarded. Through this process, the filtered downtime pump and non-downtime pump data sets are obtained.

[0138] Then, adjust the warning label of the downtime pump whose remaining life time is less than the threshold. For example, for the filtered data of downtime pumps and non-downtime pumps, each sample data contains the following fields: remaining date (Remaining_Days) and warning level (Warning_Level), where the remaining date indicates the remaining life time of the pump, and the format example is "7 days 23:59:57".

[0139] Define a threshold number of days L, and let the remaining life of the i-th record in the filtered data of the downtime pump and the non-downtime pump be represented as a string , the warning level is Parse the remaining lifetime string into days, hours, minutes, and seconds, denoted as , , and . Calculate the total number of seconds and total days , for each record, if it satisfies:

[0140] and ;

[0141] Then, Modified to 1, the corresponding formula is as follows:

[0142] ;

[0143] After processing the i-th sample data in the filtered data of the downtime pump and the non-downtime pump The value is recorded as the warning level, that is, the target label y i .

[0144] The filtered data of downtime pumps and non-downtime pumps with warning levels are divided into training samples and test samples in a ratio of 7:3. The training samples of downtime pumps are merged with the training samples of non-downtime pumps to form a complete training set; similarly, the test samples of downtime pumps are merged with the test samples of non-downtime pumps to form the final test set.

[0145] A2, according to the filtered data, preset the sliding window length and sliding step size, perform data sampling based on the sliding window, and combine the fault sample oversampling strategy to obtain sampling data of each fixed window length; wherein, the fault sample oversampling strategy is to dynamically adjust the sliding step size of the next window according to whether the starting position data in the current window is a fault sample, so as to increase the number of fault samples.

[0146] In this embodiment, a sliding window data sampling method is used to extract features and label information from the training set and the test set, and oversampling of non-downtime pump fault samples is achieved by dynamically adjusting the sliding step size to increase the number of fault samples.

[0147] In this embodiment, the fault data in the training set is sampled using sliding window sampling; for a given time series data, let its length be N, and each time point contains multiple features and a target label. In the sliding window sampling process of time series data, each sample in the data set contains multiple features and a label, which is expressed as , where x t is the feature vector at the tth time point, represents the corresponding label, and N is the total number of samples.

[0148] Define the length of the sliding window as w and the sliding step as Each sliding window starts from the first sample of the time series and sequentially intercepts subsequences of fixed length w to form a window. For the i-th window, the feature data it contains is The corresponding label is , where t i The starting position of the window.

[0149] like = 1 (i.e., faulty samples), then after the window ends, the oversampling strategy is applied to shorten the sliding step size to = 3600 seconds (1 hour) to increase the frequency of fault samples; otherwise, the step size remains at the original value The oversampling process can be expressed by the following formula:

[0150] The starting position of the next window is: ;

[0151] The step function Defined as:

[0152] ;

[0153] in, is the sliding step length, is the second sliding step length, > .

[0154] A3, taking the window as the unit, normalizes the sampled data in each window to convert the sampled data from two dimensions to one dimension and capture the local features of the data.

[0155] In this embodiment, for the training set and test set data sampled in step A3, at this time, the data set contains window data of fixed length w. , n is the total number of training set and test set data windows after sampling, is represented as a w×c matrix, where c represents the number of features. Specifically, for each window , we have:

[0156] ;

[0157] Next, each feature in each window is normalized. The specific method is to calculate the normalization factor of each feature j in the window. The mean and standard deviation , and the calculation formulas are:

[0158] ;

[0159] ;

[0160] Then, each eigenvalue in the window Normalized to , the standardized formula is:

[0161] ;

[0162] in, is the characteristic value, is the standardized value.

[0163] The data matrix after normalization The data is flattened into a one-dimensional vector for subsequent model input. Specifically, the data of each window is converted from two dimensions (time step, number of features) to one dimension (number of features * time step). The processed data can fully capture the local characteristics of the time series features, which facilitates the machine learning model to perform effective feature learning and prediction.

[0164] A4, based on the Catboost algorithm, the normalized sampling data and the adjusted cost-sensitive loss function are used to train the fault prediction model until the model meets the evaluation criteria, and a trained vacuum dry pump fault prediction model based on CMS data cost-sensitive learning is obtained.

