Fault prediction method and device based on GANs and LSTM

Abstract

The invention discloses a fault prediction method and device based on GANs and LSTM. The prediction method comprises the following steps: S1, obtaining historical operation data of target equipment; s2, constructing a data set by using the historical operation data, and training a GANs network to obtain a data generation model; s3, generating simulation data by using the data generation model, and merging the simulation data with the data set to obtain an expanded data set; s4, training an LSTM network by using the extended data set to obtain a fault prediction model; and S5, performing fault prediction on the target equipment by using the fault prediction model. According to the invention, the problem of data shortage in the fault prediction process can be overcome.

Description

Technical Field

[0001] The present invention relates to the technical field related to fault prediction, and more specifically, to a fault prediction method and device based on GANs and LSTM. Background Art

[0002] In modern industry and daily life, the widespread use of various complex equipment has greatly improved production efficiency and people's quality of life. However, these devices will inevitably fail during long-term operation. If the failure is not discovered and handled in time, it may not only cause damage to the equipment itself, but also cause serious consequences such as production interruption, service stagnation and economic losses. Therefore, accurate and efficient fault prediction has become a key link to ensure the reliable operation of equipment.

[0003] Traditional fault prediction methods are mainly based on empirical rules and simple statistical analysis; for example, maintenance personnel summarize the typical characteristics of equipment before certain faults occur based on past maintenance cases and experience knowledge, and then conduct regular manual inspections to see if there are any abnormalities in the relevant parameters of the equipment to determine whether a fault is about to occur. In addition, some statistical methods perform simple statistical analysis such as mean and variance on the equipment operation data, set fixed thresholds, and determine that a fault may occur when the data exceeds the threshold range. Obviously, these traditional methods have many limitations, such as being highly dependent on the maintenance personnel's professional knowledge and experience accumulation, being unable to adapt to the dynamic changes in the equipment's operating data under different working conditions and environments, and being prone to false alarms or missed alarms.

[0004] With the rapid development of big data technology and artificial intelligence, using deep learning methods to analyze equipment operation data and then predict faults has become a research hotspot. However, in practical applications, deep learning methods for fault prediction face a key problem - data scarcity. Although normal operation data is relatively easy to obtain, the general type is relatively single, while fault data is very scarce, which makes it difficult for the model to fully learn the corresponding features of the fault, resulting in insufficient generalization ability and limited prediction accuracy when facing actual complex working conditions.

[0005] Therefore, it is necessary to design a technical solution that can overcome the above-mentioned defects. Summary of the invention

[0006] An object of the present invention is to provide a fault prediction method and device based on GANs and LSTM, which can overcome the problem of data scarcity in the fault prediction process to a certain extent.

[0007] In order to achieve these objectives and other advantages of the present invention, according to one aspect of the present invention, the present invention provides a fault prediction method based on GANs and LSTM, including: S1: obtaining historical operation data of the target device; S2: using the historical operation data to construct a data set, training the GANs network, and obtaining a data generation model; S3: using the data generation model to generate simulation data, merging the simulation data with the data set to obtain an expanded data set; S4: using the expanded data set to train the LSTM network to obtain a fault prediction model; S5: using the fault prediction model to predict faults for the target device.

[0008] Furthermore, in S2, normal data and abnormal data in the historical operation data are extracted; the normal data and the abnormal data are clustered respectively to obtain classification clusters; when training the GANS network, different feature parameters are assigned to the normal data or the abnormal data belonging to different classification clusters.

[0009] Furthermore, the clustering method includes: dividing the normal data and the abnormal data into multiple subsets respectively, each subset containing at least one type of operation data; clustering each subset to obtain multiple sub-classification clusters; assigning a weight to each subset according to the contribution of each subset to the operation of the target device; determining the score of each normal data or abnormal data on each sub-classification cluster according to the weight of each subset, and selecting the sub-classification cluster with the highest score as the classification cluster for each normal data or abnormal data.

[0010] Furthermore, each of the subsets is clustered using a K-Means algorithm.

