A remote state monitoring method and system for hydraulic pumps based on cloud-based intelligent reconstruction of low sampling rate signals

Abstract

The present invention discloses a method and system for remote status monitoring of a hydraulic pump based on cloud-based intelligent reconstruction of low-sampling rate signals, and belongs to the field of remote status monitoring. This method effectively alleviates the high cost and delay problems faced by traditional cloud platforms when uploading large-scale high-sampling rate signals by uploading only low-sampling rate signals with a small amount of data to the cloud, thereby overcoming the key bottleneck of achieving timely remote status monitoring. After the data is uploaded, with the support of the powerful parallel computing capabilities of the cloud platform, the reconstruction model deployed in the cloud is used to accurately and quickly reconstruct the low-sampling rate signal into a high-sampling rate signal. The signal reconstructed by the model restores the high-frequency information in the original signal, thereby supporting more accurate status monitoring. In general, the proposed scheme reduces data transmission costs and delays, and ensures the accuracy of cloud signal reconstruction, so as to achieve timely and accurate remote status monitoring of the hydraulic pump.

Description

Technical Field

[0001] The present invention belongs to the field of remote status monitoring, and in particular relates to a method and system for remote status monitoring of a hydraulic pump based on cloud-based intelligent reconstruction of low sampling rate signals. Background Art

[0002] As the core power component of the hydraulic system, the condition monitoring of the hydraulic pump is crucial to ensure the reliable operation of the system. In recent years, with the rapid development of cloud computing technology, remote monitoring technology has become an important means to monitor and analyze the operating status of hydraulic pumps. In order to achieve accurate remote monitoring of the hydraulic pump status, more and more sensors are integrated into the pump body. High-frequency acquisition methods are generally used in industrial sites to obtain multiple types of signals such as vibration, pressure and torque. Although this multi-sensor and high-sampling rate data acquisition strategy provides detailed information for the operation of the pump, it also leads to the generation of massive monitoring data during operation. Although the cloud platform has powerful computing power and can use advanced algorithms to quickly analyze large-scale data from various sensors, due to the bandwidth limitation of the wireless communication channel, when uploading these huge monitoring data from the pump to the cloud, not only will the long data upload time cause monitoring delays, but also incur high transmission costs. Therefore, establishing an efficient data transmission mechanism that can ensure the integrity of status information and the timeliness of transmission is a key issue that needs to be solved in the current field of remote condition monitoring of hydraulic pumps.

[0003] Cloud-based remote condition monitoring technology is very suitable for processing large-scale hydraulic pump data due to its massive storage space and powerful computing power. However, the delay problem when uploading data to the cloud affects the timeliness of condition monitoring, which has become one of the key challenges that the technology needs to solve.

[0004] In order to solve the delay problem caused by large-scale high-sampling rate data transmission, the existing remote status monitoring data transmission method: the existing remote status monitoring methods tend to only upload key feature parameters extracted from the original data, such as root mean square value, maximum value and peak-to-peak value. Although these feature parameters can characterize the operating status of the hydraulic pump in a specific time period to a certain extent, it will inevitably cause information loss and increase the delay of data processing during the feature extraction process, resulting in incomplete status information and reducing the timeliness of data transmission, which can only meet the basic needs of time domain trend analysis and over-limit alarm. In order to ensure the integrity of status information while taking into account the timeliness of transmission, some patents propose a method of compressing and encoding high-sampling rate data using compressed sensing technology in the data acquisition stage, and decompressing and reconstructing the data after uploading it to the cloud. However, the application premise of compressed sensing technology is that the data must have a certain compressibility or sparsity, and noise will have a significant impact on the reconstruction accuracy of the data. Considering that the hydraulic system usually works in an environment with high background noise and low signal-to-noise ratio, this limits the effective application of compressed sensing technology. Therefore, although the existing methods can alleviate the pressure of data transmission to a certain extent, they are still limited in ensuring data integrity and monitoring timeliness.