[0165] In this embodiment, for two-class classification problems, positive and negative samples are represented by symbols "1" and "0" respectively. Cost-sensitive learning measures the unequal misclassification losses caused by misclassification of machine learning algorithms by defining a misclassification cost matrix.

[0166] In the cost matrix, C 00 represents the cost of correctly classifying the normal category, C 01 represents the cost of misclassification of the normal category, C 10 represents the cost of misclassifying a fault category, C 11 Indicates the cost of correctly classifying the fault category. Correspondingly, the larger the corresponding value, the greater the loss caused by the wrong classification. Among them, the negative class (warning label is 0) is identified as healthy data, and the positive class (warning label is 1) is identified as fault data;

[0167] In the binary classification problem, the standard logarithmic loss function (Log Loss) is defined as:

[0168] ;

[0169] here, is the true label of the ith sample (valued as 0 or 1), and is the probability of predicting the i-th sample as a positive class. From Table 1, we can see that the cost of misclassifying a negative sample as a positive sample is C 01 , the cost of misclassifying a positive sample as a negative sample is C 10 , define two new tuning parameters:

[0170] Weighted Average Cost :

[0171] ;

[0172] Logarithmic adjustment parameter of the cost ratio :

[0173] ;

[0174] According to the imbalance degree of the sample The misclassification cost parameter is calibrated to determine the degree of sample imbalance The calculation formula is as follows:

[0175] ;

[0176] in is the number of positive samples, is the number of negative samples.

[0177] In this embodiment, set Value is Value times, normal, default The value is 1.0.

[0178] Based on these two parameters, the model's predicted probability is adjusted to reflect the difference in the cost of different misclassifications. The calculation is as follows:

[0179] ;

[0180] Here is the predicted probability after cost-sensitive adjustment, taking into account the misclassification cost between different categories. The final cost-sensitive logarithmic loss function is:

[0181] ;

[0182] Here is the cost-sensitive adjusted loss function.

[0183] Then, the constructed cost-sensitive function is used as the loss function to train the CMS data cost-sensitive learning fault prediction model built based on the Catboost algorithm.

[0184] In this embodiment, the Catboost algorithm is used to build and train the prediction model. The Catboost algorithm improves the performance of the model by building multiple decision trees and gradually optimizing the prediction capabilities of these trees using a unique gradient boosting method. During the training process, the model uses the cost-sensitive loss function described above to optimize the difference between the predicted value and the true label, and gradually adjusts the decision tree based on gradient boosting.

[0185] In this embodiment, gradient boosting is an ensemble learning technique for regression and classification problems, which builds a strong prediction model by combining multiple weak prediction models (such as decision trees). The core idea of ​​the gradient boosting method is to gradually reduce the value of the model's loss function (such as mean square error or logarithmic loss).

[0186] During the construction of each tree, Catboost determines the split and output value of the leaf node by minimizing the cost-sensitive objective function. The cost-sensitive objective function takes into account the difference in the cost of different misclassifications, making the model more sensitive to high-cost errors. Each leaf node represents a specific segmentation area, and its output value is the predicted value of all samples in the area. The output value of the leaf node is obtained by minimizing the loss function within each leaf node, so that the prediction error of the model in each segmentation area is minimized, thereby reducing the loss and improving the prediction accuracy, as follows:

[0187] First, under the adjustment of the cost-sensitive function, the cost-sensitive objective function of the model based on the Catboost algorithm is constructed, which is expressed as:

[0188] ;

[0189] in, represents the tth decision tree generated during the training process, Represents the predicted value of the tth tree.

[0190] Then, the gradient loss function is used to calculate the residual between the predicted value and the actual value of the current model, that is, the rate of change of the model output. The expression of the gradient loss function adjusted by the cost-sensitive function is:

[0191] ;

[0192] in, Represents the gradient of the model at the tth decision tree; , , is the Sigmoid activation function; represents the first-order partial derivative;

[0193] Next, the second-order gradient Hessian function is used to measure the curvature of the loss function to correct the error of the fitting model, which has an important impact on accelerating the convergence of model optimization. Correspondingly, the second-order gradient Hessian function is also affected by the cost sensitivity adjustment, and its expression is:

[0194] ;

[0195] in, represents the second-order gradient of the t-th tree, represents the second-order partial derivative.

