Method and apparatus for fault diagnosis of industrial systems based on multi-sampling rate data

Abstract

The application discloses an industrial system fault diagnosis method and device based on multi-sampling rate data, and the method is as follows: acquiring time series data of multiple groups of different sensors of an industrial system, the multiple groups of data having different sampling rates; for the time series data of all sensors of each sampling rate, screening and reserving the time series data of part of the sensors according to correlation; normalizing the time series data of the screened and reserved sensors of each sampling rate respectively, and then splicing according to sensors; training a CNN model in an end-to-end manner by combining the spliced data corresponding to multiple sampling rates and corresponding real classification labels, to obtain a fault classifier; when fault diagnosis is needed, the trained fault classifier is used to perform fault diagnosis on the industrial system based on the multi-sampling rate data. The application can automatically mine deep features of the multi-sampling rate data, and effectively utilize information of variables with different sampling rates to improve the precision of fault diagnosis.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of industrial system fault diagnosis, and particularly relates to an industrial system fault diagnosis method and device based on multi-sampling rate data. BACKGROUND

[0002] Industry is of great importance to the development of national economy. Due to the increasingly complex structure and the increasingly large scale of industrial systems, once a fault occurs, it will have an adverse impact on the production process and personnel safety. Therefore, the safety and reliability of industrial systems have attracted great attention. Accurate fault diagnosis of the system and timely maintenance strategy can avoid major safety accidents, reduce maintenance costs and improve production efficiency. The development of computers, sensors and storage technology has brought new opportunities for data-driven fault diagnosis methods. Such methods train offline fault diagnosis models through data collected by sensors, and make diagnosis decisions based on real-time data.

[0003] Multi-sensor information fusion is the development trend in the field of fault diagnosis. Compared with a single sensor, multiple sensors can provide more comprehensive and rich system information, thereby improving the reliability of diagnosis. Multi-sensor information fusion can be carried out at different levels: data level, feature level and decision level. However, the sampling rates of different sensors in actual industrial systems are often inconsistent. On the one hand, the sampling rate of a sensor depends to some extent on the physical quantity it monitors. Specifically, for some slowly changing process variables, such as pressure, temperature and flow, their sampling rates are mostly minute-level; while for some rapidly changing variables, such as vibration, current and sound signals, their sampling rates are generally second-level. On the other hand, some components of the process industry need to be collected manually and sent to the laboratory for testing, while the variables at the scheduling level depend on the decisions of professional personnel, so the sampling rates of these variables are often minute-level, hour-level, or even day-level. Data with such different sampling rate properties are referred to as multi-sampling rate data, which have the following characteristics:

[0004] 1) Incomplete data. Multi-sampling rate data is incomplete, that is, due to the inconsistency of the sampling rate, there are missing data of low sampling rate variables at some sampling times of high sampling rate variables. This missing is an inherent characteristic of data and is not affected by human factors.

[0005] 2) Information asymmetry. In multi-sampling rate data, the sampling rate of process variables is often high, but the information contained is limited; while the sampling rate of quality benefit related variables is relatively low, but such data is more valuable for diagnosing the system.

[0006] These two characteristics make it impossible to apply traditional fault diagnosis models to multisampling rate data, because most existing methods assume that the data is uniformly complete. To solve this problem, domestic and foreign scholars have carried out a series of researches on multisampling rate data. Down-sampling and up-sampling are the most direct methods to deal with multisampling rate data. Down-sampling deletes high sampling rate data, so that the sampling rate of all variables is reduced to the minimum. Up-sampling fills in the missing data of low sampling rate data relative to high sampling rate data by using some interpolation or interpolation methods, so that the sampling rate of all variables is increased to the maximum. Liu et al. proposed a data interpolation method based on KNN and SOM. Li et al. proposed a multisampling rate data conversion method of integer I times interpolation-integral D times decimation. Although up-sampling and down-sampling are simple to operate, they have the following shortcomings: down-sampling greatly reduces the amount of data, and when the sampling rates of different variables differ greatly, it may result in a lack of sufficient data to train the model. In addition, down-sampling also reduces the real-time performance of offline diagnosis. Up-sampling may introduce some additional information when interpolating or interpolating data, which may adversely affect the diagnosis model. In order to perform fault diagnosis without changing the original data, some scholars have studied data reorganization-based methods. These methods divide multisampling rate data into multiple uniformly complete data sets and establish diagnostic models respectively. Specifically, if there are two sensors with sampling periods of k and 3k in the system, the value of the second sampling rate sensor can only be obtained at 3nk time points (n is a positive integer). Collecting the data at 3nk time points and establishing a diagnostic model (containing sensor 1 and sensor 2), the data at the remaining time points is established another diagnostic model (only sensor 1).

