Directional noise reduction method and system for pressure sensor filtering variable frequency converter interference sources

Abstract

The application relates to the technical field of pressure measurement, and discloses a directional noise reduction method and system for filtering interference sources of a pressure sensor frequency converter, which can effectively eliminate nonlinear interference generated by the frequency converter. The method comprises the following steps: inputting synchronously collected pressure sensor voltage data with noise and a frequency converter principal component matrix into two 1D convolution layers based on a CNN structure respectively to extract local features; inputting the two extracted local features into two Bi-LSTM networks respectively for processing, and then outputting a sensor voltage time sequence vector with noise and a frequency converter principal component time sequence vector; calculating the correlation between the sensor voltage time sequence vector with noise and the frequency converter principal component time sequence vector based on a cross-attention mechanism to obtain a context vector fused with the frequency converter features; and inputting the vector obtained by splicing the sensor voltage time sequence vector with noise and the context vector fused with the frequency converter features into a full connection layer to obtain pressure sensor voltage data after noise reduction.

Description

Technical Field

[0001] This invention relates to the field of pressure measurement technology, and in particular to a method and system for targeted noise reduction of pressure sensors to filter interference sources from frequency converters. Background Technology

[0002] In the industrial sector, especially in heavy industries such as steel, frequency converters are widely used. Frequency converters are common devices that generate electromagnetic noise, and the interference they generate during operation can affect surrounding electronic equipment, including the actual output of various sensors.

[0003] For example, during the cleaning process of steel plates, a frequency converter is used to adjust the high-pressure pump to generate high pressure in the pipeline, which drives the abrasive jet to clean the surface of the steel plate. In order to ensure that the cleaning quality meets the processing requirements, it is necessary to monitor the pressure in the high-pressure pipeline in real time. However, when the frequency converter is working, the voltage signal output by the pressure sensor is unstable and fluctuates irregularly, which makes it impossible to reflect the true pressure value.

[0004] To reproduce the true pressure value, filtering methods are typically used to process the pressure sensor output. Traditional data filtering methods, such as low-pass filtering, high-pass filtering, median filtering, and Kalman filtering, have the following shortcomings: First, filtering is designed based on prior knowledge of the frequency or statistical characteristics of the signal and noise. For example, low-pass filters assume that the signal frequency is low and the noise frequency is high, while Kalman filtering requires knowledge of the statistical models of the system and noise. Once the filter parameters, such as cutoff frequency, order, system matrix, and noise covariance matrix, are designed, they are usually fixed and lack adaptability. When the noise characteristics in the actual application scenario change, or when the prior knowledge is inaccurate, the data processing effect will significantly decrease. Second, filters are essentially linear systems. They perform well in processing linearly superimposed noise and linear system signals. However, interference generated by frequency converters is often complex and nonlinear, such as harmonic interference and switching noise. There is a complex nonlinear relationship between these noises and the operating state of the frequency converter, and linear filters cannot effectively model and eliminate these nonlinear interferences. Summary of the Invention

[0005] The purpose of this invention is to disclose a directional noise reduction method and system for filtering interference sources from frequency converters using pressure sensors, so as to effectively eliminate nonlinear interference generated by frequency converters.

[0006] To achieve the above objectives, the method disclosed in this invention implements the following steps based on a network model: Step S1: Acquire a set of noisy voltage data from the synchronously acquired pressure sensor and a set of characteristic data from the frequency converter; Step S2: Determine the dimensionality reduction matrix of the inverter feature variables based on principal component analysis, and obtain the principal component matrix after dimensionality reduction of the feature data based on the dimensionality reduction matrix of the inverter feature variables. Step S3: Input the noisy voltage data of the pressure sensor and the principal component matrix into two independent 1D (One-Dimensional) convolutional layers based on CNN (Convolutional Neural Networks) structure to extract local features. Step S4: Input the two sets of features extracted independently by the 1D convolutional layer into two independent Bi-LSTM (Bidirectional Long Short Term Memory) networks for group processing, model the global temporal dependency relationship, and output the sensor noisy voltage time-series vector and the inverter principal component time-series vector, respectively. Step S5: For the different outputs of the two Bi-LSTM networks, based on the cross-attention mechanism, calculate the correlation between the sensor noisy voltage time-series vector and the inverter principal component time-series vector and implement weight adjustment. The output is a context vector that integrates inverter features. Step S6: Concatenate the sensor's noisy voltage timing vector and the context vector that incorporates the inverter's features; Step S7: Input the spliced ​​vector into the fully connected layer to obtain a denoised pressure sensor voltage data with a 1-dimensional output.