[0196] During the training process of the decision tree, the update of the leaf node value directly affects the tree construction process. The update formula of is affected by cost sensitivity:

[0197] ;

[0198] in Represents the samples contained in the j-th leaf node of the decision tree.

[0199] Finally, the model update rule based on cost-sensitive adjustment is obtained as follows:

[0200] ;

[0201] in, represents the predicted value of the t+1th tree, is the learning rate, which is used to control the step size of the model update in each iteration; Represents the t-th tree for the training sample The predicted output.

[0202] The trained model is used to predict faults for new input features, and the results can be used for decision-making and early warning in practical applications.

[0203] The divided test set is used as the input of the trained CMS data-based cost-sensitive learning fault prediction model, and tests are continuously performed according to the evaluation criteria until the trained model meets the evaluation criteria.

[0204] The normalized test set is input into the trained CMS data cost-sensitive learning fault prediction model, and the model outputs the warning level label corresponding to each test sample in the test set.

[0205] According to the warning level label corresponding to each test sample, the model is judged whether it is qualified according to the evaluation criteria. If the model is qualified, the pre-trained CMS data cost-sensitive learning fault prediction model is obtained, otherwise the model is retrained.

[0206] Furthermore, the specific operations to determine whether the model meets the evaluation criteria are:

[0207] The warning level is statistically predicted using sliding window technology.

[0208] Specifically, we accumulate the prediction results in a sliding window of h hours, and the step size of the sliding window is set to 1 hour. Let p(t) be the prediction result at time point t, if p(t) = 1 means that a failure is predicted in the future, and p(t) = 0 means that no failure is predicted in the future.

[0209] For each sliding window, calculate the sum of the predicted values ​​within the window , the calculation formula is:

[0210] ;

[0211] Among them, p(t ) is the prediction result at time point t in the i-th window. If p(t ) = 1, indicating that a failure is predicted in the future; p(t ) = 0, indicating that no failure is predicted in the future;

[0212] When the Q(t) of the sliding window exceeds the set threshold for the first time When, that is:

[0213] ;

[0214] The starting time point of the current window is regarded as the first predicted downtime, that is, ;

[0215] The actual downtime of the sample that crashed in the test set is recorded as ;when The prediction is considered correct when the following conditions exist and are met:

[0216] ;

[0217] Where, d is the allowed time difference threshold;

[0218] The number of correct predictions for:

[0219] ;

[0220] Where N is the total number of samples; is an indicator function, which is 1 when the condition is met and 0 otherwise;

[0221] is the actual downtime of the i-th sample;

[0222] is the predicted downtime of the i-th sample;

[0223] If the predicted time Than the actual failure time Earlier than d, that is:

[0224] , it is judged as a morning newspaper;

[0225] The number of morning papers for:

[0226] ;

[0227] False positives refer to the time when the model predicts a failure to occur , but no failure actually occurred, that is, the actual downtime of the sample does not exist, then the number of false positives for:

[0228] ;

[0229] in, Indicates that no actual failure has occurred;

[0230] False negatives refer to the failure that actually occurred but was not predicted by the model. If it does not exist, the number of missed reports for:

[0231] ;

[0232] Then, the model prediction accuracy The calculation formula is:

[0233] ;

[0234] in, Indicates the number of all predicted downtime events;

[0235] Then, the model predicts the false alarm rate The calculation formula is:

[0236] ;

[0237] Then, the model predicts the false negative rate The calculation formula is:

[0238] ;

[0239] When the false positive rate of the model , false negative rate and accuracy When the preset requirements are met, the model evaluation is qualified. The preset value is set manually by those skilled in the art according to actual needs, and the present invention does not make any specific limitation on this.

[0240] After the model has been iteratively trained to reach the maximum number of iterations, a trained prediction model is obtained and applied to actual working conditions. The characteristic data under real-time working conditions is input into the prediction model to output the prediction results.

