CN115204272A - Industrial system fault diagnosis method and equipment based on multi-sampling rate data - Google Patents

Abstract

The invention discloses an industrial system fault diagnosis method and equipment based on multi-sampling rate data, wherein the method comprises the following steps: acquiring time sequence data of a plurality of groups of different sensors of an industrial system, wherein the plurality of groups of data have different sampling rates; for the time-series data of all the sensors of each sampling rate, screening and retaining the time-series data of partial sensors according to the correlation; the time sequence data of the sensors are screened and reserved for each sampling rate, normalization processing is carried out respectively, and then splicing is carried out according to the sensors; the method comprises the steps that spliced data corresponding to various sampling rates are combined with corresponding real classification labels, and a CNN model is trained in an end-to-end mode to obtain a fault classifier; and when fault diagnosis is needed, performing fault diagnosis on the industrial system based on the multi-sampling-rate data by using the trained fault classifier. The invention can automatically mine the deep characteristics of the multi-sampling rate data and effectively utilize the information of different sampling rate variables to improve the precision of fault diagnosis.

Description

Industrial system fault diagnosis method and equipment based on multi-sampling rate data

Technical Field

The invention belongs to the field of industrial system fault diagnosis, and particularly relates to an industrial system fault diagnosis method and equipment based on multi-sampling rate data.

Background

Industry is vital to the development of national economy. Due to the increasingly complex and large scale construction of industrial systems, failure in the event of a failure can adversely affect production and personnel safety. Thus, the safety and reliability of industrial systems are of great concern. The system is accurately diagnosed, and a maintenance strategy is adopted in time, so that major safety accidents can be avoided, the maintenance cost is reduced, and the production efficiency is improved. The development of computer, sensor and storage technologies has brought new opportunities for data-driven based fault diagnosis methods. The method trains an offline fault diagnosis model through data acquired by a sensor, and makes diagnosis decision based on real-time data.

Multi-sensor information fusion is a developing trend in the field of fault diagnosis. Compared with a single sensor, the multiple sensors can provide more comprehensive and abundant system information, so that the reliability of diagnosis is improved. Multi-sensor information fusion can be performed on different levels: a data level, a feature level, and a decision level. However, the sampling rates of different sensors in an actual industrial system are often not uniform. On the one hand, the sampling rate of a sensor depends to some extent on the physical quantity it monitors. In particular, for some slowly varying process variables, such as pressure, temperature and flow, their sampling rate is mostly on the order of minutes; for some rapidly changing variables, such as vibration, current, and acoustic signals, their sampling rate is typically on the order of seconds. On the other hand, some components of the process industry need to be collected manually and sent to a laboratory for testing, and variables at the scheduling level depend on the decision of a professional, so that the sampling rate of the variables is often in the order of minutes, hours and even days. Data having such different sample rate properties is called multi-sample rate data, which has the following characteristics:

1) The data is incomplete. Multi-sample rate data has imperfections in that low sample rate variables have missing data at some sample times of high sample rate variables due to sampling rate inconsistencies. The lack is an inherent characteristic of the data and is not interfered by human factors.

2) The information is asymmetric. In multi-sample rate data, the sample rate of a process variable is often high, but the information contained therein is limited; the sampling rate of the quality benefit related variables is relatively low, but such data is of more research value for diagnostic systems.

These two characteristics make the conventional fault diagnosis model inapplicable to multi-sample rate data because most existing methods assume that the data is uniformly complete. To solve this problem, scholars at home and abroad have conducted a series of studies on multi-sampling rate data. Downsampling and upsampling are the most straightforward methods of processing multi-sample rate data. The down-sampling method deletes the high-sampling-rate data so as to reduce the sampling rate of all variables to the minimum. The up-sampling method adopts some interpolation or interpolation methods to fill the missing of the low-sampling-rate data relative to the high-sampling-rate data, so that the sampling rate of all variables is improved to the highest. Liu et al propose a data interpolation method based on KNN and SOM. Li et al propose a multiple sampling rate data conversion method of integer I times interpolation-integer D times decimation. Although the up-sampling method and the down-sampling method are simple to operate, they have the following disadvantages: downsampling can greatly reduce the amount of data, and when the sampling rates of different variables are extremely different, it may result in a lack of sufficient data training models. In addition, downsampling also reduces the real-time performance of offline diagnostics. The upsampling method may introduce some extra information when performing data interpolation or interpolation, thereby adversely affecting the diagnostic model. In order to perform failure diagnosis without changing the original data, some researchers have studied a method based on data reorganization. These methods divide the multi-sampling rate data into a plurality of uniform and complete data sets and build diagnostic models, respectively. Specifically, assuming that there are two sensors with a sampling period of k and a sampling period of 3k in the system, the values of the sensors at the second sampling rate can only be obtained at 3nk (n is a positive integer). Data at time 3nk were collected and a diagnostic model was built (with sensor 1 and sensor 2), and data at the remaining time was built into another diagnostic model (with sensor 1 only).