[0007] Tian et al. proposed a PCA-based process fault detection method, which combines variables in pairs, so it is a repeated data reorganization. Feng et al. grouped the data at different time points according to the number of variables, and used k-neighborhood for fault detection. In addition, some scholars have proposed methods based on the probability framework, which use maximum likelihood estimation to solve model parameters, thus solving the problem of relative missing of low sampling rate data. Tian et al. proposed a fault diagnosis method based on enhanced mobile level Bayesian, which uses the EM method to calculate the likelihood probability to estimate the realization probability of the current incomplete sample. However, this method is only applicable to data at a single time point, and cannot be applied to multisampling rate time series data. In addition, this data reorganization method needs to establish multiple diagnostic models, and the number of models will increase with the increase in the number of variables with different sampling rates. Therefore, this method requires more storage space, and the training of multiple models also consumes more training time, thus increasing the time complexity and space complexity of the model.

[0008] Most of the above methods are machine learning-based methods, which often need to combine some feature extraction methods for diagnosis when processing large-scale industrial data sets. Feature extraction for different types of industrial data depends on artificial experience, which is sometimes difficult to obtain. Deep learning methods can automatically learn hidden features in data, so manual feature extraction is not required. SUMMARY

[0009] The present application provides an industrial system fault diagnosis method and device based on multi-sampling rate data, which can automatically mine deep features of multi-sampling rate data and improve the performance of fault diagnosis.

[0010] To achieve the above technical purpose, the present application adopts the following technical scheme:

[0011] An industrial system fault diagnosis method based on multi-sampling rate data, comprising:

[0012] S1, obtaining time series data of multiple groups of different sensors of an industrial system, the multiple groups of data having R different sampling rates;

[0013] S2, for all time series data of sensors of each same sampling rate, filtering and retaining part of the time series data of sensors according to correlation;

[0014] S3, normalizing the time series data of the filtered and retained sensors of each same sampling rate, and then sequentially splicing the time series data of the sensors;

[0015] S4, taking the spliced data corresponding to multiple sampling rates as input data of a network channel of a CNN model, combining the corresponding real classification labels, and training the CNN model in an end-to-end manner to obtain a fault classifier;

[0016] S5, when fault diagnosis is needed, obtaining the current time series data of the filtered and retained sensors of step S2, then processing the current time series data according to step S3, and then inputting the current time series data to the fault classifier obtained in step S4 according to the method corresponding to the sampling rate and the network channel, and outputting the current fault diagnosis result.

[0017] Further, the filtering and retaining part of the time series data of sensors according to correlation is specifically: calculating the Pearson correlation coefficient between the time series data of each two different sensors, and selecting all Pearson correlation coefficients with an absolute value greater than a correlation threshold value; then in the two sensors corresponding to each selected Pearson correlation coefficient, the data of one sensor is retained and the data of the other sensor is removed.

[0018] Furthermore, the CNN model used in the fault classifier includes a feature extraction module and a diagnostic classification module. The feature extraction module includes multiple network channels, and the number of network channels is the same as the number of sampling rates and corresponds one-to-one. Each network channel of the feature extraction module extracts features from the time series data of the corresponding sampling rate, and then inputs them into the diagnostic classification module for feature fusion and classification.

[0019] Furthermore, the different network channels of the feature extraction module have the same structure but different parameters: the number of convolution kernels set in each convolutional layer of each network channel is proportional to the number of sensors with the corresponding sampling rate of that network channel; the size of the convolution kernels on the corresponding network channel is set according to the length of the input time series data.

[0020] Furthermore, the loss function used to train the CNN model is:

[0021]

[0022] In the formula, K represents the actual classification label; K is the number of categories in the fault diagnosis classification; O j L represents the estimated probability that the CNN model's predicted value belongs to class j; L is the prediction loss.

