A device data anomaly prediction method

Abstract

The application discloses a kind of equipment data anomaly prediction method, applied to equipment data anomaly prediction system, the equipment data anomaly prediction system includes data acquisition module, data transmission module, data processing module and data prediction diagnosis module, the data acquisition module is used to collect equipment loss related data, the data of acquisition completion is transferred to data processing module by data transmission module, data is preprocessed in data processing module, and the data after preprocessing is transferred to data prediction module, and the data prediction module predicts data.The application predicts the data of equipment by prediction model, so as to compare with normal value, predict the anomaly of equipment data in advance, so as to overhaul equipment in advance, avoid some equipment failure;In the data processing process, the redundant information in original data is removed, the preprocessing of sensor data is completed, and the efficiency of data processing is improved.

Description

Technical Field

[0001] This invention belongs to the field of Internet of Things (IoT) technology, and in particular relates to a method for predicting device data anomalies. Background Technology

[0002] Data anomaly detection is a key research area in the data-driven Internet of Things (IoT) field. With the development of computer and sensor technologies, the field of mechanical equipment fault diagnosis has entered the "big data" era. Data analysis of actual equipment operation data can reveal system energy consumption and operational status, providing technical support for optimized control, energy conservation, and improved equipment safety.

[0003] Equipment loss monitoring data is characterized by its large volume, multidimensionality, and multiple metrics. Furthermore, the complexity and diversity of equipment and system configurations, along with inherent equipment malfunctions, make actual operating data highly susceptible to outliers, missing values, and inconsistencies. Low-quality data leads to low-quality analytical results. Moreover, when performing loss prediction, lifespan prediction, and failure prediction on equipment, the influence of outlier data makes it difficult to obtain accurate results. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of the aforementioned background technology by providing a method for predicting abnormal equipment data.

[0005] To achieve the above-mentioned objectives, the present invention employs the following technical solution:

[0006] A method for predicting equipment data anomalies is applied to an equipment data anomaly prediction system. The system includes a data acquisition module, a data transmission module, a data processing module, and a data prediction and diagnosis module. The data acquisition module collects equipment wear-related data. The collected data is transmitted to the data processing module via the data transmission module. The data processing module preprocesses the data and then transmits the preprocessed data to the data prediction module, which predicts the data. The equipment data anomaly prediction method includes:

[0007] Step S1: Select equipment wear and tear as a health factor reflecting the degradation trend of equipment. Use an adaptive double exponential model smoothing method to remove the fluctuations in wear and tear in the original data to obtain a sequence of wear and tear data with a relatively stable degradation trend.

[0008] Step S2: A compression algorithm is used to reduce the amount of data to achieve data compression; then, the CEEMDAM algorithm is used to decompose the sequence of the smoothed loss data to obtain residuals and simulated components that can reflect the main trend of degradation.

[0009] Step S3: Select the residual to form a training sample set, combine the neural network to establish the mapping relationship between the early and late stages of loss, and predict the unknown loss sequence. Obtain the loss prediction value through iterative prediction, and then calculate the equipment loss cycle.

[0010] Step S4: Predict the equipment wear cycle using the prediction model, obtain the predicted equipment wear cycle, set the error range, and compare it with the standard equipment wear cycle. When the error exceeds the error range, it is judged as data abnormality.

[0011] Preferably, step S1 is as follows: Step S11, the observation time data is compressed, collected and stored on a daily basis, and only the start time and stop time are recorded. The time information of the data collected during the period is replaced by hexadecimal serial numbers 1, 2, 3, ..., n, which represent the first day, the second day, the third day, ..., the nth day after the start time. The data collection time is represented by numerical intervals.

[0012] Step S12: Retain the data directly collected by the sensor and delete redundant values ​​in the collected data;

[0013] Step S13: Change the data representation method;

[0014] Step S14: Convert floating-point arithmetic to integer arithmetic. Separate the integer and fractional parts of each floating-point measurement value and perform calculations separately, thereby converting floating-point arithmetic into integer arithmetic, as shown in the following formula:

[0015] int = int(A)

[0016] Dec = (A - Int)·2 n

[0017] In the formula, A is the original data of the parameter, Int is the integer part of the data, Dec is the decimal part of the data, and n is the number of decimal places of the parameter;

[0018] Finally, arrange the integer parts of each parameter in the order they were transmitted, followed by the fractional parts.