Congress et al propose a PCA-based process fault detection method that combines the variables in the system pairwise, thus allowing for repeated data reorganization. Feng et al, however, group data at different times according to their variables, and perform failure detection using k-neighbors. In addition, some scholars propose probabilistic framework-based methods that solve model parameters using maximum likelihood estimation, thereby solving the problem of relative absence of low-sampling rate data. Tian et al propose a fault diagnosis method based on Bayes with enhanced mobility level, which calculates likelihood probability by using EM method, and performs probability estimation on the current incomplete sample. However, this method is only suitable for data at a single time, and cannot be applied to multi-sampling rate time series data. In addition, this method of data reorganization requires the establishment of multiple diagnostic models, and the number of models increases as the number of different sampling rate variables increases. Therefore, this method requires a large amount of memory space, and the training of multiple models consumes more training time, thereby increasing the time and space complexity of the models.

Most of the above methods are machine learning based methods, and when processing large-scale industrial data sets, the methods often need to be combined with some feature extraction methods for diagnosis. Characterizing different types of industrial data relies on manual experience, which is sometimes difficult to obtain. The deep learning method can automatically learn hidden features in data, so that manual feature extraction is not needed.

Disclosure of Invention

The invention provides a fault diagnosis method and equipment for an industrial system based on multi-sampling rate data, which can automatically mine deep features of the multi-sampling rate data and improve the performance of fault diagnosis.

In order to achieve the technical purpose, the invention adopts the following technical scheme:

an industrial system fault diagnosis method based on multi-sampling rate data comprises the following steps:

s1, acquiring time sequence data of a plurality of groups of different sensors of an industrial system, wherein R different sampling rates exist in the plurality of groups of data;

s2, screening and reserving time sequence data of partial sensors according to the correlation aiming at the time sequence data of all the sensors with the same sampling rate;

s3, screening each type of time sequence data of the reserved sensors with the same sampling rate, respectively normalizing the time sequence data, and then sequentially splicing the time sequence data according to the sensors;

s4, taking spliced data corresponding to various sampling rates as input data of a network channel of the CNN model, and training the CNN model in an end-to-end mode by combining corresponding real classification labels to obtain a fault classifier;

and S5, when fault diagnosis is needed, acquiring current time sequence data of the screened and reserved sensor in the step S2, processing the current time sequence data according to the step S3, inputting the current time sequence data into the fault classifier obtained by training in the step S4 according to a method of corresponding sampling rate to a network channel, and outputting to obtain a current fault diagnosis result.

Further, screening the time-series data of the remaining sensors according to the correlation, specifically: calculating Pearson correlation coefficients between the time-series data of each two different sensors, and selecting all Pearson correlation coefficients with absolute values larger than a correlation threshold value; and then selecting one sensor from the two sensors corresponding to each selected Pearson correlation coefficient to retain data, and removing the data of the other sensor without using the data.

Further, the CNN model adopted by the fault classifier comprises a feature extraction module and a diagnosis classification module; the characteristic extraction module comprises a plurality of network channels, and the number of the network channels is the same as the number of sampling rates and corresponds to the sampling rates one by one; each network channel of the feature extraction module extracts features from the time sequence data of the corresponding sampling rate, and then the features are input into the diagnosis classification module for feature fusion and classification.

Further, different network channels of the feature extraction module have the same structure but different parameters: the number of convolution kernels arranged on each convolution layer on each network channel is proportional to the number of sensors with the sampling rate corresponding to the network channel; and setting the size of a convolution kernel on a corresponding network channel according to the length of the input time series data.

Further, the loss function used to train the CNN model is:

in the formula (I), the compound is shown in the specification,

a true category label; k is the number of categories of fault diagnosis classification; o is _j Representing the estimation probability of the input data prediction value belonging to the category j by the CNN model; l is the predicted loss;

updating the weights of the CNN model by calculating partial derivatives of the predicted loss L to the weights in the CNN model in multiple iterations:

where α is the learning rate.