[0007] Preferably, during the training of the network model in steps S1 to S7, the sample data includes: synchronously acquired pressure sensor voltage data with noise, inverter characteristic data, and actual pressure values ​​that are not disturbed by noise. The actual pressure values ​​that are not disturbed by noise are converted into actual voltage outputs for use in loss function calculation.

[0008] Preferably, the actual pressure value, which is not affected by noise, is collected using a mechanical pressure gauge.

[0009] Preferably, the process of processing sample data and real-time detection data further includes: first cleaning and standardizing the collected data, and then performing local feature extraction.

[0010] Preferably, step S2 specifically includes: Step S21: Obtain the characteristic data of the standardized frequency converter. , For the sample size, R represents the characteristic data dimension of the frequency converter, and R denotes a real number. Step S22, Calculation covariance matrix Capture the linear relationship between different characteristic variables of the frequency converter: ; Step S23: Adjust the covariance matrix Perform eigenvalue decomposition to obtain eigenvalues. and eigenvectors , ;in, , , , The feature space represents the feature vectors formed by the inverter. This represents the eigenvalues ​​arranged in descending order; Step S24: Calculate the cumulative contribution rate based on the magnitude of the eigenvalues. , Select cumulative contribution rate Before reaching the set threshold The dimensionality reduction matrix of the inverter's characteristic variables, consisting of the principal component eigenvalues ​​and their corresponding eigenvectors, ; Step S25: Collect the standardized inverter data Dimensionality reduction matrix using inverter characteristic variables By projection, the principal component matrix of the inverter's characteristic data is obtained:

[0011] in, This represents the principal component matrix after dimensionality reduction of the inverter's feature data.

[0012] Preferably, the method of the present invention is applied to a steel plate cleaning scenario, wherein the frequency converter is used to adjust the high-pressure pump to generate high pressure in the high-pressure pipeline to drive the abrasive to perform jet cleaning on the steel plate surface, and the pressure sensor is used to monitor the pressure in the high-pressure pipeline in real time. Based on the aforementioned scenario, the characteristic data of the frequency converter collected includes any one or any combination of operating frequency, bus voltage, output voltage, output current, output power, output torque, and each major harmonic.

[0013] To achieve the above objectives, the present invention also discloses a directional noise reduction system for filtering interference sources from frequency converters using pressure sensors, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the above-described method.

[0014] Furthermore, the system of the present invention includes: A data processing subsystem for completing the functions of data alignment, abnormal data handling, data standardization, and principal component extraction; The model training subsystem has functions for data storage and loading, model parameter setting, model training and evaluation. It is used to load datasets in a dynamic batch loading manner, dynamically adjust model parameters according to loss function and optimizer, and complete the training of denoising model. The online noise reduction subsystem is used to add real-time pressure sensor signals with noise and inverter characteristic data to the end of a time-series data window through a sliding window, delete the earliest set of data in the sequence, and input it into the deployed noise reduction model to output the noise-reduced pressure sensor signal.

[0015] Preferably, the system of the present invention further includes a dynamic monitoring subsystem for the characteristic principal components of the frequency converter, which is used to track and analyze whether the characteristic principal components of the frequency converter have changed. If they have changed, new sample data is collected again based on the same method to train, deploy and perform online noise reduction processing of the noise reduction model.