[0019] Preferably, in step S2, the specific formula is as follows:

[0020] Let the original data sequence be X1, X2, ..., X... m First, the data is differentially calculated.

[0021] A0 = X0

[0022] A m =|X m -X m-1 |(m≥1)

[0023] S m =Sign(X) m -X m-1 )

[0024] When X m -X m-1 When S > 0, m =01; when X m -X m-1 When <0, S m =00;

[0025] After differentiation, the encoding is: A0, S1+A1, ..., S m +A m ;

[0026] Remove duplicate data.

[0027] The original sequence is X1, X2, ..., X m If the sequence contains a data source: X i =X i+1 =X i+2 =...=X i+m =0, where m≤15, then this data is encoded as 0Q, which means there are Q repeated 0s, where Q is a hexadecimal number;

[0028] At this point, the original data becomes X1, X2, ..., 0Q, ..., X m Then repeat the deduplication step to obtain the processed code.

[0029] Preferably, step S2 further includes the following steps:

[0030] Gaussian white noise ε0ω is added to the acquired raw signal S(t). i (t), i = 1, 2, ..., I, to obtain

[0031] S i (t)=S(t)+ε0ω i (t)

[0032] For S i (t) is decomposed to obtain the i modal components of the first stage. right The first modal component IMF1 is obtained by taking the average value.

[0033]

[0034] Calculate the first residual.

[0035] R i (t)=S i (t)-IMF1

[0036] E k (·) represents the k-th modal component obtained after processing, for the signal R1(t)+ε1E1ω i The second modal component is obtained by decomposing (t).

[0037]

[0038] For k = 1, 2, ..., n, calculate the nth residual as follows:

[0039] R k (t)=R k-1 (t)-IMF k

[0040] For signal R k (t)+ε k E k ω i (t) decomposition yields the (k+1)th modal component as

[0041]

[0042] Repeat the calculation until R is reached. k (t) cannot be decomposed, resulting in the final residual as

[0043]

[0044] In the formula, b is the total number of modal components.

[0045] Preferably, in step S3, the prediction model is as follows:

[0046] C i+1 =LSTM(C i-L+1 C i-L+2 ,...,C i-1 C i )

[0047] This prediction model is based on an LSTM artificial neural network, where C... i+1 The predicted value for the next equipment wear cycle; the smoothed data is used as the training set C. train and test set C test Test set C test The data is used as input data to perform iterative predictions in the model.

[0048] Preferably, it also includes the training of artificial neural networks, the specific process of which is as follows:

[0049] The network structure was constructed and the parameters were initialized. Multiple combinations of convolutional sizes were used in the feature extraction layer to find the optimal size. Based on network construction experience, the sizes of the max pooling and average pooling layers were chosen to be 6, with a dropout factor of 0.2. The learning rate was set to exponential decay at an initial value of 0.02 with a decay factor of 0.90. A total of 80 training iterations were performed, with 10 samples input per iteration from the training set C. train Randomly selected from the list.

[0050] Compared with the prior art, the present invention, employing the above technical solution, has the following beneficial effects:

[0051] 1. This invention provides a method for predicting equipment data anomalies. By using a prediction model to predict equipment data and comparing it with normal values, anomalies in equipment data can be predicted in advance, allowing for early maintenance of the equipment and avoiding some equipment failures.

[0052] 2. This invention provides a method for predicting abnormal equipment data. In the data processing process, redundant information in the original data is removed, and the sensor data is preprocessed. In the initial processing stage, some abnormal data is removed, thereby improving the efficiency of data processing.

[0053] 3. This invention provides a method for predicting abnormal equipment data, which effectively saves data storage space, improves the throughput of communication networks per unit time, and also helps to improve the computing speed of marine underwater mooring embedded systems and reduce system power consumption.