[0023] The weights of the CNN model are updated by calculating the partial derivative of the prediction loss L with respect to the weights in the CNN model over multiple iterations.

[0024]

[0025] In the formula, α is the learning rate.

[0026] Furthermore, an industrial system can be a hydraulic system.

[0027] Furthermore, the sensor data acquired in step S1 includes physical quantities such as pressure, power, flow rate, temperature, vibration, cooling efficiency, cooling capacity, and efficiency factor.

[0028] Furthermore, there are 6 sensors for monitoring pressure, 2 sensors for monitoring flow rate, 4 sensors for monitoring temperature, and 1 sensor for each of the other physical quantities.

[0029] Furthermore, the types of fault diagnosis include: various degrees of decrease in condenser cooling effect, various degrees of degradation in hydraulic valve reversing function, various degrees of internal pump leakage, and various degrees of accumulator gas leakage.

[0030] An industrial system fault diagnosis device based on multi-sampling rate data comprises a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to realize the industrial system fault diagnosis method of any one of the technical solutions.

[0031] Advantages

[0032] The original multi-sampling rate time sequence does not need to be up-sampled and down-sampled, so that the original information is retained as much as possible, and the utilization rate of the multi-sampling rate time sequence is improved. In addition, the method can process multi-sampling rate data by establishing only one model, and compared with the data reorganization method, the time complexity and space complexity of the classification model are greatly reduced, and the information of variables with different sampling rates is effectively utilized to improve the fault diagnosis accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 is a technical line diagram of the method described in the embodiments of the application;

[0034] Figure 2 is a fault diagnosis result of the method described in the embodiments of the application under different noises. DETAILED DESCRIPTION

[0035] The embodiments of the application are described in detail below. The embodiments are based on the technical solutions of the application, and give detailed implementation modes and specific operation processes, and further explain and describe the technical solutions of the application.

[0036] The application provides an industrial system fault diagnosis method based on multi-sampling rate data. In the embodiments, the method is applied to fault diagnosis of a hydraulic system, and possible fault types include: multiple degrees of condenser cooling reduction (close to complete failure, power reduction, full power), multiple degrees of hydraulic valve reversing function degradation (normal reversing function, small reversing delay, severe reversing delay, close to complete failure), multiple degrees of pump internal leakage (no leakage, slight leakage, severe leakage), and multiple degrees of accumulator gas leakage (proper pressure, slight pressure reduction, severe pressure reduction, close to complete failure).

[0037] The fault diagnosis method of the embodiments, as shown in Figure 1 comprises the following steps:

[0038] S1, acquiring time sequence data of multiple groups of different sensors of an industrial system, the multiple groups of data having R different sampling rates.

[0039] In the embodiments, multiple sensors are used to monitor multiple different physical quantities of the hydraulic system at respective sampling rates, as shown in the following table:

[0040] Table 1 Hydraulic system monitoring data

[0041]

[0042]

[0043] In Table 1, the sensors with a sampling rate of 100 Hz include 6 sensors PS1-PS6 for monitoring pressure, 1 sensor EPS1 for monitoring power; the sensors with a sampling rate of 10 Hz include 2 sensors FS1-FS2 for monitoring flow; the sensors with a sampling rate of 1 Hz include 4 sensors TS1-TS4 for monitoring temperature, 1 sensor VS1 for monitoring vibration, 1 sensor CE for monitoring cooling efficiency, 1 sensor CP for monitoring cooling capacity, and 1 sensor SE for monitoring efficiency factor.

[0044] S2, for each same sampling rate, time series data of all sensors are screened according to correlation to retain time series data of part of the sensors.

[0045] For all sensors of each sampling rate, the Pearson correlation coefficient ρ between time series data of each two sensors is calculated to quantify the correlation between the two corresponding sensors monitoring variables:

[0046]

[0047] wherein x i ,x j represent time series data monitored by two sensors under the same sampling rate, cov(x i ,x j ) is the covariance between x i ,x j , and σ i and σ j are standard deviations of x i ,x j . The value of the Pearson correlation coefficient is between -1 and 1, and the greater the absolute value, the stronger the correlation.

[0048] Then, from all the calculated Pearson correlation coefficients, the Pearson correlation coefficients with an absolute value greater than a correlation threshold (set to 0.95 in this embodiment) are selected, and in each two sensors corresponding to the selected Pearson correlation coefficients, the data of one sensor is retained and the data of the other sensor is removed.