[0005] The process of reconstructing low-sampling rate data into high-sampling rate data is called upsampling. Upsampling technology usually uses interpolation methods to estimate the values ​​at the newly added data points. Traditional time domain data interpolation methods can be divided into two categories: deterministic methods and random methods. Deterministic methods assume that the data is accurate and estimate the unknown values ​​between known data points by constructing numerical functions. Common methods include nearest neighbor interpolation, linear interpolation, polynomial interpolation, and spline interpolation. These methods are simple and computationally efficient, but may amplify noise and measurement errors in the signal. In contrast, random methods take into account the randomness and uncertainty in the data and use statistical methods to estimate the values ​​of unknown points and their corresponding confidence intervals, such as Kriging interpolation and Bayesian interpolation. However, the implementation of such random methods requires specific statistical assumptions and requires the prior distribution of the data to be known in advance. For time domain signals with strong randomness, it is extremely challenging to accurately obtain prior information. It is worth noting that although traditional interpolation methods can improve the temporal resolution of low-sampling rate data and make the signal appear more continuous and smooth, they cannot generate high-frequency components beyond the spectrum range of the original data. Therefore, traditional interpolation technology cannot provide an effective solution to the problem of high-frequency component loss caused by insufficient sampling rate. Summary of the invention

[0006] The purpose of the present invention is to address the deficiencies in the prior art and to propose a method and system for remote status monitoring of a hydraulic pump based on cloud-based intelligent reconstruction of low sampling rate signals.

[0007] The object of the present invention is to achieve the following technical solution: a method for remote state monitoring of a hydraulic pump based on cloud-based intelligent reconstruction of low sampling rate signals, the method comprising:

[0008] (1) Collect the original status monitoring data of the hydraulic pump;

[0009] (2) Extracting features from the collected original state data to obtain the first feature;

[0010] (3) The first feature is divided into two parts along the channel direction. The first part extracts local features, and the second part performs feature enhancement. The output results of the two parts are spliced ​​along the channel direction to obtain the second feature.

[0011] (4) The first feature is extracted through Swin Transformer layers of different window scales to obtain the third feature;

[0012] (5) After fusing the second feature and the third feature, one-dimensional pixel rearrangement is performed to obtain high sampling rate state monitoring data relative to the original state data.

[0013] Furthermore, the collected original status monitoring data is uploaded to the cloud and upsampled by a reconstruction model deployed in the cloud. The reconstruction model executes steps (2) to (5) to upsample the low sampling rate signal into a high sampling rate signal.

[0014] Furthermore, in step (1), the original state data of the hydraulic pump is a low sampling rate signal collected in real time when the hydraulic pump is in operation.

[0015] Furthermore, in step (2), shallow feature extraction of the original state data is achieved through a single convolutional layer.

[0016] Furthermore, in step (3), local feature extraction is implemented through a single convolutional layer, and high-frequency information is enhanced through a maximum pooling layer, and the extracted high-frequency features are obtained after splicing.

[0017] Furthermore, in step (4), the number of Swin Transformer layers is set according to the required interpolation accuracy, and the window size of the Swin Transformer layer is increased successively to extract multi-scale low-frequency features.

[0018] Furthermore, DropKey is introduced in the attention calculation stage of the Swin Transformer layer to encourage the network to capture the global features of the time series by randomly discarding some keys.

[0019] Furthermore, in step (5), an interactive attention mechanism is used to realize feature fusion, and then an upsampling factor r is set. The number of channels of the feature map is expanded by r times through a convolution operation, and then the feature map is divided into groups of r channels, and the r channels in each group are tiled and rearranged in a periodic manner to obtain a high sampling rate signal.

[0020] On the other hand, the present invention also provides a hydraulic pump remote state monitoring system based on cloud-based intelligent reconstruction of low sampling rate signals, the system comprising:

[0021] The equipment-side state perception layer is composed of hydraulic pumps and various sensors. Various sensors collect the original state monitoring data of the hydraulic pumps at a low sampling rate.