[0241] like Figure 3 As shown, in this embodiment, the state information time series data collected by CMS per second when the vacuum dry pump is running under a certain process condition is selected. , including running time , lower pump power , Pump power , lower pump speed , lower pump temperature .

[0242] After preprocessing, the input data is obtained }, where the sequence in brackets represents data fragments of multiple time windows. Input the data into the trained CMS data cost-sensitive learning fault prediction model and output the predicted downtime .in, Indicates the application under working conditions.

[0243] In another preferred embodiment, in order to verify the effectiveness of the model and model solution proposed in the present invention, the historical operation data of a vacuum dry pump under a certain process condition of a certain company is selected as a research object.

[0244] The data information includes 74 pumps operating normally and 41 pumps that are down from December 2023 to May 2024, a total of 115 vacuum dry pumps. 70% of the vacuum dry pumps operating normally and the vacuum dry pumps that are down are selected as training sets, and 30% of the vacuum dry pumps operating normally and the vacuum dry pumps that are down are selected as test sets, that is, there are 28 down pumps and 52 normal pumps, a total of 80 vacuum dry pumps as training sets, 12 down pumps and 23 normal pumps, a total of 35 vacuum dry pumps as test sets, and the operating characteristic data includes the operating time R 1 , lower pump power R 2 , upper pump power R 3 , lower pump speed R 4 and the lower pump temperature R 5 , see Table 1 for details.

[0245] Table 1. Operational characteristic data of a vacuum dry pump

[0246]

[0247] For downtime pumps, set the data labels of vacuum dry pump remaining life Remaining_Days≤8 to 1, and the rest to 0. For non-downtime pumps, discard the data of vacuum dry pump remaining life Remaining_Days≤8, and the rest to 0. Set the input window to 8 hours, the sliding sampling distance to 3 hours, and the sliding oversampling sliding step to 1 hour.

[0248] The experimental parameters of the present invention are set as shown in Table 2 below:

[0249] Table 2. Experimental parameter settings

[0250]

[0251] Thus, the accuracy rate, false alarm rate and missed alarm rate are calculated according to the formula:

[0252] ;

[0253] ;

[0254] ;

[0255] The allowed time difference threshold d is set to 31 days, that is, an early warning within 0 to 31 days of the downtime of the vacuum dry pump is considered a correct prediction, more than 31 days is considered a false alarm, no early warning is considered a missed alarm, and an early warning for a vacuum dry pump that is not down is considered a false alarm.

[0256] The cost-sensitive Catboost method of the present invention and the traditional Catboost algorithm are used to train and test the model, and their performance is observed. The alarm situation of the vacuum dry pump is shown in Table 3, and the calculation results of the accuracy rate, false alarm rate and missed alarm rate are shown in Table 4.

[0257] Table 3. Alarm conditions of vacuum dry pump

[0258]

[0259] Table 4. Calculation results of accuracy, false alarm rate and missed alarm rate

[0260]

[0261] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0262] The present invention combines sliding window sampling, cost-sensitive learning and Catboost algorithm to solve the problems of model overfitting and prediction result deviation when the data is unbalanced (i.e., normal data is far more than fault data). The present invention combines the fault sample oversampling strategy to dynamically adjust the sliding step size of the window and increase the number of fault sample labels; through normalization operations, the local characteristics of the time series features are fully captured, which facilitates the machine learning model to perform effective feature learning and prediction; through cost-sensitive learning, the model is more sensitive to the minority class, i.e., fault data, when processing unbalanced data. Compared with the traditional Catboost method, the method of the present invention (cost-sensitive Catboost) can overcome the data imbalance problem in traditional prediction methods and can provide more accurate and reliable fault warning predictions in practical applications.