Further, the industrial system may be a hydraulic system.

Further, in the sensor data obtained in step S1, the monitored physical quantities include pressure, power, flow, temperature, vibration, cooling efficiency, cooling capacity, and efficiency factor.

Further, the number of sensors for monitoring pressure is 6, the number of sensors for monitoring flow is 2, the number of sensors for monitoring temperature is 4, and the number of sensors for monitoring other physical quantities is 1.

Further, the types of fault diagnosis include: the cooling effect of the condenser is reduced in various degrees, the reversing function of the hydraulic valve is degenerated in various degrees, the leakage in the pump is in various degrees, and the gas leakage of the energy accumulator is in various degrees.

An industrial system fault diagnosis device based on multi-sampling rate data comprises a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the processor realizes the industrial system fault diagnosis method according to any one of the technical schemes.

Advantageous effects

The invention does not need to carry out any up-sampling and down-sampling operation on the original multi-sampling rate time sequence, thereby retaining the original information as much as possible and improving the utilization rate of the multi-sampling rate time sequence. In addition, the method can process multi-sampling rate data only by establishing one model, greatly reduces the time complexity and the space complexity of a classification model compared with a data recombination method, and effectively utilizes information of different sampling rate variables to improve the precision of fault diagnosis.

Drawings

FIG. 1 is a technical circuit diagram of a method according to an embodiment of the present application;

fig. 2 shows the results of fault diagnosis under different noises by the method according to the embodiment of the present application.

Detailed Description

The following describes embodiments of the present invention in detail, which are developed based on the technical solutions of the present invention, and give detailed implementation manners and specific operation procedures to further explain the technical solutions of the present invention.

The invention provides an industrial system fault diagnosis method based on multi-sampling rate data, which is applied to fault diagnosis of a hydraulic system in the embodiment and can obtain possible fault types: the cooling effect of the condenser is reduced to various degrees (close to complete failure, power reduction and full power), the reversing function of the hydraulic valve is degraded to various degrees (normal reversing function, small reversing delay, severe reversing delay and close to complete failure), the internal leakage of the pump to various degrees (no leakage, slight leakage and severe leakage) and the gas leakage of the energy accumulator to various degrees (proper pressure, slight pressure reduction, severe pressure reduction and close to complete failure).

The fault diagnosis method of the present embodiment, as shown in fig. 1, includes the following steps:

s1, acquiring time sequence data of a plurality of groups of different sensors of an industrial system, wherein R different sampling rates exist in the plurality of groups of data.

In this embodiment, a plurality of sensors are used to monitor a plurality of different physical quantities of the hydraulic system at respective sampling rates, as shown in the following table:

TABLE 1 Hydraulic System monitoring data

In table 1, a sensor with a sampling rate of 100Hz includes: 6 sensors PS1-PS6 for monitoring pressure, and 1 sensor EPS1 for monitoring electric power; the sensor with the sampling rate of 10Hz comprises: 2 sensors FS1-FS2 for monitoring flow; the sensor with the sampling rate of 1Hz comprises: 4 sensors TS1-TS4 to monitor temperature, 1 sensor VS1 to monitor vibration, 1 sensor CE to monitor cooling efficiency, 1 sensor CP to monitor cooling capacity, 1 sensor SE to monitor efficiency factor.

And S2, screening and reserving the time series data of partial sensors according to the correlation aiming at the time series data of all the sensors with the same sampling rate.

For all sensors at each sampling rate, a pearson correlation coefficient ρ is calculated between the time series data for each two sensors to quantify the correlation between the monitored variables for the two sensors:

wherein x is _i ,x _j Representing time series data monitored by two sensors at the same sampling rate, cov (x) _i ,x _j ) Is x _i ,x _j Covariance between, σ _i And σ _j Are each x _i ,x _j Standard deviation of (2). The pearson correlation coefficient has a value between-1 and 1, and the larger the absolute value, the stronger the correlation.

Then, from all the calculated pearson correlation coefficients, pearson correlation coefficients having an absolute value larger than a correlation threshold (set to 0.95 in the present embodiment) are selected, and of the two sensors corresponding to each selected pearson correlation coefficient, data of one sensor is selected to be retained, and data of the other sensor is not removed.