[0022] The edge signal transmission layer, which consists of a data acquisition card and a gateway, is used to upload the data collected by the device status perception layer to the cloud;

[0023] The cloud signal reconstruction layer includes a reconstruction model to realize signal upsampling. The reconstruction model includes a shallow feature extraction module, a deep feature extraction module and a reconstruction module; the shallow feature extraction module is used to extract features from the original state data with a low sampling rate to obtain a first feature; the deep feature extraction module is used to divide the first feature into two parts along the channel direction, the first part extracts local features, and the second part performs feature enhancement. The output results of the two parts are spliced ​​along the channel direction to obtain a second feature; and the first feature is sequentially extracted through SwinTransformer layers of different window scales to obtain a third feature, and is fused with the second feature; the reconstruction module is used to perform one-dimensional pixel rearrangement on the fused feature to obtain high sampling rate state monitoring data;

[0024] The hydraulic pump condition monitoring cloud platform is used to realize data storage, management, analysis, visualization and fault warning.

[0025] Furthermore, the deep feature extraction module includes a high-frequency feature extraction module, a low-frequency feature extraction module and a fusion module;

[0026] The high-frequency feature extraction module includes a parallel local feature extraction branch and a high-frequency enhancement branch, which are used to perform deep feature extraction on the first feature to obtain a second feature;

[0027] The low-frequency feature extraction module includes Swin Transformer layers with different window sizes, which are used to perform deep feature extraction on the first feature to obtain the third feature;

[0028] The fusion module is constructed based on an interactive attention mechanism to achieve feature fusion.

[0029] Beneficial effects of the invention: The invention proposes a new mechanism: using low sampling rate to collect and transmit signals, and completing the reconstruction of high sampling rate signals in the cloud. Compared with high sampling rate data, the amount of data in the same time interval is reduced exponentially by low sampling frequency data, thereby reducing the data transmission burden and transmission delay. However, precisely because the recorded data points are relatively small, the signal collected by the low sampling rate may lose the high-frequency components in the original signal, thereby affecting the accuracy of condition monitoring.

[0030] In order to accurately reconstruct low sampling rate data into high sampling rate data, so as to improve the time resolution of the monitoring signal and restore its high-frequency components, the present invention proposes a signal intelligent reconstruction technology based on the adaptive multi-scale feature modulation Swin Transformer (AMFM-SwinT) model. By using machine learning algorithms, the mapping relationship between low sampling rate signals and high sampling rate signals is learned by training neural networks. While improving the time resolution of low sampling rate signals, it can capture and restore the missing high-frequency components in low sampling rate signals, thereby realizing the accurate reconstruction of the corresponding high sampling rate signals. After sufficient training, the AMFM-SwinT model demonstrates the ability to accurately reconstruct high sampling rate signals.

[0031] On this basis, the present invention only needs to upload low-sampling rate signals with a small amount of data to the cloud, which effectively solves the delay problem caused by large amounts of data transmission. Once these low-sampling rate data arrive at the cloud, with the support of the powerful computing power of the cloud platform and combined with the AMFM-SwinT model deployed on the cloud, the low-sampling rate signals can be quickly and accurately upsampled to high-sampling rate signals. Ultimately, these reconstructed high-sampling rate signals provide sufficient detailed information for subsequent operating status analysis. The present invention achieves accurate and timely remote status monitoring of hydraulic pumps by reducing data transmission delays and ensuring high-quality signal reconstruction. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0033] Figure 1 It is composed of a remote condition monitoring system for a hydraulic pump.

[0034] Figure 2 It is the signal conversion process.

[0035] Figure 3 Swin Transformer model framework for adaptive multi-scale feature modulation.

[0036] Figure 4 The high-frequency feature extraction module in the Swin Transformer model is modulated by adaptive multi-scale features.

[0037] Figure 5 Modulate the Swin Transformer layers in the Swin Transformer model for adaptive multi-scale features.

[0038] Figure 6 Fusion module in Swin Transformer model for adaptive multi-scale feature modulation.

[0039] Figure 7 One-dimensional pixel reordering in the Swin Transformer model for adaptive multi-scale feature modulation. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with the accompanying drawings and implementation examples. It should be understood that the specific implementation examples described herein are only used to explain the present invention and are not used to limit the present invention.