[0263] Embodiment 2

[0264] like Figure 4 As shown, the second embodiment of the present invention further provides a vacuum dry pump fault prediction device based on CMS data cost-sensitive learning, comprising:

[0265] An acquisition unit 201 is used to acquire CMS monitoring data when the vacuum dry pump is running;

[0266] The model application unit 202 is used to input the CMS monitoring data into a pre-trained vacuum dry pump fault prediction model based on CMS data cost-sensitive learning, and output a prediction result; wherein the prediction result includes the warning level and predicted downtime of the vacuum dry pump, and the warning level is used to indicate the time from the downtime moment;

[0267] The pre-trained vacuum dry pump fault prediction model based on CMS data cost-sensitive learning is trained by the following steps:

[0268] A1, based on the acquired historical CMS monitoring data of the vacuum dry pump, preprocessing is performed to obtain preprocessed filtered data;

[0269] A2, according to the filtered data, preset the sliding window length and sliding step length, perform data sampling based on the sliding window, and combine the fault sample oversampling strategy to obtain the sampling data of each fixed window length; wherein the fault sample oversampling strategy is to dynamically adjust the sliding step length of the next window according to whether the starting position data in the current window is a fault sample, so as to increase the number of fault samples;

[0270] A3, taking the window as the unit, normalizes the sampled data in each window to convert the sampled data from two dimensions to one dimension and capture the local features of the data;

[0271] A4, based on the Catboost algorithm, the normalized sampling data and the adjusted cost-sensitive loss function are used to train the fault prediction model until the model meets the evaluation criteria, and a trained vacuum dry pump fault prediction model based on CMS data cost-sensitive learning is obtained.

[0272] like Figure 5 As shown, the model application unit 202 includes:

[0273] The training data processing module 301 is used to obtain preprocessed filtered data through preprocessing according to the acquired historical CMS monitoring data of the vacuum dry pump;

[0274] The sliding window sampling module 302 is used to preset the sliding window length and sliding step length according to the filtered data, perform data sampling based on the sliding window, and combine the fault sample oversampling strategy to obtain the sampling data of each fixed window length; wherein the fault sample oversampling strategy is to dynamically adjust the sliding step length of the next window according to whether the starting position data in the current window is a fault sample, so as to increase the number of fault samples;

[0275] A normalization module 303 is used to perform normalization processing on the sampled data in each window in units of windows, so as to convert the sampled data from two dimensions to one dimension and capture the local features of the data;

[0276] The model training module 304 is used to train the fault prediction model based on the Catboost algorithm using the normalized sampled data and the adjusted cost-sensitive loss function until the model meets the evaluation criteria, thereby obtaining a trained vacuum dry pump fault prediction model based on CMS data cost-sensitive learning.

[0277] Embodiment 3

[0278] The third embodiment of the present invention also provides a vacuum dry pump fault prediction device based on CMS data cost-sensitive learning, which includes a memory and a processor, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement the vacuum dry pump fault prediction method based on CMS data cost-sensitive learning as described above.

[0279] Embodiment 4

[0280] The fourth embodiment of the present invention also provides a computer-readable storage medium, on which computer-readable instructions are stored. When the computer-readable instructions are executed by a processor of a device where the computer-readable storage medium is located, the vacuum dry pump fault prediction method based on CMS data cost-sensitive learning as described above is implemented.

[0281] In several embodiments provided in the embodiments of the present invention, it should be understood that the disclosed apparatus and method can also be implemented in other ways. The apparatus and method embodiments described above are merely schematic. For example, the flowcharts in the accompanying drawings show the possible architecture, functions and operations of the apparatus, method and computer program product according to multiple embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, program segment or a part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0282] In addition, the functional modules in the various embodiments of the present invention may be integrated together to form an independent part, or each module may exist independently, or two or more modules may be integrated to form an independent part.

[0283] If the function 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 the part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, an electronic device, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program code. It should be noted that in this article, the term "include", "include" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements that are not explicitly listed, or also includes elements inherent to such process, method, article or device. Without more constraints, an element defined by the phrase "comprising a..." does not exclude the existence of other identical elements in the process, method, article or apparatus comprising the element.

[0284] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The singular forms "a", "said" and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings.

[0285] It should be understood that the term "and / or" used in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0286] The word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to determining" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)", depending on the context.

[0287] The "first\second" mentioned in the embodiments is only to distinguish similar objects, and does not represent a specific order for the objects. It is understandable that the "first\second" can be interchanged with the specific order or sequence where permitted. It should be understood that the objects distinguished by "first\second" can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than those illustrated or described herein.