[0016] The present invention has the following beneficial effects: This method explicitly utilizes information about the inverter's interference sources. Based on PCA (Principal Components Analysis) and a cross-attention mechanism, it can automatically extract noise and signal features from the data and establish a complex relationship between noise and interference sources. It considers the sources and variation patterns of interference, thus better adapting to complex and unknown noise environments. It overcomes the limitations of traditional linear filters and the difficulty in effectively modeling and eliminating nonlinear noise. Furthermore, it uses a combination of long short-term memory to analyze the time dependence between data samples, making it more suitable than traditional filtering methods for handling time-series data sequences with inaccurate sensor data caused by noise. This improves the stability of filtering performance.

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is an architecture diagram of the noise reduction model disclosed in Embodiment 1 of the present invention.

[0019] Figure 2 This is a schematic diagram of the core steps in the directional noise reduction method for filtering interference sources from frequency converters using pressure sensors, as disclosed in Embodiment 2 of the present invention.

[0020] Figure 3 This is part of the data in Embodiment 1 of the present invention. Detailed Implementation

[0021] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0022] Example 1 This embodiment discloses a directional noise reduction method for filtering interference sources from frequency converters using a pressure sensor, which is applied in a steel plate cleaning scenario. The frequency converter is used to adjust a high-pressure pump to generate high pressure in a high-pressure pipeline to drive abrasives to perform jet cleaning on the surface of the steel plate, and the pressure sensor is used to monitor the pressure in the high-pressure pipeline in real time.

[0023] In this embodiment, a multi-layered deep network model is established, combining PCA and attention mechanisms to achieve noise reduction of the pressure sensor output, referring to... Figure 1 The diagram shows the architecture of the noise reduction model; specifically, it includes the following: 1) Synchronously collect the noisy voltage data of the pressure sensor, the characteristic data of the frequency converter, and the actual pressure value that is not disturbed by noise, and perform data processing.

[0024] 2) Using Principal Component Analysis (PCA), the dimensionality-reduced matrix of the inverter's characteristic variables and the principal component matrix after dimensionality reduction of the characteristic data are obtained. The characteristic variables refer to the selected inverter specifications such as operating frequency, bus voltage, output voltage, etc., while the characteristic data refers to the corresponding numerical values ​​of the characteristic variables.

[0025] 3) Local Feature Network Layer: The front layer of the noise reduction model consists of two 1D convolutional layers based on the CNN structure (i.e., the CON1D CNN network layer in the figure). The noisy voltage data of the pressure sensor and the principal component matrix of the frequency converter are respectively input into the convolutional layer for local feature extraction.

[0026] 4) Temporal Feature Network Layer: The intermediate layer of this model consists of two Bi-LSTM network layers. The features extracted by the 1D convolutional layer are input into the two Bi-LSTM networks for processing to model the global temporal dependencies. The outputs are the sensor noisy voltage temporal vector and the inverter principal component temporal vector, respectively.

[0027] 5) Attention layer: After the intermediate layer of this model is the attention layer. Based on the attention mechanism, it calculates the correlation of the above time-series vectors and adjusts the weights for the output of the Bi-LSTM network. The output is a context vector that incorporates the features of the frequency converter.

[0028] 6) Feature splicing and fusion layer: It is located after the attention layer. Its input is the sensor's noisy voltage time-series vector and the context vector that fuses the inverter features. Its output is the splicing vector.

[0029] 7) Dense Fully Connected Layer: The output layer of this model is a fully connected layer with a 1-dimensional output. Its input is the spliced ​​vector output by the feature splicing and fusion layer, and its output is the denoised pressure sensor voltage data.

[0030] In this embodiment, when the noise reduction model is performing real-time data noise reduction, the input data is the noisy voltage data of the pressure sensor over a period of time and the reduced principal component matrix of the frequency converter. The earliest input data is discarded and the latest input is introduced in a sliding window-based manner to obtain the latest output data.