[0041] like Figure 1 and Figure 2 As shown, the present invention provides a method for remote state monitoring of a hydraulic pump based on cloud-based intelligent reconstruction of low-sampling rate signals. When the hydraulic pump is in operation, the original state monitoring data is sampled at a low sampling rate and uploaded to the cloud. The reconstruction model deployed in the cloud is used for upsampling processing, and the low-sampling rate signal is reconstructed into a high-sampling rate signal. These reconstructed high-sampling rate signals provide high-frequency detail information for subsequent operation state analysis. The reconstruction model is an adaptive multi-scale feature modulation Swin Transformer (AMFM-SwinT) model. The signal reconstruction capability of the model is obtained through offline training. In the offline training stage, the low-sampling rate signal of the hydraulic pump state monitoring is used as input, and the corresponding high-sampling rate signal is used as the expected output, and the AMFM-SwinT model is iteratively optimized and trained. After sufficient training, the reconstruction model AMFM-SwinT can not only improve the time resolution of the low-sampling rate signal, but also effectively restore the missing high-frequency details therein, thereby realizing the accurate reconstruction of the high-sampling rate signal of the corresponding hydraulic pump state monitoring, thereby ensuring the accuracy of state monitoring.

[0042] The reconstruction model takes the low sampling rate signal as input and extracts shallow features. These shallow features are input into the parallel high-frequency feature extraction module and low-frequency feature extraction module to extract deep high-frequency and low-frequency features respectively. The extracted low-frequency deep features and high-frequency deep features are then integrated in the fusion module. Finally, the upsampling of the time series is finally achieved through the one-dimensional pixel rearrangement technology to generate a high sampling rate signal.

[0043] like Figure 3 As shown in the figure, the specific composition of the reconstruction model is as follows:

[0044] 1) Shallow feature extraction module

[0045] The low sampling rate signal of the hydraulic pump status monitoring input is passed through a The convolution layer is processed to extract the shallow features as the first features.

[0046] 2) High-frequency feature extraction module

[0047] like Figure 4 As shown, the high-frequency feature extraction module is intended to extract high-frequency features. The high-frequency feature extraction module includes a parallel local feature extraction branch and a high-frequency enhancement branch. The present invention divides the input shallow features into two parts along the channel direction, and the two branches perform parallel processing:

[0048]

[0049] in, is the feature map after shallow feature extraction, They represent the input of the local feature extraction branch and the high-frequency enhancement branch respectively.

[0050] In the local feature extraction branch, the present invention utilizes Convolutional layer to extract local high-frequency features:

[0051]

[0052] in, Indicates that the convolution kernel size is The convolutional layer, represents the ReLU activation function, Represents the output of the local feature extraction branch.

[0053] In the high-frequency enhancement branch, the present invention uses a maximum pooling layer to retain and highlight the high-frequency features extracted from the input. Convolutional layers are used to further enhance high-frequency information:

[0054]

[0055] in, represents the maximum pooling layer, Indicates that the convolution kernel size is The convolutional layer, represents the ReLU activation function, represents the output of the high-frequency enhancement branch.

[0056] Finally, the outputs of the two branches are concatenated along the channel direction to obtain the extracted high-frequency features as the second feature:

[0057]

[0058] in, Represents a concatenation operation.

[0059] 3) Low-frequency feature extraction module

[0060] like Figure 5 As shown in Figure 1, the low-frequency feature extraction module is designed to extract multi-scale low-frequency features as the third feature. According to the required interpolation accuracy, the low-frequency feature extraction module contains three layers of one-dimensional Swin Transformer layers (STL) with different window sizes. The one-dimensional time series input, STL first divides the input into non-overlapping, length M The number of local windows is , reshape the input into Then, the self-attention mechanism of each window is calculated separately. For the i-th local window feature , the query matrix, key matrix and value matrix can be calculated as:

[0061]

[0062]

[0063]

[0064] in, , and is the projection matrix, denote the query matrix, key matrix, and value matrix respectively, and d is the projection dimension.