[0288] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A vacuum dry pump fault prediction method based on CMS data cost-sensitive learning, characterized in that: include: S1, obtain CMS monitoring data when the vacuum dry pump is running; S2, inputting the CMS monitoring data into a pre-trained vacuum dry pump fault prediction model based on CMS data cost-sensitive learning, and outputting a prediction result; wherein the prediction result includes the warning level and predicted downtime of the vacuum dry pump, and the warning level is used to indicate the time until the downtime moment; The pre-trained vacuum dry pump fault prediction model based on CMS data cost-sensitive learning is trained by the following steps: A1, based on the historical CMS monitoring data of the vacuum dry pump, preprocessing is performed to obtain preprocessed filtered data; A2, according to the filtered data, preset the sliding window length and sliding step length, perform data sampling based on the sliding window, and combine the fault sample oversampling strategy to obtain the sampling data of each fixed window length; wherein the fault sample oversampling strategy is to dynamically adjust the sliding step length of the next window according to whether the starting position data in the current window is a fault sample, so as to increase the number of fault samples; A3, taking the window as the unit, normalizes the sampled data in each window to convert the sampled data from two dimensions to one dimension and capture the local features of the data; A4, based on the Catboost algorithm, the normalized sampling data and the adjusted cost-sensitive loss function are used to train the fault prediction model until the model meets the evaluation criteria, and a trained vacuum dry pump fault prediction model based on CMS data cost-sensitive learning is obtained; wherein, during the model training process, the difference between the predicted value and the true label is optimized by the adjusted cost-sensitive loss function, and the expression of the adjusted cost-sensitive loss function is: ； in, is the adjusted cost-sensitive loss function, is the true label of the i-th sample, is the predicted probability after cost-sensitive adjustment, The formula is: ； ； ； in, is the weighted average cost; is the logarithmic adjustment parameter of the cost ratio; is the probability of predicting the i-th sample as a positive class; C 01 is the cost of misclassifying a negative sample as a positive sample; C 10 The cost of misclassifying a positive sample as a negative sample; According to the imbalance degree of the sample The misclassification cost parameter is set, and the degree of sample imbalance The calculation formula is: ； in, is the number of positive samples, is the number of negative samples; set Value is Value times; Based on the Catboost algorithm, the adjusted cost-sensitive function is used as the loss function, and the loss function of the model is minimized by gradient boosting, thereby optimizing the performance of the vacuum dry pump fault prediction model, as follows: First, the cost-sensitive objective function of the Catboost algorithm is constructed, and its expression is: ； in, represents the tth decision tree generated during the training process, represents the predicted value of the tth decision tree, and N is the total number of samples; Then, the gradient loss function is used to calculate the residual between the predicted value and the actual value of the current model, that is, the rate of change of the model output. The expression of the gradient loss function adjusted by the cost-sensitive loss function is: ； in, Represents the gradient loss function of the model at the tth decision tree; , , is the Sigmoid activation function; represents the first-order partial derivative; Next, the second-order gradient Hessian function is used to measure the curvature of the loss function to correct the error of the fitting model. The expression of the second-order gradient Hessian function adjusted by the cost-sensitive loss function is: ； in, Represents the second-order gradient Hessian function of the t-th decision tree; represents the second-order partial derivative; Finally, update the leaf node value of the model, the leaf node value The update formula of is affected by cost sensitivity: ； in Represents the samples contained in the jth leaf node of the decision tree; The model update rule based on cost-sensitive adjustment is obtained as follows: ； in, is the learning rate, which is used to control the step size of the model update in each iteration; Represents the t-th decision tree for the training sample The predicted output.

2. A vacuum dry pump fault prediction method based on CMS data cost-sensitive learning according to claim 1, characterized in that ,The CMS monitoring data is the time series data of the status ,information collected by the CMS every second when the vacuum dry pump is running, including the ,operating time, lower pump power, upper pump power, lower pump speed and lower pump temperature.