[0049] S3, time series data of the screened sensors of each same sampling rate are normalized respectively and then spliced in sequence according to the sensors.

[0050] Assume that there are R sampling rates in the data, the data of the rth sampling rate is concatenated in the last dimension to form the input of network channel r. The definition is as follows:

[0051]

[0052] where M r is the number of sensors at the sampling rate r, and n represents the number of samples.

[0053] S4, the concatenated data corresponding to multiple sampling rates is taken as the input data of a network channel of the CNN model, and the corresponding true classification label is combined to train the CNN model in an end-to-end manner to obtain a fault classifier.

[0054] The offline training fault diagnosis framework mainly consists of a feature extraction module and a diagnosis classification module. The feature extraction module contains multiple network channels, which can extract features from multi-sampling rate raw data. Then the diagnosis classification module concatenates the features extracted by these network channels to form a global feature representation. Finally, the trained fault classifier can identify faults according to these representations.

[0055] (1) Feature extraction module

[0056] The feature extraction module is mainly composed of a convolutional neural network, as shown in FIG. 1, and its advantages are as follows. First, the CNN has the characteristics of local connection and weight sharing. Therefore, compared with ordinary multi-layer fully connected neural networks, it has fewer trainable parameters and faster training speed. Second, most industrial monitoring data have local correlation, and CNN can learn and represent local correlation information, so it is suitable for industrial fault diagnosis. Finally, industrial monitoring data are easily contaminated by noise, and CNN can extract shift-invariant features to enhance the robustness of the diagnosis model. Figure 2

[0057] The feature extraction module is composed of multiple network channels, each of which extracts features of data of one sampling rate. Different network channels have the same structure, but the parameters constituting the network channel are set differently. The basic structure of each network channel includes a convolutional layer, a ReLU activation layer, a batch normalization (Batch Normalization) layer, and a maximum pooling (Max-pooling) layer.

[0058] ​The convolutional layer is the core unit of CNN, which is composed of multiple convolution kernels with weighted parameters. These convolution kernels will be convolved with the input to extract shift-invariant features. The design of the convolutional layer should match the features of the underlying data, so the parameters need to be set according to the following principles. First, the number of convolution kernels needs to be proportional to the number of sensors. The more sensors the network channel input samples correspond to, the more convolution kernels are set to handle more sensor data. In addition, the number of convolution kernels is adjusted according to different industrial systems. Generally speaking, 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 to the network channel. The underlying network (close to the input end) should use larger convolution kernels relative to the top layer network, because its receptive field is wider and can extract low-frequency features such as periodic changes in the signal. In addition, large convolution kernels can also suppress high-frequency noise in the signal, which is of great significance to fault diagnosis in actual industrial systems. Smaller convolution kernels are used in the top layer network (close to the output end) to enable the model to capture local features. The output of the convolutional layer can be represented as:

[0059]

[0060] where C q represents the feature extracted by the qth convolution kernel, W p,q and b q represent the weight and bias of the convolution kernel respectively, p = 1, 2, …, P, P represents the number of convolution channels, q = 1, 2, …, Q, Q is the number of convolution kernels set. The symbol * represents the convolution operator.

[0061] After that, the feature extraction module inputs the features extracted by the convolutional layer to the batch normalization layer. In offline training, by normalizing each batch of data, the variation of the input distribution of each layer can be controlled, and the internal covariate shift can be reduced. Therefore, the batch normalization layer can prevent overfitting to some extent.

[0062] Assuming c i is a batch of data of the above-mentioned extracted feature C, the output h i of the batch normalization layer is:

[0063]

[0064]

[0065] where μ and σ 2These are the mean and variance of the mini-batch data. ε is a very small value close to 0 to prevent division by zero errors. The symbols γ and β represent the learnable parameters. By using batch normalization, the mini-batch data will affect the parameters γ and β. This function is equivalent to introducing some perturbation into the training samples, thereby performing data augmentation and avoiding overfitting.