S3, screening each type of time sequence data of the sensors with the same sampling rate, respectively normalizing the time sequence data, and sequentially splicing the time sequence data according to the sensors.

And (3) assuming that the data comprises R sampling rates, splicing the data of the R sampling rate on the last dimension to form the input of a network channel R. It is defined as follows:

wherein M is _r Is the number of sensors at the sampling rate r, and n represents the number of samples.

And S4, taking the spliced data corresponding to the various sampling rates as input data of one network channel of the CNN model, and training the CNN model in an end-to-end mode by combining the corresponding real classification labels to obtain the fault classifier.

The fault diagnosis framework of the off-line training mainly comprises a feature extraction module and a diagnosis classification module. The feature extraction module includes a plurality of network channels capable of extracting features from raw data at multiple sample rates. And then the diagnosis classification module splices the features extracted by the network channels to form a global feature representation. Finally, the trained fault classifier can identify faults based on these representations.

(1) Feature extraction module

The feature extraction module is mainly composed of a convolutional neural network, and as shown in fig. 2, the following advantages are obtained. First, CNN is characterized by local connectivity and weight sharing. Therefore, compared with the common multilayer fully-connected neural network, the trainable parameters are fewer and the training speed is higher. Secondly, most industrial monitoring data have local relevance, and the CNN can learn and characterize locally relevant information, so that the method is suitable for industrial fault diagnosis. Finally, industrial monitoring data is easily polluted by noise, and CNN can extract the characteristics of invariance and enhance the robustness of a diagnostic model.

The feature extraction module is composed of a plurality of network channels, and each network channel extracts features of sampling rate data. Different network channels have the same structure, but the parameters constituting the network channels are set differently. The basic composition structure of each network channel is convolutional layer, reLU activation layer, batch Normalization (Batch Normalization) layer and Max-pooling (Max-Pooling) layer.

Convolutional layers are the core elements of a CNN, which consist of a number of convolutional kernels with weighting parameters. These convolution kernels will be convolved with the input to extract the translation invariant features. The design of the convolutional layer should match the characteristics of the underlying data, and therefore, the parameters need to be set according to the following principles. Firstly, the number of convolution kernels needs to be proportional to the number of sensors, and the number of sensors corresponding to the network channel input sampling data is more, the number of convolution kernels is more, so that more sensor data can be processed; in addition, the number of convolution kernels is also adjusted according to different industrial systems. In general, the deeper the convolutional layer, the better the feature extraction capability of the model. Second, the size of the convolution kernel is set according to the dimension (i.e., length) of the time-series data input by the network channel. The lower network (close to the input) should use a larger convolution kernel than the upper network because its receiving domain is wider and can extract low-frequency features, such as periodic variations in the signal; in addition, the large convolution kernel can also suppress high-frequency noise in the signal, which is of great significance to fault diagnosis of an actual industrial system. Smaller convolution kernels need to be employed in the top-level network (near the output) so that the model can capture local features. The output of the convolutional layer can be expressed as:

wherein C is ^q Features representing the q-th convolution kernel extraction, W ^p,q And b ^q Respectively representing the weights and offsets of the convolution kernels, P =1,2, \ 8230;, P, P representing the number of convolution channels, Q =1,2, \ 8230;, Q, Q being the set number of convolution kernels. Symbol denotes the convolution operator.

The feature extraction module then inputs the features extracted by the convolution layer into the batch normalization layer. By normalizing the data of each batch in the off-line training, the variation of the input distribution of each layer can be controlled, and the drift of the internal covariates is reduced. Thus, the batch normalization layer may prevent overfitting to some extent.

Let c _i The output h of the batch normalization layer is the data of one batch of the extracted features C _i Comprises the following steps:

in the formula, mu and sigma ² Is the mean and variance of the small batch of data. ε is a minimum value near 0 to prevent divide by 0 errors. The symbols γ and β represent learnable parameters. By using batch normalization, the small batch data will affect the parameters γ and β. This function amounts to introducing some interference in the training samples, thereby performing data enhancement and avoiding overfitting.