3. A vacuum dry pump fault prediction method based on CMS data cost-sensitive learning according to claim 2, characterized in that The filtering data includes the filtering data of the downtime pump and the filtering data of the non-downtime pump; the downtime pump is a vacuum dry pump that has been downtime before the specified time, and the data filtering operation is specifically as follows: Delete the abnormal data, and then use the valid data of the preset time period before the vacant moment to fill it; if the vacant moment is the starting moment of the data, then discard this moment and use the valid data of the next moment as the starting moment; wherein, the abnormal data is that within the preset unit time length, the maximum values ​​of the upper pump power and the lower pump power of the vacuum dry pump exceed the threshold value; Adjust the warning label of downtime pumps whose remaining life time is less than the threshold.

4. A vacuum dry pump fault prediction method based on CMS data cost-sensitive learning according to claim 3, characterized in that ,When combining the fault sample oversampling strategy, for the data in each window, if the data at the starting position in the current window is a fault sample, the sliding step of the next window is adjusted to the second sliding step, and the starting position data of the next window is further judged to be a fault sample. If not, the original sliding step is restored until all windows are traversed. The expression is as follows: ； in, is the starting position of the next window, is the starting position of the current window, for Warning labels at all times, is the step size function, defined as: ； in, is the sliding step length, is the second sliding step length, > ; represents a normal sample; represents a fault sample; Each sample data in the dataset contains multiple features and a warning label, expressed as: ； in, is the feature vector at the tth time point, express The corresponding warning label, N is the total number of samples.

5. A vacuum dry pump fault prediction method based on CMS data cost-sensitive learning according to claim 4, characterized in that ,The specific steps of normalizing the sampled data in units of windows are: For each feature j in each window, calculate its data in the window The mean and standard deviation , and the calculation formulas are: ； ； is a w×c matrix, expressed as: ； in, , n is the total number of windows of sampled data, w is the window length, and c represents the number of features; is the value of the jth feature at the tth time point; is the value of the cth feature of the ith window; Then, each eigenvalue in the window Normalized to , the standardized formula is: ； in, is the standardized value.

6. A vacuum dry pump fault prediction method based on CMS data cost-sensitive learning according to claim 5, characterized in that ,The specific operations to determine whether the model meets the evaluation criteria are: For each sliding window, calculate the sum of the predicted values ​​within the window , the calculation formula is: ； Among them, p(t ) is the prediction result at time point t in the i-th window. If p(t ) = 1, indicating that a failure is predicted in the future; p(t ) = 0, indicating that no failure is predicted in the future; When the Q(t) of the sliding window exceeds the set threshold for the first time When, that is: ； The starting time point of the current window is regarded as the first predicted downtime, that is, ; The actual downtime of the sample that crashed in the test set is recorded as ;when The prediction is considered correct when the following conditions exist and are met: ； Where, d is the allowed time difference threshold; The number of correct predictions for: ； Where N is the total number of samples; is an indicator function, which is 1 when the condition is met and 0 otherwise; is the actual downtime of the i-th sample; is the predicted downtime of the i-th sample; If the predicted time Than the actual failure time Earlier than d, that is: , it is judged as a morning newspaper; The number of morning papers for: ； False positives refer to the time when the model predicts a failure to occur , but no failure actually occurred, that is, the actual downtime of the sample does not exist, then the number of false positives for: ； in, Indicates that no actual failure has occurred; False negatives refer to the failure that actually occurred but was not predicted by the model. If it does not exist, the number of missed reports for: ； Then, the model prediction accuracy The calculation formula is: ； in, Indicates the number of all predicted downtime events; Then, the model predicts the false alarm rate The calculation formula is: ； Then, the model predicts the false negative rate The calculation formula is: ； When the false positive rate of the model , false negative rate and accuracy The model is evaluated as qualified when it meets the preset requirements.