[0054] 4. This invention provides a method for predicting equipment data anomalies, which smooths and reduces noise in the original capacity data so that the processed data has a monotonically decreasing and stable trend. Attached Figure Description

[0055] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0056] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] A method for predicting equipment data anomalies is applied to an equipment data anomaly prediction system. The system includes a data acquisition module, a data transmission module, a data processing module, and a data prediction and diagnosis module. The data acquisition module collects equipment wear-related data. The collected data is transmitted to the data processing module via the data transmission module. The data processing module preprocesses the data and then transmits the preprocessed data to the data prediction module, which predicts the data. The equipment data anomaly prediction method includes:

[0059] Step S1: Select equipment wear and tear as a health factor reflecting the degradation trend of equipment. Use an adaptive double exponential model smoothing method to remove the fluctuations in wear and tear in the original data to obtain a sequence of wear and tear data with a relatively stable degradation trend.

[0060] Step S2: A compression algorithm is used to reduce the amount of data to achieve data compression; then, the CEEMDAM algorithm is used to decompose the sequence of the smoothed loss data to obtain residuals and simulated components that can reflect the main trend of degradation.

[0061] Step S3: Select the residual to form a training sample set, combine the neural network to establish the mapping relationship between the early and late stages of loss, and predict the unknown loss sequence. Obtain the loss prediction value through iterative prediction, and then calculate the equipment loss cycle.

[0062] Step S4: Predict the equipment wear cycle using the prediction model, obtain the predicted equipment wear cycle, set the error range, and compare it with the standard equipment wear cycle. When the error exceeds the error range, it is judged as data abnormality.

[0063] During equipment operation, wear and tear occurs in various parts. Therefore, the data collected by the data acquisition module varies at different times. When a component experiences excessive wear or malfunctions, abnormal data will be generated. The damage to a component is a process, and when the data acquisition module collects the first abnormal data, the prediction module makes a prediction. It will show a change in the wear cycle. By comparing the error range, it can be determined whether a component is damaged or about to be damaged, thus reminding the staff to inspect the equipment and avoid the probability of safety accidents or other incidents.

[0064] As a specific implementation method, step S1 is as follows: Step S11, the observation time data is compressed, collected and stored on a daily basis, and only the start time and stop time are recorded. The time information of the data collected during the period is replaced by hexadecimal serial numbers 1, 2, 3, ..., n, which represent the first day, the second day, the third day, ..., the nth day after the start time. The data collection time is represented by numerical intervals.

[0065] Step S12: Retain the data directly collected by the sensor and delete redundant values ​​in the collected data;

[0066] Step S13: Change the data representation method;

[0067] Step S14: Convert floating-point arithmetic to integer arithmetic. Separate the integer and fractional parts of each floating-point measurement value and perform calculations separately, thereby converting floating-point arithmetic into integer arithmetic, as shown in the following formula:

[0068] int = int(A)

[0069] Dec = (A - Int)·2 n

[0070] In the formula, A is the original data of the parameter, Int is the integer part of the data, Dec is the decimal part of the data, and n is the number of decimal places of the parameter;

[0071] Finally, following the transmission order of each parameter, the integer parts are arranged first, followed by the decimal parts; this can simplify the calculation process and improve calculation accuracy to a certain extent.

[0072] As a specific implementation method, the specific formula in step S2 is as follows:

[0073] Let the original data sequence be X1, X2, ..., X... m First, the data is differentially calculated.

[0074] A0 = X0

[0075] A m =|X m -X m-1 |(m≥1)

[0076] S m =Sign(X) m -X m-1 )

[0077] When X m -X m-1 When S > 0, m =01; when X m -X m-1When <0, S m =00;

[0078] After differentiation, the encoding is: A0, S1+A1, ..., S m +A m Since differential compression algorithms only involve subtraction and not multiplication or division, the calculation process is simple and the execution efficiency is high, which helps to reduce the workload and working time of embedded systems and reduce system power consumption.

[0079] Remove duplicate data.