[0065] In order to prevent the overfitting problem, DropKey is introduced in the attention calculation stage. DropKey encourages the network to capture the global features of the time series by randomly discarding some keys. The random mask matrix of the same shape is , then the self-attention mechanism in the local window is:

[0066]

[0067]

[0068] in, represents element-wise multiplication, is the key matrix after mask operation, and B is the learnable relative position encoding.

[0069] Subsequently, a multilayer perceptron with two fully connected layers is used for further feature transformation. Layer normalization is added before the windowed multi-head self-attention mechanism and the multilayer perceptron module, and residual connections are applied around these two modules. The j-th layer STL process can be expressed as:

[0070]

[0071]

[0072] Among them, LN represents layer normalization operation, W-MSA represents windowed multi-head self-attention mechanism, and MLP represents multi-layer perceptron.

[0073] In order to promote information transfer between windows, a shift window multi-head self-attention mechanism is introduced to move the window forward along the time series. Finally, regular windows and shifted windows are used in pairs to promote the connection between different windows, thereby enhancing the model's ability to capture local and global dependencies.

[0074] Set the three STL windows to different sizes, , and , through the design of multi-level receptive fields, the model's ability to capture information at different time scales is significantly enhanced. In addition, the application of residual connections accelerates model convergence. The entire low-frequency feature extraction process can be expressed as:

[0075]

[0076] in, , , Represents STL with different window sizes respectively.

[0077] 4) Fusion module

[0078] like Figure 6As shown in the figure, in order to better integrate the high-frequency and low-frequency features extracted by the high-frequency feature extraction module and the low-frequency feature extraction module, a fusion module is designed. The output of the high-frequency feature extraction module is used as the query, the output of the low-frequency feature extraction module is used as the key and value, and the interactive attention is calculated to fuse the high-frequency feature (the second feature) and the low-frequency feature (the third feature) to obtain the fused feature. In order to prevent overfitting, DropKey is also introduced in the attention calculation stage.

[0079]

[0080] in, Represents the interactive attention mechanism.

[0081] The output of the interactive attention mechanism is further explored through a lightweight feed-forward network composed of deep separable convolutions. Specifically, deep convolution is first used to extract the spatial local features of each input channel, and then point-by-point convolution is used to achieve information fusion between channels.

[0082]

[0083] Among them, IMLP stands for improved multi-layer perceptron.

[0084] 5) One-dimensional pixel rearrangement

[0085] like Figure 7 As shown in Figure 1, one-dimensional pixel rearrangement is used to achieve upsampling of the time series. The feature map after the low sampling rate signal fusion is taken as input, through The convolution operation transforms it into Feature map of channels , where r is the set upsampling factor. Then, the feature map is divided into groups of r channels, and the r channels in each group are tiled and rearranged in a periodic manner, finally obtaining a size of high sampling rate signal.

[0086]

[0087] in, is the input low sampling rate signal characteristic diagram, Output high sampling rate signal characteristic diagram. Indicates that the output feature map is at position and Channel The value at . Indicates that the input feature map is at position and Channel The value at .

[0088] The reconstruction model training process is as follows:

[0089] The signal reconstruction capability of the AMFM-SwinT model is obtained through offline training. In the offline training stage, the two channels of the data acquisition card are used to synchronously collect the signals output by the sensor at high and low sampling rates. On this basis, the AMFM-SwinT model is iteratively optimized and trained with the low sampling rate signal as input and the corresponding high sampling rate signal as the expected output. Through iterative training, the neural network gradually learns the mapping relationship between the low sampling rate signal and the high sampling rate signal, so that it can capture and restore the high-frequency details missing in the low sampling rate signal.

[0090] After sufficient iterative training, the AMFM-SwinT model not only learned how to infer the original high-frequency components from the low-sampling rate signal, but also accurately reconstructed the input low-sampling rate signal into a high-sampling rate signal in practical applications. This method ensures that the key features of the signal can be effectively preserved and restored even under limited transmission conditions, thereby providing high-quality data reconstruction results.