[0011] Furthermore, the characteristic parameter is the mean square error of each type of operation data included in the sub-classification cluster.

[0012] Furthermore, S4 includes: training a labeling model, adding labels to the simulation data, and then merging the labels with the data set to obtain an expanded data set; using the expanded data set to train the LSTM network to obtain the fault prediction model, the LSTM network includes two LSTM layers and one fully connected layer.

[0013] Furthermore, the labeled data set is used to train a random forest model and a support vector machine model to obtain a first labeled model and a second labeled model, and the label of the simulation data is determined according to the first labeled model and the second labeled model.

[0014] According to another aspect of the present invention, a fault prediction device based on GANs and LSTM is also provided, including a processor and a memory, wherein the memory is used to store program instructions, and the processor is used to call the program instructions to execute the fault prediction method based on GANs and LSTM.

[0015] According to another aspect of the present invention, a computer-readable storage medium is also provided, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the fault prediction method based on GANs and LSTM is implemented.

[0016] The present invention has at least the following beneficial effects: The present invention uses historical operation data to construct a data set, trains a GANs network, obtains a data generation model, uses the data generation model to generate simulation data, merges the simulation data with the data set to obtain an expanded data set, uses the expanded data set to train an LSTM network, obtains a fault prediction model, and performs fault prediction; the present invention uses a data generation model to generate simulation data for training a fault prediction model, overcomes the problem of data scarcity, and improves generalization ability and prediction accuracy.

[0017] Other advantages, objectives and features of the present invention will be embodied in part through the following description, and in part will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of an embodiment of the present application. DETAILED DESCRIPTION

[0019] The present invention is further described in detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0020] It should be understood that the terms such as "having", "including" and "comprising" used in the embodiments of the present application do not exclude the existence or addition of one or more other elements or their combinations. All directional indications (such as up, down, left, right, front, back...) in the embodiments of the present application are only used to explain the relative position relationship, movement, etc. between the components in a certain specific posture. If the specific posture changes, the directional indication will also change accordingly. When an element is referred to as "fixed on" or "set on" another element, it can be directly on the other element or there may be a centering element at the same time. When an element is referred to as "connecting" another element, it can be directly connected to another element or it can be indirectly connected to another element through a centering element. The description of "first", "second", etc. in the embodiments of the present application is only for descriptive purposes, and cannot be understood as indicating or implying its relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" can explicitly or implicitly include at least one of the features.

[0021] It should be noted that the technical solutions between the various embodiments of the present application can be combined with each other, but it must be based on the fact that ordinary technicians in the field can implement it. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by this application.