[0080] The original sequence is X1, X2, ..., X m If the sequence contains a data source: X i =X i+1 =X i+2 =...=X i+m =0, where m≤15, then this data is encoded as 0Q, which means there are Q repeated 0s, where Q is a hexadecimal number;

[0081] At this point, the original data becomes X1, X2, ..., 0Q, ..., X m The deduplication process is then repeated to obtain the processed code. These data processing methods can filter out outlier data, thereby avoiding its impact on the prediction model and improving its accuracy.

[0082] In one specific implementation, step S2 further includes the following steps:

[0083] Gaussian white noise ε0ω is added to the acquired raw signal S(t). i (t), i = 1, 2, ..., I, to obtain

[0084] S i (t)=S(t)+ε0ω i (t)

[0085] For S i (t) is decomposed to obtain the i modal components of the first stage. right The first modal component IMF1 is obtained by taking the average value.

[0086]

[0087] Calculate the first residual.

[0088] R i (t)=S i (t)-IMF1

[0089] E k(·) represents the k-th modal component obtained after processing, for the signal R1(t)+ε1E1ω i The second modal component is obtained by decomposing (t).

[0090]

[0091] For k = 1, 2, ..., n, calculate the nth residual as follows:

[0092] R k (t)=R k-1 (t)-IMF k

[0093] For signal R k (t)+ε k E k ω i (t) decomposition yields the (k+1)th modal component as

[0094]

[0095] Repeat the calculation until R is reached. k (t) cannot be decomposed, resulting in the final residual as

[0096]

[0097] In the formula, b is the total number of modal components.

[0098] As a specific implementation method, in step S3, the prediction model is as follows:

[0099] C i+1 =LSTM(C i-L+1 C i-L+2 ,...,C i-1 C i )

[0100] This prediction model is based on an LSTM artificial neural network, where C... i+1 The predicted value for the next equipment wear cycle; the smoothed data is used as the training set C. train and test set C test Test set C test The data is used as input data to perform iterative predictions in the model.

[0101] As a specific implementation method, it also includes the training of artificial neural networks, the specific process of which is as follows:

[0102] The network structure was constructed and the parameters were initialized. Multiple combinations of convolutional sizes were used in the feature extraction layer to find the optimal size. Based on network construction experience, the sizes of the max pooling and average pooling layers were chosen to be 6, with a dropout factor of 0.2. The learning rate was set to exponential decay at an initial value of 0.02 with a decay factor of 0.90. A total of 80 training iterations were performed, with 10 samples input per iteration from the training set C. train The network is trained using randomly selected samples from the input training set. After multiple backpropagation iterations, the network selects the optimal parameters, outputs the fault type and diagnostic accuracy, and uses T-SNE to visualize the network's internal feature learning performance.