[0091] On the other hand, corresponding to the above-mentioned embodiment of a hydraulic pump remote state monitoring method based on cloud-based intelligent reconstruction of low sampling rate signals, the present invention also provides an embodiment of a hydraulic pump remote state monitoring system based on cloud-based intelligent reconstruction of low sampling rate signals, the system comprising four parts: a device-side state perception layer, an edge-side signal transmission layer, a cloud-based signal reconstruction layer and a hydraulic pump state monitoring cloud platform.

[0092] The equipment-side status perception layer is composed of a hydraulic pump and various sensors. It uses a variety of sensor elements such as accelerometers, torque sensors, pressure sensors and flow sensors to monitor the key operating status signals such as vibration, torque, pressure and flow of the hydraulic pump in real time at a low sampling rate to obtain the original status monitoring data.

[0093] The edge data transmission layer consists of data acquisition cards and gateways. The data acquisition cards collect and preliminarily process the output signals from various sensors at a low sampling rate, converting continuous analog signals into discrete low-sampling-rate digital signals. Subsequently, the gateway aggregates the multi-source sensor data collected and preprocessed by the data acquisition card, and performs the necessary protocol conversion and data optimization to minimize transmission costs and delays, and finally transmits these low-sampling-rate data securely and reliably to the cloud.

[0094] The cloud data reconstruction layer includes a reconstruction model AMFM-SwinT. With the help of the powerful computing resources of the cloud platform, the low-sampling rate data transmitted to the cloud is efficiently upsampled through the AMFM-SwinT model, and is quickly and accurately reconstructed into high-sampling rate data. The reconstruction model is an adaptive multi-scale feature modulation Swin Transformer (AMFM-SwinT) model. The AMFM-SwinT model consists of three parts: a shallow feature extraction module, a deep feature extraction module, and a reconstruction module. The shallow feature extraction module includes a single convolutional layer, which is used to extract features from the original state data with a low sampling rate to obtain the first feature. The deep feature extraction module consists of a high-frequency feature extraction module, a low-frequency feature extraction module, and a fusion module. The high-frequency feature extraction module includes a parallel local feature extraction branch and a high-frequency enhancement branch. The two branches are used to extract the first feature, and the output results are spliced ​​along the channel direction to obtain the second feature. The low-frequency feature extraction module includes three layers of Swin Transformer layers with different window sizes, which perform feature extraction to obtain the third feature. The fusion module is built based on the interactive attention mechanism, which fuses the second feature and the third feature to obtain the fusion feature. The convolution layer and one-dimensional pixel rearrangement are used to obtain high sampling rate state monitoring data based on fusion features.

[0095] The hydraulic pump condition monitoring platform consists of a data storage and management module, a data analysis and visualization module, an alarm and notification system, a user interface, and a security management and permission control module. It receives high-sampling rate signals reconstructed by the cloud, uses advanced analysis algorithms for in-depth analysis, and accurately and timely feeds back detailed analysis results together with specific operation suggestions to the equipment end, ensuring that on-site operators can respond quickly, thereby achieving fine monitoring and effective management of the hydraulic pump's operating status. In the hydraulic pump condition monitoring platform, the data storage and management module receives and processes high-sampling rate signals restored by the cloud data reconstruction layer. The data analysis and visualization module uses advanced analysis algorithms to comprehensively analyze these high-sampling rate signals. The analysis results include real-time status assessment, predictive maintenance suggestions, and performance optimization tips, and intuitively display data change trends and system health status by generating charts, dashboards, etc. The alarm and notification system automatically triggers the alarm mechanism when an abnormal situation is detected or action is required, and promptly informs relevant personnel. The user interface allows users to view the status information of the hydraulic pump, receive alarm notifications, and perform remote control operations anytime and anywhere. The security management and permission control module ensures that only authorized users can access sensitive information or control systems.