Subsequently, a non-linear activation layer is added in the feature extraction module, so that a non-linear representation of the features is obtained, and the representation and the recognition capability of the features are enhanced. If the function is not activated, the output of the feature extraction module is only a simple linear function, and complex data cannot be learned and simulated. Taking into account the convergence speed and gradient vanishing issues, a corrective linear unit (ReLU) function is employed as the activation function. Compared with the common Sigmoid and tanh activation functions, the ReLU function has a faster convergence rate, so that the gradient saturation can be prevented when the model is deep. The ReLU function is defined as follows:

T ^r ＝ReLU(H ^r )＝max(H ^r ,0)

wherein H ^r Represents the output of the batch normalization layer of the r-th network channel, and T ^r Is the output result of the ReLU activation layer.

Finally, the import pool layer handles the output of the ReLU activation. The pooling layer may down-sample the output of neighboring cell groups in the same feature map to extract the most representative features. The pooling layer can reduce the feature dimension and network parameters, improve the robustness of feature extraction and accelerate the calculation speed. The most common maximum pool layer adopted by the patent has the following expression:

Z ^r ＝max(T ^r )

the model can automatically learn features from the multi-sample rate data by a feature extraction module, thereby avoiding manual feature extraction requiring expert knowledge.

(2) Diagnostic classification module

After feature extraction, it needs to be fused at the feature level and the fused representation is sent to the fault classifier. The diagnostic classifier module is composed of a full-connected layer (full-connected layer), a batch normalization layer, a ReLU activation layer, and a classification layer, as shown with reference to fig. 2.

In order to extract a global representation of features learned from different sample rate data, the features Zr learned from the various network channels are fused to form a global feature Z. Since the extracted features are two-dimensional, while the fully connected layer can only process one-dimensional data, it is necessary to stretch the features into one-dimensional vectors before fusion. The global representation is then fed into a fully connected layer that is responsible for mapping the multidimensional input to the low dimensional data. In order to obtain better detection performance, a batch normalization layer and a ReLU activation layer are also adopted, which are defined as follows:

Z＝[Z ¹ ,Z ² ,…,Z ^R ]

G＝ReLU(BN(W _g Z+b _g ))

wherein W _g And b _g Representing the weight and bias of the fully connected layer. ReLU and BN represent the ReLU activation layer and batch normalization layer, respectively. In some complex fault detection problems, the performance of the diagnostic model can be improved by increasing the number of fully connected layers.

Finally, the representation is input to a classification layer, comprising K neurons and a Softmax function. The output of the Softmax function is a vector, where each number represents a target probability, whose function is defined as follows:

wherein O is _j Represents the estimated probability, θ, of the class j ^(j) Is a parameter of the Softmax function and K is the number of fault types. The method can not only avoid the defects of the up-sampling and down-sampling methods, but also effectively fuse the information contained in the multi-rate data samples, thereby improving the performance of fault diagnosis.

In the present embodiment, a hydraulic system is taken as an example of an industrial system, and neurons provided in a classification layer correspond to the following various failure types: various degrees of condenser cooling degradation (near complete failure, power droop, full power), various degrees of hydraulic valve reversal function degradation (normal reversal function, small reversal delay, severe reversal delay, near complete failure), various degrees of pump internal leakage (no leakage, slight leakage, severe leakage), and various degrees of accumulator gas leakage (proper pressure, slight depressurization, severe depressurization, near complete failure). The concrete steps are shown in a table 2:

TABLE 2 Hydraulic System Fault cases

Based on the fault diagnosis framework formed by the feature extraction module and the diagnosis classification module, the industrial system fault classifier can be obtained by using training data to train in an end-to-end mode.

Wherein each training sample comprises time series data X corresponding to various sampling rates ^r And corresponding real classification labels, the prediction loss of model training can be calculated by the cross entropy loss of the prediction output and the real classification labels, which is defined as follows:

wherein L is the predicted loss;

a true category label; k is the number of categories of fault diagnosis classification; o is _j Representing the estimated probability of the CNN model for the input data prediction belonging to category j.

The parameters of the model are updated by calculating the partial derivatives of the predicted loss L on the parameters in a number of iterations. Wherein the weights are updated as follows:

where α is the learning rate. In order to optimize the propagation process, the present embodiment employs an Adaptive Moment Estimation (Adam) stochastic optimization algorithm.

And S5, when fault diagnosis is needed, acquiring current time sequence data of the screened and reserved sensor in the step S2, processing the current time sequence data according to the step S3, inputting the current time sequence data into the fault classifier obtained by training in the step S4 according to a method of corresponding sampling rate and network channel, and outputting the current fault diagnosis result of the industrial system.