[0031] More specifically, the method in this embodiment is divided into the following steps: Step A: Under the condition of inverter operation, collect inverter characteristic data, pressure sensor output voltage signal, and simultaneously collect undisturbed actual pressure value. The specific process is as follows: (1) Within the set time step T, the step size is increased according to the set frequency. The frequency converter will start from the set starting frequency. Adjust to Simultaneously adjust the actual pressure value P of the pipeline; collect characteristic data of the frequency converter. Noisy output voltage of pressure sensor Among them, the characteristic data of the frequency converter This can include operating frequency, bus voltage, output voltage, output current, output power, output torque, and major harmonics (which can be freely selected; the third harmonic is used in this embodiment, but the third, fifth, and seventh harmonics are commonly used in industry. In this embodiment, the dimensionality of the characteristic variables will be reduced after principal component analysis, so some variables may be retained or eliminated). This data is collected during the data acquisition process. and Simultaneously collect the voltage value corresponding to the actual pressure in the pipeline. : ; in, and This represents the extreme value of the measurement range of the pressure sensor. and This represents the extreme value of the pressure sensor's output range. This represents the actual pressure value of the pipeline; this value is a noise-free signal. For example, if the output of a certain sensor is 0~10V, the corresponding pressure range is 0~2MPa, then a pressure within the 0~2MPa range... The corresponding voltage can be calculated using the above formula, which is essentially a linear interpolation.

[0032] In this embodiment, the dimensional distribution of the collected data is as follows: ;in, It is a sensor output voltage signal with noise. It is the voltage signal corresponding to the actual pressure in the pipeline. It is the output signal of the frequency converter. It is the number of dimensions of the selected inverter feature data. That is the number of samples.

[0033] (3) Check the data samples in order of timestamp. , and If outliers, missing values, or invalid data are found, the data is cleaned according to timestamp order, and then... , and Z-score standardization is used to transform the data into a distribution with a mean of 0 and a standard deviation of 1. The processing method is as follows: ;in, represent , and Any data in the, yes The mean, yes The mean squared error, yes Standardized data, recorded , and They are respectively , and Standardized data.

[0034] Step B: Standardize the inverter characteristic data PCA modeling is performed for principal component analysis. The specific process is as follows: Choose the dimension of the inverter output feature data, assuming the inverter's feature data dimension is... ,but For example, if the characteristic variables of the frequency converter are selected as {operating frequency, bus voltage, output voltage, output current, output power, output torque, third harmonic voltage, and third harmonic current}, then the dimension of the characteristic data is... , It is an N×8 dimensional matrix.

[0035] calculate covariance matrix Capture the linear relationship between different characteristic variables of the frequency converter: .

[0036] For covariance matrix Perform eigenvalue decomposition to obtain eigenvalues. and eigenvectors ;in: ; , , , The feature space represents the feature vectors formed by the inverter. This represents the eigenvalues ​​arranged from largest to smallest.

[0037] Calculate the cumulative contribution rate based on the magnitude of the eigenvalues. , Select cumulative contribution rate Before reaching the set threshold The dimensionality reduction matrix consisting of the principal component eigenvalues ​​and their corresponding eigenvectors When the CCR reaches a preset threshold, it means that only the first k eigenvalues ​​and their corresponding eigenvectors are retained to form the dimensionality reduction matrix of the inverter's characteristic variables. ,Right now: .

[0038] The collected standardized inverter data Dimensionality reduction matrix using inverter characteristic variables By projection, the principal component matrix of the inverter's characteristic data is obtained: ;in, This represents the principal component matrix after dimensionality reduction of the inverter's feature data.

[0039] Step C: Constructing a deep learning noise reduction model: A hierarchical deep learning network is used as the core architecture, referring to... Figure 1 The model consists of two CNN 1D convolutional layers, two Bi-LSTM temporal network layers, one attention layer, one feature concatenation and fusion layer, and one fully connected layer. It is used to model the temporal dependencies of pressure sensor signals and incorporates an attention mechanism for adaptive denoising. Specifically: The network has a 1D CNN convolutional layer at the beginning as a local temporal feature extraction layer, which is used to extract noisy pressure sensor signals. The principal component matrix of the frequency converter after dimensionality reduction by principal component analysis Perform a 1D convolution operation to obtain and .

[0040] The intermediate layers of the network consist of two bidirectional LSTM networks (Bi-LSTM) that process the temporal features extracted by the convolutional layers, capturing long-term dependencies in the sequence and leveraging bidirectionality to consider both past and future information. and The inputs are fed into the Bi-LSTM layer respectively, and the outputs are respectively and .