[0022] like Figure 1 As shown, the embodiment of the present application provides a fault prediction method based on GANs and LSTM, including: S1: Obtain historical operation data of the target device; Exemplarily, the target device refers to air conditioning equipment, ground source heat pump, fresh air equipment, etc. The following mainly takes air conditioning equipment as an example for explanation. Given that those skilled in the art have corresponding professional knowledge and experience, they can easily extrapolate the relevant principles and methods to ground source heat pump and fresh air equipment. Exemplarily, historical operation data is collected from various sensors of the target device, such as air conditioners, such as temperature sensors (such as indoor evaporator temperature, air outlet temperature, outdoor condenser temperature, compressor temperature, etc.), pressure sensors (high and low pressure side pressures of the refrigeration system), current sensors (compressor, indoor and outdoor fan motor currents), and wind speed sensors (indoor and outdoor fan outlet wind speeds), etc.; further, since the magnitude of data collected by different sensors varies greatly, the Min-Max normalization method is used to unify the data features into the 0-1 interval, so that data of different dimensions can participate in subsequent model training on the same scale, which is convenient for the model to better learn data features and perform operations such as parameter updates; S2: Use the historical operation data to construct a data set, train the GANs network, and obtain a data generation model; Exemplarily, the GANs network includes a generator and a discriminator. The generator includes an input layer, three hidden layers and an output layer. The input layer is used to receive a 100-dimensional random noise vector. Each element of the noise vector is randomly sampled from a standard normal distribution. The first hidden layer has 512 neurons, and a ReLU activation function is used for nonlinear transformation to perform feature mapping and expansion on the input noise vector. The second hidden layer has 256 neurons, and the ReLU activation function is also used to further extract and integrate features. The third hidden layer has 128 neurons, and feature transformation is continued. The number of neurons in the output layer is the same as the dimension of the running data (e.g., 7 dimensions), and the final output is the generated simulation data. The discriminator includes an input layer, two hidden layers and an output layer. The data dimension directly received by the input layer is the same as the dimension of the real data and the simulated data output by the generator, that is, 7 dimensions, covering the various air-conditioning operation-related features mentioned above. The first hidden layer is set with 256 neurons, and the LeakyReLU activation function is used here. When the discriminator faces complex tasks of distinguishing true and false data, it can capture data features more delicately and enhance its nonlinear expression ability, so as to better judge the authenticity of the input data; the second hidden layer is equipped with 128 neurons, and the LeakyReLU activation function is also used. , further extract the feature information in the data, and analyze the correlation and changes between different features to comprehensively judge whether the input data is the real air-conditioning operation data or the simulated data generated by the generator, so as to continuously improve the discriminator's accuracy in distinguishing true and false data; the output layer has only one neuron, and the Sigmoid activation function is used to compress the output result to the [0,1] interval. When the output value is close to 1, it means that the discriminator believes that the input data is real historical operation data, and when it is close to 0, it means that the discriminator determines that it is simulated data generated by the generator, so as to achieve quantitative judgment of the true and false of the input data; During specific training, a batch of real data is randomly extracted from the preprocessed real historical running data set and input into the discriminator. At this time, it is hoped that the discriminator can accurately judge that these real data are true, that is, the probability output by the discriminator indicating that the data is real data (the value output by the Sigmoid function in the interval [0,1]) is as close to 1 as possible; At the same time, a batch of noise vectors are randomly generated from the standard normal distribution, and these noise vectors are input into the generator to generate a batch of simulated data, and then these simulated data are also input into the discriminator. At this time, it is expected that the discriminator can accurately identify that these simulated data are false, that is, the probability of the output being judged as real data is as close to 0 as possible; The discriminator's judgment results on real data and simulated data are compared with the corresponding real labels. The accuracy of the discriminator's judgment is measured by calculating the binary cross entropy loss function to obtain a total loss value, which comprehensively reflects the discriminator's performance in distinguishing true and false data. Then, based on this loss value, the gradient of each parameter of the discriminator is calculated through the back propagation algorithm, and then the selected Adam optimizer is used to update the parameters of the discriminator according to the calculated gradient direction and size, so that its ability to distinguish true and false data in the future can be continuously enhanced. After completing the training of the discriminator, the discriminator is fixed and the generator is trained. A batch of noise vectors are generated from the standard normal distribution again and input into the generator to generate simulated data. These simulated data are input into the discriminator that has been trained for one round (with fixed parameters). At this time, for the generator, the greater the probability that the discriminator judges that these simulated data are real data, the better. In other words, the more the generated simulated data can deceive the discriminator, the closer the generated data is to the real data. According to the judgment result of the discriminator on the simulated data (the output probability value), the gradient of the generator parameters is calculated through the back propagation algorithm, and then the Adam optimizer is used to update the parameters of the generator, so that the generator can generate simulated data with higher quality and more difficult to distinguish by the discriminator. This is repeated. In each round of training, the generator and the discriminator alternately carry out such a training process. Through continuous confrontation and game, the performance of the two will gradually improve. After 200 rounds of training, the generator can finally generate simulated data with a certain quality and a distribution similar to the real historical running data. S3: Generate simulated data using the data generation model, and merge the simulated data with the data set to obtain an expanded data set; In this step, the simulated data generation method mentioned in S2 is used to generate simulated data, which is then merged with the original historical operation data set to construct an expanded data set. When merging, the consistency of the data format must be ensured so that the expanded data set can still be easily processed and analyzed later; S4: Using the expanded data set to train the LSTM network to obtain a fault prediction model; Exemplarily, the LSTM network includes an input layer, a hidden layer, and an output layer. Ten key features such as evaporator temperature and condenser temperature are selected from the expanded data set as input. Considering the data of the past 30 time points, two hidden layers are set. The first layer uses 128 LSTM units, and the activation function selects the hyperbolic tangent function (tanh) and outputs the hidden state of each time step. The second layer has 64 LSTM units, and the tanh activation function is also used. Only the hidden state of the last time step is output. The output layer can be based on the four situations of normal operation, refrigerant leakage failure, compressor failure, and fan failure, so 4 neurons are set. The output is converted into probability distribution of each category using the Softmax activation function. The model is built and the adam optimizer is specified. During training, the training set data is input into the network according to the batch size, forward propagation, loss calculation, and back propagation to update parameters. After each round, the performance is evaluated on the validation set, and the parameters are adjusted according to the situation. After 100 rounds of training, the fault prediction model is obtained; S5: Use the fault prediction model to predict the fault of the target device, that is, collect the current operation data of the target device in real time, take the data of the past 30 time points as input, input the fault prediction model, and obtain the probability values ​​of four situations: normal operation, refrigerant leakage failure, compressor failure, and fan failure. If the fault probability value exceeds the threshold, an alarm signal is issued, and an email or SMS notification is sent to the operation and maintenance personnel to facilitate timely maintenance and processing; It can be seen that this embodiment uses a data generation model to generate simulated data for training a fault prediction model, which overcomes the problem of data scarcity and improves generalization ability and prediction accuracy.