In order to verify the versatility and robustness of the method proposed in this patent, the following experiments were performed.

1. Fault diagnosis of different components

First, to verify the versatility of the proposed method of this patent, four elements of the hydraulic system were fault diagnosed, the specific fault types of which are shown in table 1. And (3) screening the sensors by using the Pearson correlation coefficient: for data with a sampling rate of 100Hz, the selected sensors are PS1-PS5. For data with a sampling rate of 10Hz, the sensors selected are FS1 and FS2. For data with a sample rate of 1Hz, the selected sensors are TS1, VS, CP, and SE. There were 2205 experimental samples, each collected during a 60 second hydraulic system duty cycle. 80% of the data were randomly divided into training sets and the remaining 20% were divided into test sets.

Considering three sampling rates for hydraulic system data, the number of channels for the diagnostic framework is set to three. Channels 1,2 and 3 process data at sampling rates of 100Hz, 10Hz and 1Hz, respectively. Each channel consists of two sets of structures. By setting different parameters, such as the size of the convolution filter, the model can extract features from the data at different sampling rates. During training, the learning rate of the Adam optimization algorithm is set to 0.001. The number of training iterations is set to vary from 10 to 50 depending on the diagnostic goal. To avoid the specificity and the contingency of the fault diagnosis results, 50 repeated experiments were performed. The average results are reported in table 3 in the format: mean test accuracy (%) ± standard deviation (%).

TABLE 3 Fault diagnosis accuracy under different elements

Condenser	Hydraulic valve	Hydraulic pump	Energy accumulator
				100.00±0.00	100.00±0.00	98.98±0.16	99.35±0.38

The fault diagnosis method provided by the invention can obtain good fault diagnosis results in different elements of the hydraulic system, and the method has strong universality.

2. Fault diagnosis in noisy environments

The experiment will verify the diagnostic effect of the method in a noisy environment. White gaussian noise with power Pn is added to the original signal to simulate a noisy environment. The power of the noise is proportional to the power of the original signal Ps, and a signal-to-noise ratio (SNR) is introduced to describe the ratio of Ps and Pn. The signal-to-noise ratio and power are calculated as follows:

SNR＝10lg(Ps/Pn)

wherein, the symbol x _i Data representing a signal or noise, and N represents a length of the data.

Experiments were performed on data with a signal-to-noise ratio of 20 to 40, in other words, the noise power increased from 0.01% to 1% of the signal power. The results are shown in FIG. 2. Generally, the accuracy of fault diagnosis improves as the signal-to-noise ratio increases. The noise environment has different effects on the fault diagnosis results of different components. In fault detection of coolers and valves, noise has little influence on results, and the diagnostic accuracy rate of the fault detection method is close to 100.0%. In the fault diagnosis of the pump and the energy accumulator with a complex fault scene, the influence of a noise environment on the diagnosis precision is larger, but the accuracy of the method still reaches more than 90%. With the increase of the signal-to-noise ratio, the diagnosis precision of the method is close to the precision in a noise-free environment, and the robustness of the method is proved.

3. Single fault and multiple fault concurrency diagnostics

In an actual hydraulic system, the failures of the four components may occur individually or simultaneously. When a compound fault occurs, the fault signature becomes more complex. In order to verify the performance of the model in both cases, single fault diagnosis and multiple fault diagnosis were performed on the hydraulic system.

The purpose of single fault detection is to diagnose which component in the hydraulic system is faulty. Specifically, the diagnostic model needs to identify the following four cases: single failure of the cooler, main pump, hydraulic valve and accumulator. The purpose of multiple fault detection is to identify whether a compound fault has occurred in the system. Two components of the hydraulic system are selected to detect the following four conditions: no failure, single failure per component, and composite failure of two components. In experiments, the main pump and accumulator were chosen because their failure characteristics were more covert than those of the cooler and valves, which made multiple failure detection more challenging. Thus, the model needs to identify conditions including normal conditions, main pump failure, accumulator failure, and a combined failure of both components. It should be noted that these two experiments only identified the faulty component, rather than measuring its severity. The results of the experiment are shown in table 4.

TABLE 4 Fault diagnosis of Single and multiple Fault concurrency

Single fault	Multiple fault concurrency
		98.12±0.98	99.97±0.18

Experimental results show that the method has better fault identification capability and stability. This is due to the deep structure of the method, which can mine richer detection information and hidden coupling relation from multi-sampling rate data.