[0041] The middle layer is followed by the attention layer, which learns how to adapt to the characteristics of the frequency converter. Let's focus on the signal characteristics of noisy sensors. Different parts of the attention mechanism are used to more effectively reduce noise. The expression for the attention mechanism is: ;in: It is a query vector, and its values ​​are the output data of the Bi-LSTM. , It is a key vector, with values... , It is a value vector, and its values ​​are the same as those of the key vector. , yes The dimension of the vector will and Perform multiplication to obtain the output of the attention layer. matrix.

[0042] Following the attention layer is a feature concatenation and fusion layer, whose input is... and the context that incorporates inverter characteristics Its output is the result of concatenating the two. matrix.

[0043] The network's output layer is a fully connected layer, which receives the output of the attention layer. The matrix is ​​mapped to the final output dimension, which is 1D, i.e., the restored denoised pressure sensor signal.

[0044] In this embodiment, sample data is first constructed based on the noisy voltage data from the synchronously acquired pressure sensor, the characteristic data of the frequency converter, and the actual pressure value undisturbed by noise for training a noise reduction model. The actual pressure value undisturbed by noise needs to be converted into an actual voltage output for use in loss function calculation. After the noise reduction model training is completed, it can be deployed to perform online noise reduction processing on the real-time acquired data from the pressure sensor based on the principal component features of the synchronous frequency converter. Optionally, in this embodiment, the actual pressure value undisturbed by noise is acquired using a mechanical pressure gauge.

[0045] Some of the data and experimental results used in this embodiment are as follows: Figure 3 As shown, by Figure 3 The calculations for some of the data are as follows:

[0046] Compared with traditional filtering methods, the method of this invention can significantly reduce errors and peak errors, especially in terms of nonlinear interference of the signal (third harmonic current and output frequency fluctuation in this example), which is more superior. This is because the model learns the working characteristics of the frequency converter and makes reasonable use of cross attention, which can accurately identify and suppress complex interference from the frequency converter.

[0047] Example 2 This embodiment discloses a directional noise reduction system for filtering interference sources from a frequency converter using a pressure sensor. The system includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the computer program, it implements the method disclosed in the above embodiment. Its core includes at least the following: Figure 2 The following steps are shown: Step S1: Acquire a set of noisy voltage data (or "noisy pressure sensor data") from the synchronously acquired pressure sensor and a set of characteristic data from the frequency converter.

[0048] Step S2: Determine the dimensionality reduction matrix of the inverter's feature variables based on principal component analysis (PCA). Then, obtain the principal component matrix after dimensionality reduction of the feature data based on this matrix. PCA extracts the inverter noise features into low-dimensional principal components, reducing redundancy and improving training efficiency and real-time performance of online noise reduction.

[0049] Step S3: Input the noisy voltage data and principal component matrix (or "dimensionality-reduced inverter principal component data") of the collected pressure sensor into two independent 1D convolutional layers based on CNN structure for local feature extraction.

[0050] Step S4: Input the two sets of features extracted independently by the 1D convolutional layer into two independent Bi-LSTM networks for group processing, model the global temporal dependency relationship, and output the sensor noisy voltage temporal vector and the inverter principal component temporal vector, respectively.

[0051] In the two steps above, CNN extracts local patterns and Bi-LSTM captures temporal dependencies. The fusion of the two types of information is more comprehensive and avoids missing important features in a single path.

[0052] Step S5: For the different outputs of the two Bi-LSTM networks, based on the cross-attention mechanism, calculate the correlation between the sensor's noisy voltage time-series vector and the inverter's principal component time-series vector and adjust the weights. The output is a context vector that incorporates inverter features. This allows the network to guide attention based on the inverter's principal components, enhancing the model's sensitivity to interference patterns. During noise reduction, it can more effectively identify which components are harmful and which are useful, achieving targeted denoising.

[0053] Step S6: Concatenate the sensor's noisy voltage time-series vector and the context vector fused with inverter features. By fusing information from different dimensions, the model's nonlinear modeling capability and complex pattern capture capability are improved.

[0054] Step S7: Input the spliced ​​vector into the fully connected layer to obtain a denoised pressure sensor voltage data with a 1-dimensional output.