[0103] It also includes error data training, which involves using abnormal data from some devices as a training set to input the prediction model, obtaining prediction results, checking for consistency, and thus determining the accuracy of the model's predictions.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting equipment data anomalies, applied to an equipment data anomaly prediction system, characterized in that, The equipment data anomaly prediction system includes a data acquisition module, a data transmission module, a data processing module, and a data prediction and diagnosis module. The data acquisition module is used to collect equipment loss-related data. The collected data is transmitted to the data processing module through the data transmission module. The data processing module preprocesses the data and transmits the preprocessed data to the data prediction module. The data prediction module makes predictions based on the data. The device data anomaly prediction method includes: Step S1: Select equipment wear and tear as a health factor reflecting the degradation trend of equipment. Use an adaptive double exponential model smoothing method to remove the fluctuations in wear and tear in the original data to obtain a sequence of wear and tear data with a relatively stable degradation trend. The specific steps of S1 are as follows: Step S11, the observation time data is compressed, collected and stored on a daily basis, and only the start time and stop time are recorded. The time information of the data collected during the period is replaced by hexadecimal serial numbers 1, 2, 3, ..., n, which represent the first day, the second day, the third day, ..., the nth day after the start time. The data collection time is represented by numerical intervals. Step S12: Retain the data directly collected by the sensor and delete redundant values ​​in the collected data; Step S13: Change the data representation method; Step S14: Convert floating-point arithmetic to integer arithmetic. Separate the integer and fractional parts of each floating-point measurement value and perform calculations separately, thereby converting floating-point arithmetic into integer arithmetic, as shown in the following formula: int = int(A) Dec=(A-Int)·2 n In the formula, A is the original data of the parameter, Int is the integer part of the data, Dec is the decimal part of the data, and n is the number of decimal places of the parameter; Finally, arrange the integer parts first, then the decimal parts of each parameter, according to the transmission order. Step S2: A compression algorithm is used to reduce the amount of data to achieve data compression; then, the CEEMDAM algorithm is used to decompose the sequence of the smoothed loss data to obtain residuals and simulated components that can reflect the main trend of degradation. In step S2, the specific formula is as follows: Let the original data sequence be X1, X2, ..., X... m First, the data is differentially calculated. A0 = X0 A m =|X m -X m-1 |(m≥1) S m =Sign(X m -X m-1 ) When X m -X m-1 When S > 0, m =01; when X m -X m-1 When <0, S m =00; After differentiation, the encoding is: A0, S1+A1, ..., S m +A m ; Remove duplicate data. The original sequence is X1, X2, ..., X m If the sequence contains a data source: X i =X i+1 =X i+2 =...=X i+m =0, where m≤15, then this data is encoded as 0Q, which means there are Q repeated 0s, where Q is a hexadecimal number; At this point, the original data becomes X1, X2, ..., 0Q, ..., X m Repeat the deduplication step to obtain the processed code; Step S3: Select the residual to form a training sample set, combine the neural network to establish the mapping relationship between the early and late stages of loss, and predict the unknown loss sequence. Obtain the loss prediction value through iterative prediction, and then calculate the equipment loss cycle. Step S4: Predict the equipment wear cycle using the prediction model, obtain the predicted equipment wear cycle, set the error range, and compare it with the standard equipment wear cycle. When the error exceeds the error range, it is judged as data abnormality.

2. The method for predicting equipment data anomalies as described in claim 1, characterized in that, Step S2 further includes the following steps: Gaussian white noise ε0w is added to the acquired raw signal S(t). i (t), i = 1, 2, ..., I, we obtain S i (t)=S(t)+ε0ω i (t) For S i (t) is decomposed to obtain the i modal components of the first stage. right The first modal component IMF1 is obtained by taking the average value. Calculate the first residual. R i (t)=S i (t)-IMF1 E K (·) represents the k-th modal component obtained after processing, for the signal R1(t)+ε1E1ω i The second modal component is obtained by decomposing (t). For k = 1, 2, ..., n, calculate the nth residual as follows: R k (t)=R k-1 (t)-IMF k For signal R k (t)+ε k E k ω i (t) decomposition yields the (k+1)th modal component as Repeat the calculation until R is reached. k (t) cannot be decomposed, resulting in the final residual as In the formula, b is the total number of modal components.

3. The method for predicting equipment data anomalies as described in claim 1, characterized in that, In step S3, the prediction model is as follows: C i+1 =LSTM(C i-L+1 ,C i-L+2 ,···,C i-1 ,C i ) This prediction model is based on an LSTM artificial neural network, where C... i+1 The predicted value for the next equipment wear cycle; the smoothed data is used as the training set C. train and test set C test Test set C test The data is used as input data to perform iterative predictions in the model.

4. The method for predicting equipment data anomalies as described in claim 3, characterized in that, It also includes the training of artificial neural networks, the specific process of which is as follows: The network structure was constructed and the parameters were initialized. Multiple combinations of convolutional sizes were used in the feature extraction layer to find the optimal size. Based on network construction experience, the sizes of the max pooling and average pooling layers were chosen to be 6, with a dropout factor of 0.

2. The learning rate was set to exponential decay at an initial value of 0.02 with a decay factor of 0.

90. A total of 80 training iterations were performed, with 10 samples input per iteration from the training set C. train Randomly selected from the list.

Images