[0023] In another embodiment, in S2, normal data and abnormal data in the historical operation data are extracted; the normal data and the abnormal data are clustered respectively to obtain classification clusters; when training the GANS network, different feature parameters are respectively assigned to the normal data or the abnormal data belonging to different classification clusters; Exemplarily, the distinction between normal data and abnormal data is determined according to a preset threshold value, and the K-Means clustering algorithm is used to cluster normal data. The best number of clusters is determined to be 3 classification clusters through the elbow method. The data characteristics in the first cluster are low wind speed and low load operation, the evaporator temperature is close to 5°C, the condenser temperature is about 30°C, and the compressor current is small; the second cluster is normal wind speed and medium load operation, the evaporator temperature is about 10°C, the condenser temperature is around 40°C, and the compressor current is moderate; the third cluster is high wind speed and high load operation, the evaporator temperature is about 15°C, the condenser temperature is close to 50°C, and the compressor current is large; the abnormal data clustering also uses the K-Means clustering algorithm to determine 2 classification clusters, one cluster represents refrigerant leakage fault-related anomalies, and the data characteristics are that the evaporator temperature continues to rise, the condenser temperature drops, the compressor current gradually decreases, etc., and the other cluster corresponds to the fan failure anomaly, which is manifested in unstable fan speed and occasional instantaneous fluctuations in compressor current, etc. Different characteristic parameters are assigned to the first, second, and third clusters in the normal data, so that the generator can generate data that fits the corresponding normal working conditions based on different initial conditions. Different characteristic parameters are assigned to the first and second clusters of the abnormal data to guide the generation of data that simulates different fault characteristics. The operation of classifying and clustering historical operation data and assigning feature parameters makes the use of data more targeted. Instead of using all historical data for GANs network training in general, the model learning is guided according to the characteristics of different operating states and failure modes, avoiding information redundancy and mutual interference in the data, and fully tapping the value contained in each type of data. This is in line with the situation in the fault prediction process where the amount of data is relatively limited but comprehensive coverage of multiple faults is required.