[0055] Furthermore, the system in this embodiment includes: a data processing subsystem for aligning collected data, handling abnormal data, standardizing data, and extracting principal components; a model training subsystem with data storage and loading, model parameter setting, model training, and evaluation functions, used to load datasets through dynamic batch loading, dynamically adjust model parameters according to loss function and optimizer, and complete the training of the denoising model; and an online denoising subsystem for adding real-time collected pressure sensor signals with noise and inverter feature data to the end of a time series data window through a sliding window, deleting the earliest set of data in the sequence, and inputting it into the deployed denoising model to output the denoised pressure sensor signal.

[0056] Furthermore, the system in this embodiment also includes a dynamic monitoring subsystem for the characteristic principal components of the frequency converter, which is used to track and analyze whether the characteristic principal components of the frequency converter have changed. If they have changed, new sample data is collected again based on the same method to train, deploy and perform online noise reduction processing of the noise reduction model.

[0057] In summary, the core of the methods and systems disclosed in the embodiments of this invention lies in the following: Noisy voltage data from the pressure sensor and the principal component matrix of the frequency converter are simultaneously acquired and input into two 1D convolutional layers based on a CNN structure for local feature extraction; the two extracted local features are then processed by two Bi-LSTM networks, outputting the noisy voltage time-series vector from the sensor and the principal component time-series vector from the frequency converter, respectively; the correlation between the noisy voltage time-series vector from the sensor and the principal component time-series vector from the frequency converter is calculated based on a cross-attention mechanism to obtain a context vector incorporating frequency converter features; the concatenated vector of the noisy voltage time-series vector from the sensor and the context vector incorporating frequency converter features is input into a fully connected layer to obtain the denoised pressure sensor voltage data. In contrast, traditional self-attention denoising models are simple but cannot model specific interference, relying solely on inferring noise characteristics from the signal itself. Their effectiveness is limited under strong interference environments and they are prone to failure after temporal changes in noise. Therefore, the methods and systems disclosed in this embodiment have at least the following beneficial effects: This invention explicitly utilizes information about inverter interference sources. Based on PCA and cross-attention mechanisms, it automatically extracts noise and signal features from data and establishes a complex relationship between noise and interference sources. It considers the sources and variation patterns of interference, thus better adapting to complex and unknown noise environments. It overcomes the limitations of traditional linear filters and their difficulty in effectively modeling and eliminating nonlinear noise. Furthermore, it employs a combination of long short-term memory (LSTM) and other methods to analyze the time dependencies between data samples, making it more suitable than traditional filtering methods for handling time-series data sequences with inaccurate sensor data caused by noise. This improves the stability of filtering performance. In other words, this invention uses external interference data for specific modeling, accurately eliminating nonlinear noise caused by inverters while maintaining effective signals, showing significant advantages, especially in complex electromagnetic environments.