7. A vacuum dry pump fault prediction device based on CMS data cost-sensitive learning, characterized in that: include: An acquisition unit, used to acquire CMS monitoring data when the vacuum dry pump is running; A model application unit, used to input the CMS monitoring data into a pre-trained vacuum dry pump fault prediction model based on CMS data cost-sensitive learning, and output a prediction result; wherein the prediction result includes an early warning level and a predicted downtime of the vacuum dry pump, and the early warning level is used to indicate the time until the downtime moment; The pre-trained vacuum dry pump fault prediction model based on CMS data cost-sensitive learning is trained by the following steps: A1, based on the historical CMS monitoring data of the vacuum dry pump, preprocessing is performed to obtain preprocessed filtered data; A2, according to the filtered data, preset the sliding window length and sliding step length, perform data sampling based on the sliding window, and combine the fault sample oversampling strategy to obtain the sampling data of each fixed window length; wherein the fault sample oversampling strategy is to dynamically adjust the sliding step length of the next window according to whether the starting position data in the current window is a fault sample, so as to increase the number of fault samples; A3, taking the window as the unit, normalizes the sampled data in each window to convert the sampled data from two dimensions to one dimension and capture the local features of the data; A4, based on the Catboost algorithm, the normalized sampling data and the adjusted cost-sensitive loss function are used to train the fault prediction model until the model meets the evaluation criteria, and a trained vacuum dry pump fault prediction model based on CMS data cost-sensitive learning is obtained; wherein, during the model training process, the difference between the predicted value and the true label is optimized by the adjusted cost-sensitive loss function, and the expression of the adjusted cost-sensitive loss function is: ； in, is the adjusted cost-sensitive loss function, is the true label of the i-th sample, is the predicted probability after cost-sensitive adjustment, The formula is: ； ； ； in, is the weighted average cost; is the logarithmic adjustment parameter of the cost ratio; is the probability of predicting the i-th sample as a positive class; C 01 is the cost of misclassifying a negative sample as a positive sample; C 10 The cost of misclassifying a positive sample as a negative sample; According to the imbalance degree of the sample The misclassification cost parameter is set, and the degree of sample imbalance The calculation formula is: ； in, is the number of positive samples, is the number of negative samples; set Value is Value times; Based on the Catboost algorithm, the adjusted cost-sensitive function is used as the loss function, and the loss function of the model is minimized by gradient boosting, thereby optimizing the performance of the vacuum dry pump fault prediction model, as follows: First, the cost-sensitive objective function of the Catboost algorithm is constructed, and its expression is: ； in, represents the tth decision tree generated during the training process, represents the predicted value of the tth decision tree, and N is the total number of samples; Then, the gradient loss function is used to calculate the residual between the predicted value and the actual value of the current model, that is, the rate of change of the model output. The expression of the gradient loss function adjusted by the cost-sensitive loss function is: ； in, Represents the gradient loss function of the model at the tth decision tree; , , is the Sigmoid activation function; represents the first-order partial derivative; Next, the second-order gradient Hessian function is used to measure the curvature of the loss function to correct the error of the fitting model. The expression of the second-order gradient Hessian function adjusted by the cost-sensitive loss function is: ； in, Represents the second-order gradient Hessian function of the t-th decision tree; represents the second-order partial derivative; Finally, update the leaf node value of the model, the leaf node value The update formula of is affected by cost sensitivity: ； in Represents the samples contained in the jth leaf node of the decision tree; The model update rule based on cost-sensitive adjustment is obtained as follows: ； in, is the learning rate, which is used to control the step size of the model update in each iteration; Represents the t-th decision tree for the training sample The predicted output.

8. A vacuum dry pump fault prediction device based on CMS data cost-sensitive learning according to claim 7, characterized in that , the model application unit includes: The training data processing module is used to obtain preprocessed filtered data through preprocessing based on the acquired historical CMS monitoring data of the vacuum dry pump; A sliding window sampling module is used to preset a sliding window length and a sliding step size according to the filtered data, perform data sampling based on the sliding window, and combine the fault sample oversampling strategy to obtain sampling data of each fixed window length; wherein the fault sample oversampling strategy is to dynamically adjust the sliding step size of the next window according to whether the starting position data in the current window is a fault sample, so as to increase the number of fault samples; The normalization module is used to normalize the sampled data in each window in units of windows, so as to convert the sampled data from two dimensions to one dimension and capture the local features of the data; The model training module is used to train the fault prediction model based on the Catboost algorithm, using the normalized sampling data and the adjusted cost-sensitive loss function until the model meets the evaluation criteria, thereby obtaining a trained vacuum dry pump fault prediction model based on CMS data cost-sensitive learning.

Images