[0066] Subsequently, a non-linear activation layer is added to the feature extraction module to obtain a non-linear representation of the features, enhancing their representation and recognition capabilities. Without an activation function, the output of the feature extraction module is merely a simple linear function, unable to learn and simulate complex data. Considering convergence speed and the vanishing gradient problem, the Corrected Linear Unit (ReLU) function is used as the activation function. Compared to common sigmoid and tanh activation functions, the ReLU function has a faster convergence speed, thus preventing gradient saturation in deeper models. The ReLU function is defined as follows:

[0067] T r =ReLU(H r ) = max(H r ,0)

[0068] Among them, H r T represents the output of the batch normalization layer for the r-th network channel, while T r This is the output of the ReLU activation layer.

[0069] Finally, a pooling layer is introduced to process the output of ReLU activation. Pooling layers downsample the outputs of adjacent unit groups within the same feature map to extract the most representative features. Pooling layers reduce feature dimensionality and network parameters, improve the robustness of feature extraction, and accelerate computation. This patent uses the most commonly used max pooling layer, whose expression is as follows:

[0070] Z r =max(T) r )

[0071] Through the feature extraction module, the model can automatically learn features from multi-sampling rate data, thus avoiding manual feature extraction that requires expert knowledge.

[0072] (2) Diagnostic classification module

[0073] After feature extraction, the features need to be fused at the feature level, and the fused representation is then fed into the fault classifier. The diagnostic classification module consists of a fully-connected layer, a batch normalization layer, a ReLU activation layer, and a classification layer, as shown in the reference. Figure 2 As shown.

[0074] To extract a global representation of features learned from data at different sampling rates, the features Zr learned from each network channel are fused to form the global feature Z. Since the extracted features are two-dimensional, and fully connected layers can only process one-dimensional data, the features need to be stretched into a one-dimensional vector before fusion. The global representation is then fed into a fully connected layer, which maps the multi-dimensional input to low-dimensional data. To achieve better detection performance, batch normalization layers and ReLU activation layers are also used, defined as follows:

[0075] Z = [Z 1 Z 2 ,…,Z R ]

[0076] G = ReLU(BN(W) g Z+b g ))

[0077] Among them W g and b g This represents the weights and biases of the fully connected layer. ReLU and BN represent 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.

[0078] Finally, the above representation is input into the classification layer, which consists of K neurons and a Softmax function. The output of the Softmax function is a vector, where each number represents the probability of the target, and its function is defined as follows:

[0079]

[0080] Among them O j θ represents the estimated probability of category j. (j) Here, K is the parameter of the Softmax function, and K is the number of fault types. This method not only avoids the shortcomings of upsampling and downsampling methods, but also effectively fuses the information contained in multi-rate data samples, thus improving the performance of fault diagnosis.

[0081] In this embodiment, the hydraulic system is used as an example of an industrial system. The neurons set in the classification layer correspond to the following various fault types: multiple degrees of condenser cooling effect reduction (near-complete failure, power reduction, full power); multiple degrees of hydraulic valve directional function degradation (normal directional function, slight directional delay, severe directional delay, near-complete failure); multiple degrees of pump internal leakage (no leakage, slight leakage, severe leakage); and multiple degrees of accumulator gas leakage (appropriate pressure, slight pressure drop, severe pressure drop, near-complete failure). See Table 2 for details.

[0082] Table 2 Hydraulic System Fault Conditions

[0083]

[0084] Based on the fault diagnosis framework consisting of the feature extraction module and the diagnostic classification module, the industrial system fault classifier can be obtained by training the data in an end-to-end manner.

[0085] Each training sample includes time-series data X corresponding to multiple sampling rates. r Given the corresponding true classification labels, the prediction loss during model training can be calculated using the cross-entropy loss between the predicted output and the true classification labels, as defined below:

[0086]

[0087] In the formula, L represents the prediction loss; K represents the actual classification label; K is the number of categories in the fault diagnosis classification; O j This represents the estimated probability that the CNN model predicts the value of the input data as belonging to category j.

[0088] The model parameters are updated by calculating the partial derivatives of the prediction loss L with respect to the parameters over multiple iterations. The weights are updated as follows:

[0089]

[0090] Where α is the learning rate. To optimize the headway propagation process, this embodiment employs the Adaptive Moment Estimation (Adam) stochastic optimization algorithm.

[0091] S5. When fault diagnosis is required, obtain the current time series data of the sensor that was filtered and retained in step S2, then process the current time series data according to step S3, and then input it into the fault classifier trained in S4 according to the method of sampling rate and network channel correspondence. The output is the current fault diagnosis result of the industrial system.