[0058] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for targeted noise reduction of a pressure sensor filtering frequency converter interference source, characterized in that, Based on the network model, the following steps are implemented: Step S1: Acquire a set of noisy voltage data from the synchronously acquired pressure sensor and a set of characteristic data from the frequency converter; Step S2: Determine the dimensionality reduction matrix of the inverter's characteristic variables based on principal component analysis, and obtain the principal component matrix after dimensionality reduction of the characteristic data based on the dimensionality reduction matrix of the inverter's characteristic variables; specifically including: Step S21: Standardize the acquired inverter feature data to obtain standardized inverter feature data. , For the sample size, R represents the characteristic data dimension of the frequency converter, and R denotes a real number. Step S22, Calculation covariance matrix Capture the linear relationship between different characteristic variables of the frequency converter: ; Step S23: Adjust the covariance matrix Perform eigenvalue decomposition to obtain eigenvalues. and eigenvectors , ;in, , , , The feature space represents the feature vectors formed by the inverter. This represents the eigenvalues ​​arranged in descending order; Step S24: Calculate the cumulative contribution rate based on the magnitude of the eigenvalues. , Select cumulative contribution rate Before reaching the set threshold The dimensionality reduction matrix of the inverter's characteristic variables, consisting of the principal component eigenvalues ​​and their corresponding eigenvectors, ; Step S25: Collect the standardized inverter data Dimensionality reduction matrix using inverter characteristic variables By projection, the principal component matrix of the inverter's characteristic data is obtained: ； in, This represents the principal component matrix after dimensionality reduction of the inverter's feature data; Step S3: Input the noisy voltage data of the pressure sensor and the principal component matrix into two independent 1D convolutional layers based on CNN structure for local feature extraction. Step S4: Input the two sets of features extracted independently by the 1D convolutional layer into two independent Bi-LSTM networks for group processing, model the global temporal dependency relationship, and output the sensor noisy voltage temporal vector and the inverter principal component temporal vector, respectively. Step S5: For the different outputs of the two Bi-LSTM networks, based on the cross-attention mechanism, calculate the correlation between the sensor noisy voltage time-series vector and the inverter principal component time-series vector and implement weight adjustment. The output is a context vector that integrates inverter features. Step S6: Concatenate the sensor's noisy voltage timing vector and the context vector that incorporates the inverter's features; Step S7: Input the spliced ​​vector into the fully connected layer to obtain a denoised pressure sensor voltage data with a 1-dimensional output.

2. The directional noise reduction method for filtering interference sources from frequency converters using pressure sensors according to claim 1, characterized in that, During the training of the network model in steps S1 to S7 above, the sample data includes: noisy voltage data of the pressure sensor collected synchronously, characteristic data of the frequency converter, and actual pressure values ​​that are not disturbed by noise. The actual pressure values ​​that are not disturbed by noise are converted into actual voltage outputs for use in loss function calculation.

3. The directional noise reduction method for filtering interference sources from frequency converters using pressure sensors according to claim 2, characterized in that, The actual pressure values, undisturbed by noise, were collected using a mechanical pressure gauge.

4. The directional noise reduction method for filtering interference sources from frequency converters using pressure sensors according to any one of claims 1 to 3, characterized in that, The processing of sample data and real-time detection data also includes: The collected data is first cleaned and standardized, and then local feature extraction is performed.

5. The directional noise reduction method for filtering interference sources from frequency converters using pressure sensors according to any one of claims 1 to 3, characterized in that, The method is applied to a steel plate cleaning scenario, wherein the frequency converter is used to adjust the high-pressure pump to generate high pressure in the high-pressure pipeline to drive the abrasive to perform jet cleaning on the steel plate surface, and the pressure sensor is used to monitor the pressure in the high-pressure pipeline in real time.

6. The method for targeted noise reduction of pressure sensor filtering inverter interference sources according to any one of claims 1 to 3, characterized in that, The characteristic data of the frequency converter collected in step S1 includes any one or any combination of the following: operating frequency, bus voltage, output voltage, output current, output power, output torque, and major harmonics.

7. A directional noise reduction system for filtering interference sources from a frequency converter using a pressure sensor, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the method described in any one of claims 1 to 6.

8. The directional noise reduction system for filtering interference sources from frequency converters using pressure sensors according to claim 7, characterized in that, include: A data processing subsystem for completing the functions of data alignment, abnormal data handling, data standardization, and principal component extraction; The model training subsystem has functions for data storage and loading, model parameter setting, model training and evaluation. It is used to load datasets in a dynamic batch loading manner, dynamically adjust model parameters according to loss function and optimizer, and complete the training of denoising model. The online noise reduction subsystem is used to add the real-time acquired pressure sensor signal with noise and the characteristic data of the frequency converter to the end of a time-series data window through a sliding window, and after deleting the earliest set of data in the sequence, input it into the deployed noise reduction model and output the noise-reduced pressure sensor signal.

9. The directional noise reduction system for filtering interference sources from frequency converters using pressure sensors according to claim 8, characterized in that, Also includes: The characteristic principal component dynamic monitoring subsystem of the frequency converter is used to track and analyze whether the characteristic principal components of the frequency converter have changed. If they have changed, new sample data is collected again based on the same method to train, deploy and perform online noise reduction processing of the noise reduction model.

Images