The fault diagnosis method of the multi-sampling rate data feature level fusion can be applied to the fault diagnosis problem of a complex industrial system. The method can automatically mine deep features of the multi-sampling rate data by utilizing the convolutional neural network, and can extract hidden information which cannot be captured by the traditional statistical features, such as fault features in a noise environment. In application, the method can handle sampling rate ratios as high as 1:100, respectively.

The above embodiments are preferred embodiments of the present application, and those skilled in the art can make various changes or modifications without departing from the general concept of the present application, and such changes or modifications should fall within the scope of the protection claimed in the present application.

Claims (10)

1. A method for fault diagnosis in an industrial system based on multi-sampling rate data, comprising:

s1, acquiring time sequence data of a plurality of groups of different sensors of an industrial system, wherein R different sampling rates exist in the plurality of groups of data;

s2, screening and reserving time series data of partial sensors according to the correlation aiming at the time series data of all the sensors with the same sampling rate;

s3, screening each type of time sequence data of the reserved sensors with the same sampling rate, respectively normalizing the time sequence data, and then sequentially splicing the time sequence data according to the sensors;

s4, taking spliced data corresponding to various sampling rates as input data of one network channel of the CNN model, and training the CNN model in an end-to-end mode by combining corresponding real classification labels to obtain a fault classifier;

and S5, when fault diagnosis is required, acquiring current time sequence data of the screened and reserved sensor in the step S2, processing the current time sequence data according to the step S3, inputting the current time sequence data into a fault classifier obtained by training in the step S4 according to a method of corresponding sampling rate to a network channel, and outputting to obtain a current fault diagnosis result.

2. The failure diagnosis method according to claim 1, wherein the time-series data of the remaining partial sensors are screened according to the correlation, specifically: calculating a Pearson correlation coefficient between the time-series data of each two different sensors, and selecting all Pearson correlation coefficients with absolute values larger than a correlation threshold value; and then selecting one sensor from the two sensors corresponding to each selected Pearson correlation coefficient to retain data, and removing the data of the other sensor without using the data.

3. The fault diagnosis method according to claim 1, wherein the CNN model adopted by the fault classifier comprises a feature extraction module and a diagnosis classification module; the feature extraction module comprises a plurality of network channels, and the number of the network channels is the same as the number of the sampling rates and corresponds to the sampling rates one by one; each network channel of the feature extraction module extracts features from the time sequence data of the corresponding sampling rate, and then the features are input into the diagnosis classification module for feature fusion and classification.

4. The fault diagnosis method according to claim 3, characterized in that the different network channels of the feature extraction module have the same structure but different parameters: the number of convolution kernels arranged on each convolution layer on each network channel is proportional to the number of sensors with the sampling rate corresponding to the network channel; and setting the size of the convolution kernel on the corresponding network channel according to the length of the input time sequence data.

5. The fault diagnosis method according to claim 3, characterized in that the loss function used for training the CNN model is:

in the formula (I), the compound is shown in the specification,

a true category label; k is the number of categories of fault diagnosis classification; o is _j Representing the estimation probability of the input data prediction value belonging to the category j by the CNN model; l is the predicted loss;

updating the weights of the CNN model by calculating partial derivatives of the predicted loss L to the weights in the CNN model in multiple iterations:

where α is the learning rate.

6. The fault diagnosis method according to claim 1, characterized in that the industrial system may be a hydraulic system.

7. The failure diagnosing method according to claim 6, wherein the monitored physical quantities of the sensor data acquired in step S1 include pressure, power, flow, temperature, vibration, cooling efficiency, cooling capacity, and efficiency factor.

8. The failure diagnosing method according to claim 7, wherein the number of sensors for monitoring pressure is 6, the number of sensors for monitoring flow rate is 2, the number of sensors for monitoring temperature is 4, and the number of sensors for monitoring other physical quantities is 1, respectively.

9. The fault diagnosis method according to claim 6, wherein the types of fault diagnosis include: various degrees of condenser cooling degradation, various degrees of hydraulic valve commutation function degradation, various degrees of pump internal leakage, and various degrees of accumulator gas leakage.

10. An industrial system fault diagnosis device based on multisampling rate data, comprising a memory and a processor, the memory having stored therein a computer program, characterized in that the computer program, when executed by the processor, causes the processor to carry out the method according to any one of claims 1 to 9.

Images