[0092] To verify the universality and robustness of the method proposed in this patent, the following experiments were conducted.

[0093] 1. Fault diagnosis of different components

[0094] First, to verify the universality of the method proposed in this patent, fault diagnosis was performed on four components of the hydraulic system, and the specific fault types are shown in Table 1. Pearson correlation coefficients were used for sensor selection: for data with a sampling rate of 100Hz, sensors PS1-PS5 were selected. For data with a sampling rate of 10Hz, sensors FS1 and FS2 were selected. For data with a sampling rate of 1Hz, sensors TS1, VS, CP, and SE were selected. A total of 2205 experimental samples were collected, each sample being collected during a 60-second hydraulic system work cycle. 80% of the data was randomly assigned to the training set, and the remaining 20% ​​was assigned to the test set.

[0095] Considering the three sampling rates of the hydraulic system data, the number of channels in the diagnostic framework was set to three. Channels 1, 2, and 3 processed data at sampling rates of 100Hz, 10Hz, and 1Hz, respectively. Each channel consisted of two sets of structures. By setting different parameters, such as the size of the convolutional filter, the model could extract features from the data at different sampling rates. During training, the learning rate of the Adam optimization algorithm was set to 0.001. The number of training iterations varied from 10 to 50 depending on the diagnostic objective. To avoid the particularity and randomness of the fault diagnosis results, 50 repeated experiments were conducted. The average results are recorded in Table 3, in the format: average test accuracy (%) ± standard deviation (%).

[0096] Table 3 Fault diagnosis accuracy for different components

[0097] Condenser Hydraulic valve Hydraulic pump Accumulator 100.00±0.00 100.00±0.00 98.98±0.16 99.35±0.38

[0098] It can be seen that the fault diagnosis method provided by the present invention has achieved good fault diagnosis results in different components of the hydraulic system, indicating that the method has strong versatility.

[0099] 2. Fault diagnosis in noisy environments

[0100] This experiment will verify the diagnostic effectiveness of this method in noisy environments. Gaussian white 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 the signal-to-noise ratio (SNR) is introduced to describe the ratio of Ps to Pn. The calculation of SNR and power is as follows:

[0101] SNR = 10lg(Ps / Pn)

[0102]

[0103] Where, symbol x i The data represents the signal or noise, where N represents the length of the data.

[0104] Experiments were conducted on data with a signal-to-noise ratio (SNR) of 20–40, in other words, the noise power increased from 0.01% to 1% of the signal power. The results are as follows: Figure 2 As shown in the figure, the accuracy of fault diagnosis generally increases with the increase of the signal-to-noise ratio. The noise environment has different effects on the fault diagnosis results of different components. In the fault detection of coolers and valves, noise has little impact on the results, and the diagnostic accuracy is close to 100.0%. However, in the fault diagnosis of pumps and accumulators, which have more complex fault scenarios, the noise environment has a greater impact on the diagnostic accuracy, but the accuracy of this method still reaches over 90%. With the increase of the signal-to-noise ratio, the diagnostic accuracy of this method approaches that of a noise-free environment, demonstrating its robustness.

[0105] 3. Single fault and concurrent fault diagnosis

[0106] In real hydraulic systems, failures of the four components can occur individually or simultaneously. When a compound failure occurs, the fault characteristics become more complex. To verify the model's performance in both cases, single-fault and multi-fault diagnostics were performed on the hydraulic system.

[0107] The purpose of single fault detection is to diagnose which component in the hydraulic system is faulty. Specifically, the diagnostic model needs to identify four scenarios: a single fault in the cooler, main pump, hydraulic valve, and accumulator. The purpose of multi-fault detection is to identify whether a compound fault has occurred in the system. Two components of the hydraulic system are selected to detect four scenarios: no fault, a single fault in each component, and a compound fault in both components. In the experiments, the main pump and accumulator were chosen because their fault characteristics are more subtle than those of the cooler and valves, making multi-fault detection more challenging. Therefore, the model needs to identify the following conditions: normal conditions, main pump failure, accumulator failure, and a compound fault in both components. It should be noted that these two experiments only identify the faulty component, not measure its severity. The experimental results are shown in Table 4.