[0024] In another embodiment, the clustering method includes: dividing the normal data and the abnormal data into a plurality of subsets, each of which contains at least one type of operation data; clustering each of the subsets to obtain a plurality of sub-classification clusters; assigning a weight to each of the subsets according to the contribution of each of the subsets to the operation of the target device; determining a score of each of the normal data or the abnormal data on each of the sub-classification clusters according to the weight of each of the subsets, and selecting the sub-classification cluster with the highest score as the classification cluster of each of the normal data or the abnormal data; Exemplarily, clustering is performed on each of the subsets using a K-Means algorithm; Exemplarily, the characteristic parameter is the mean square error of each type of operation data included in the sub-classification cluster; Exemplarily, the normal data is divided into three subsets, the first subset includes the evaporator temperature and the condenser temperature, two operating data related to heat exchange, the second subset includes the compressor current, and the third subset includes the fan speed data; similarly, the abnormal data is divided into two subsets, the first abnormal data subset includes the evaporator temperature and the compressor current, and the second abnormal data subset covers the fan speed data; For normal data, the weight of subset 1 is set to 0.4, because the speed and temperature have a significant impact on the stable operation and heat dissipation of the motor; the weight of subset 2 is 0.35, and the current and operating time are related to the load and overall usage; the weight of subset 3 is 0.25, mainly focusing on the wind speed of the fan; for abnormal data, the weight of subset 1 is set to 0.6, and current and temperature abnormalities are often key fault manifestations; the weight of subset 2 is 0.4, reflecting the wind speed of the fan; The five subsets are clustered using the K-Means algorithm. The cluster centers (corresponding to the running data) selected for each clustering are the same, and multiple sub-classification clusters are obtained. For a piece of normal operation data X, determine the classification cluster to which it belongs in each subset clustering process. For example, if it belongs to cluster A in both subsets 1 and 2, and to cluster B in subset 2, but does not belong to cluster C, then the score of cluster A is 0.75, the score of cluster B is 0.25, and the score of cluster C is 0. At this time, it is determined that the operation data belongs to cluster A. For an abnormal operation data Y, it belongs to cluster C in subset 1 and cluster D in subset 2, then the score of cluster C is 0.6, and the score of cluster D is 0.4. At this time, it is also determined that the operation data belongs to cluster C. Repeat the above operation to determine the classification cluster for each operation data, and then use the mean square error of all operation data in the corresponding classification cluster as the feature parameter. When using the GANs network for subsequent training, it can guide the generator to learn to generate data that meets the characteristics of this type of normal operation status, so that the generated data is more in line with the actual situation.

[0025] In another embodiment, in order to improve the accuracy and efficiency of labeling, it further includes: training a labeling model and adding labels to the simulation data, and then merging it with the data set to obtain an expanded data set; using the expanded data set to train the LSTM network to obtain the fault prediction model, the LSTM network includes two LSTM layers and one fully connected layer.

[0026] Exemplarily, a random forest model and a support vector machine model are trained using the labeled data set to obtain a first labeled model and a second labeled model, and labels of the simulated data are determined according to the first labeled model and the second labeled model; Specifically, the existing historical operation data set is first divided into a training set and a test set, for example, at a ratio of 70% and 30%. The training set is used to train the annotation model, and the test set is used to evaluate the model effect. This historical operation data set contains various operation data when the air conditioner is operating normally and when different fault conditions occur, such as refrigerant leakage, compressor failure, and fan failure. Each piece of data has a corresponding accurate fault label (for example, 0 represents normal operation, 1 represents refrigerant leakage failure, 2 represents compressor failure, and 3 represents fan failure). The data features in the training set are used as input, and the corresponding fault labels are used as output labels. The random forest algorithm is used for training. Similarly, the running data features in the training set are used as input, and the fault labels are used as output labels. The support vector machine algorithm is used for training. Determine the labels of the simulated data A batch of simulated data is extracted from the simulated data set previously generated by the data generation model. These simulated data contain various parameter values ​​of the simulated air-conditioning equipment under different operating conditions, but initially there is no fault label.