[0096] The above embodiments are used to illustrate the present invention rather than to limit the present invention. Any modification and change made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A remote condition monitoring method for a hydraulic pump based on cloud-based intelligent reconstruction of low sampling rate signals, characterized in that: The method includes: (1) Collecting the original state monitoring data of the hydraulic pump and uploading it to the cloud, upsampling the low sampling rate signal to a high sampling rate signal through the reconstruction model deployed on the cloud, and performing the following steps (2) to (5); (2) Extracting features from the collected original state data to obtain the first feature; (3) The first feature is divided into two parts along the channel direction. The first part extracts local features through a single convolution layer, and the second part is enhanced through a maximum pooling layer to enhance high-frequency information. The output results of the two parts are spliced ​​along the channel direction to obtain the extracted high-frequency features as the second feature; (4) The first feature is extracted through Swin Transformer layers of different window scales to obtain the third feature; (5) After fusing the second and third features, perform one-dimensional pixel rearrangement to obtain high-sampling rate state monitoring data relative to the original state data. Specifically, the interactive attention mechanism is used to achieve feature fusion, and then the upsampling factor is set. r , through the convolution operation, the number of channels of the feature map is expanded r times, and then the feature map is r The channels are divided into groups, and the r The channels are tiled and rearranged in a periodic manner to obtain a high sampling rate signal.

2. According to claim 1, a method for remote status monitoring of a hydraulic pump based on cloud-based intelligent reconstruction of low sampling rate signals is characterized in that: In step (1), the original state data of the hydraulic pump is a low sampling rate signal collected in real time when the hydraulic pump is in operation.

3. The method for remote status monitoring of a hydraulic pump based on cloud-based intelligent reconstruction of low sampling rate signals according to claim 1 is characterized in that: In step (2), shallow feature extraction of the original state data is achieved through a single convolutional layer.

4. The method for remote status monitoring of a hydraulic pump based on cloud-based intelligent reconstruction of low sampling rate signals according to claim 1 is characterized in that: In step (4), the number of Swin Transformer layers is set according to the required interpolation accuracy, and the window size of the Swin Transformer layer is increased successively to extract multi-scale low-frequency features.

5. The method for remote status monitoring of a hydraulic pump based on cloud-based intelligent reconstruction of low sampling rate signals according to claim 4 is characterized in that: DropKey is introduced in the attention calculation stage of the Swin Transformer layer to encourage the network to capture the global features of the time series by randomly discarding some keys.

6. A hydraulic pump remote state monitoring system that implements the hydraulic pump remote state monitoring method based on cloud-based intelligent reconstruction of low sampling rate signals as described in any one of claims 1 to 5, characterized in that: The system includes: The equipment-side state perception layer is composed of hydraulic pumps and various sensors. Various sensors collect the original state monitoring data of the hydraulic pumps at a low sampling rate. The edge signal transmission layer, which consists of a data acquisition card and a gateway, is used to upload the data collected by the device status perception layer to the cloud; The cloud signal reconstruction layer includes a reconstruction model to realize signal upsampling. The reconstruction model includes a shallow feature extraction module, a deep feature extraction module and a reconstruction module; the shallow feature extraction module is used to extract features from the original state data with a low sampling rate to obtain a first feature; the deep feature extraction module is used to divide the first feature into two parts along the channel direction, the first part extracts local features, and the second part performs feature enhancement, and the output results of the two parts are spliced ​​along the channel direction to obtain a second feature; and the first feature is sequentially extracted through Swin Transformer layers of different window scales to obtain a third feature, and is fused with the second feature; the reconstruction module is used to perform one-dimensional pixel rearrangement on the fused feature to obtain high sampling rate state monitoring data; Hydraulic pump condition monitoring cloud platform is used to realize data storage, management, analysis, visualization and fault warning.

7. The hydraulic pump remote status monitoring system according to claim 6, characterized in that: The deep feature extraction module includes a high-frequency feature extraction module, a low-frequency feature extraction module and a fusion module; The high-frequency feature extraction module includes a parallel local feature extraction branch and a high-frequency enhancement branch, which are used to perform deep feature extraction on the first feature to obtain a second feature; The low-frequency feature extraction module includes Swin Transformer layers with different window sizes, which are used to perform deep feature extraction on the first feature to obtain the third feature; The fusion module is constructed based on an interactive attention mechanism to achieve feature fusion.

Images