[0108] Table 4 Fault diagnosis for single faults and concurrent multiple faults

[0109] Single fault Multiple fault concurrency 98.12±0.98 99.97±0.18

[0110] Experimental results demonstrate that the method of this invention possesses good fault identification capabilities and stability. This is attributed to the method's deep structure, which allows it to extract richer detection information and hidden coupling relationships from multi-sampling-rate data.

[0111] The fault diagnosis method based on hierarchical fusion of multi-sampling-rate data features proposed in this invention can be applied to fault diagnosis problems in complex industrial systems. Utilizing convolutional neural networks, this method can automatically mine deep features from multi-sampling-rate data and extract hidden information that traditional statistical features cannot capture, such as fault features in noisy environments. In practice, this method can handle data with sampling rate ratios as high as 1:100.

[0112] The above embodiments are preferred embodiments of this application. Those skilled in the art can make various changes or improvements based on them. Without departing from the overall concept of this application, such changes or improvements should fall within the scope of protection claimed in this application.

Claims

1. A fault diagnosis method for industrial systems based on multi-sampling-rate data, characterized in that, include: S1, acquire time series data from multiple different sensors in the industrial system, where there are R different sampling rates for these multiple sets of data; S2, for time series data of all sensors with the same sampling rate, filter and retain time series data of some sensors based on correlation; S3, the time series data of each sensor with the same sampling rate are filtered and retained, normalized separately, and then spliced ​​together according to the sensor in sequence; S4. The concatenated data corresponding to multiple sampling rates is used as the input data of one network channel of the CNN model. Combined with the corresponding real classification labels, the CNN model is trained in an end-to-end manner to obtain the fault classifier. The CNN model used in the fault classifier includes a feature extraction module and a diagnostic classification module. The feature extraction module includes multiple network channels, and the number of network channels is the same as the number of sampling rates and corresponds one-to-one. Each network channel of the feature extraction module extracts features from the time series data of the corresponding sampling rate, and then inputs them into the diagnostic classification module for feature fusion and classification. Different network channels in the feature extraction module have the same structure but different parameters: the number of convolutional kernels in each convolutional layer of each network channel is proportional to the number of sensors at the corresponding sampling rate of that network channel; the size of the convolutional kernels on the corresponding network channel is set according to the length of the input time series data. S5. When fault diagnosis is required, obtain the current time series data of the sensor that was filtered and retained in step S2, then process the current time series data according to step S3, and then input it into the fault classifier trained in S4 according to the method of sampling rate and network channel correspondence, and the output is the current fault diagnosis result.

2. The fault diagnosis method according to claim 1, characterized in that, The time series data of some sensors are retained based on correlation. Specifically, the Pearson correlation coefficient between time series data of two different sensors is calculated, and all Pearson correlation coefficients with absolute values ​​greater than the correlation threshold are selected. Then, among the two sensors corresponding to each selected Pearson correlation coefficient, the data of one sensor is selected and retained, while the data of the other sensor is removed.

3. The fault diagnosis method according to claim 1, characterized in that, The loss function used to train the CNN model is: ； In the formula, These are real category tags; K This refers to the number of categories in fault diagnosis classification; This represents the estimated probability that the CNN model predicts the value of the input data as belonging to class j; To predict losses; By calculating the prediction loss over multiple iterations The weights of the CNN model are updated using the partial derivatives of the weights in the CNN model: ； In the formula, It is the learning rate.

4. The fault diagnosis method according to claim 1, characterized in that, Industrial systems are hydraulic systems.

5. The fault diagnosis method according to claim 4, characterized in that, The sensor data acquired in step S1 includes physical quantities such as pressure, power, flow rate, temperature, vibration, cooling efficiency, cooling capacity, and efficiency factor.

6. The fault diagnosis method according to claim 5, characterized in that, There are 6 sensors for monitoring pressure, 2 sensors for monitoring flow rate, 4 sensors for monitoring temperature, and 1 sensor for each of the other physical quantities.

7. The fault diagnosis method according to claim 4, characterized in that, The types of fault diagnosis include: various degrees of decrease in condenser cooling effect, various degrees of degradation in hydraulic valve reversing function, various degrees of internal pump leakage, and various degrees of accumulator gas leakage.

8. An industrial system fault diagnosis device based on multi-sampling rate data, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the computer program is executed by the processor, it causes the processor to implement the method as described in any one of claims 1 to 7.

Images