[0027] Input each piece of simulation data into the trained first labeling model (random forest model) and the second labeling model (support vector machine model) respectively, and obtain the fault prediction results (that is, the corresponding label prediction values) given by the two models; if the prediction results of the two models are consistent, then the result is directly used as the label of the simulation data; if they are inconsistent, manual intervention is performed; using this method to determine the label for each simulation data, the random forest may have a better grasp of the overall trend, but may not be sensitive enough to some subtle feature differences, while the support vector machine may have more advantages in capturing local features. The combination of the two can more comprehensively consider the characteristics of the data, thereby improving the accuracy of labeling; The following table gives a comparison of fault prediction models obtained with different data and the same model.

[0028] An embodiment of the present application also provides a fault prediction device based on GANs and LSTM, including a processor and a memory, wherein the memory is used to store program instructions, and the processor is used to call the program instructions to execute the fault prediction method based on GANs and LSTM; the device of this embodiment can be a mobile phone, a laptop computer, a tablet computer, etc., which has a memory and a processor arranged therein to execute the fault prediction method of the above embodiment.

[0029] An embodiment of the present application also provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, it implements the fault prediction method based on GANs and LSTM; in this embodiment, the computer-readable medium can be a portable computer disk, a hard disk, a random access memory, a read-only memory, an erasable programmable read-only memory, an optical storage device, a magnetic storage device, etc., which can store a computer program to execute the fault prediction method of the above embodiment.

[0030] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and the implementation modes, and they can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and the illustrations shown and described herein.

Claims

1. The fault prediction method based on GANs and LSTM is characterized by: include: S1: Obtain historical operation data of the target device; S2: Use the historical operation data to construct a data set, train the GANs network, and obtain a data generation model; S3: Generate simulated data using the data generation model, and merge the simulated data with the data set to obtain an expanded data set; S4: Using the expanded data set to train the LSTM network to obtain a fault prediction model; S5: Use the fault prediction model to perform fault prediction on the target device.

2. The fault prediction method based on GANs and LSTM as claimed in claim 1, characterized in that: In S2, normal data and abnormal data in the historical operation data are extracted; Clustering the normal data and the abnormal data respectively to obtain classification clusters; When training the GANS network, different feature parameters are respectively assigned to the normal data or the abnormal data belonging to different classification clusters.

3. The fault prediction method based on GANs and LSTM as claimed in claim 2, characterized in that: Clustering methods include: Dividing the normal data and the abnormal data into a plurality of subsets respectively, each of the subsets comprising at least one type of operation data; Clustering each of the subsets separately to obtain multiple sub-classification clusters; Assigning a weight to each of the subsets according to the contribution of each of the subsets to the operation of the target device; According to the weight of each of the subsets, the score of each piece of the normal data or the abnormal data on each of the sub-classification clusters is determined, and the sub-classification cluster with the highest score is selected as the classification cluster for each piece of the normal data or the abnormal data.

4. The fault prediction method based on GANs and LSTM as claimed in claim 3, characterized in that: Each of the subsets is clustered using the K-Means algorithm.

5. The fault prediction method based on GANs and LSTM as claimed in claim 3, characterized in that: The characteristic parameter is the mean square error of each type of operation data included in the sub-classification cluster.

6. The fault prediction method based on GANs and LSTM as claimed in claim 1, characterized in that: The S4 includes: Training the annotation model, adding labels to the simulated data, and then merging the simulated data with the data set to obtain an expanded data set; The LSTM network is trained using the expanded data set to obtain the fault prediction model, wherein the LSTM network includes two LSTM layers and one fully connected layer.

7. The fault prediction method based on GANs and LSTM as claimed in claim 6, characterized in that: The labeled data set is used to train a random forest model and a support vector machine model to obtain a first labeled model and a second labeled model, and the label of the simulation data is determined according to the first labeled model and the second labeled model.

8. A fault prediction device based on GANs and LSTM, characterized in that: It includes a processor and a memory, wherein the memory is used to store program instructions, and the processor is used to call the program instructions to execute the fault prediction method based on GANs and LSTM as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the fault prediction method based on GANs and LSTM according to any one of claims 1 to 7 is implemented.

Images