Method, system and electronic device for identifying the rotational speed of a rotating mechanical device

Abstract

This application proposes a method, system, and electronic device for identifying the rotational speed of rotating machinery, relating to the field of rotational speed measurement. The method includes: acquiring raw vibration waveform data and performing a discrete Fourier transform to obtain spectral data containing multiple frequency components arranged in frequency order; assigning voting weights to each of the multiple frequency components, including itself, as a voter; the sum of the voting weights cast by each voter is a fixed value, while the sum of the weights obtained by each frequency component as a voted-for item is a non-fixed value, where the voting weights characterize the correlation strength between the voter and the voted-for item; identifying the target rotational frequency component from the multiple frequency components based on the sum of the voting weights obtained by each frequency component as a voted-for item, and thus determining the real-time rotational speed of the rotating machinery. This scheme can accurately identify the real-time rotational speed of rotating machinery under complex operating conditions without increasing hardware costs.

Description

Technical Field

[0001] This application relates to the field of rotational speed measurement, and more specifically, to a method, system, and electronic device for identifying the rotational speed of rotating machinery. Background Technology

[0002] The rotational speed of rotating machinery (such as fans, pumps, and motors) is a fundamental parameter for equipment condition monitoring and fault diagnosis. For example, there is a definite mathematical relationship between the fault characteristic frequency of rolling bearings and the equipment rotational speed; accurately obtaining the rotational speed is a prerequisite for accurate fault diagnosis.

[0003] The traditional method for obtaining the rotational speed of rotating machinery is to install a dedicated speed sensor. Speed ​​sensors include contact and non-contact sensors, which accurately obtain the rotational speed by detecting the periodic changes in the equipment. However, installing a dedicated speed sensor has drawbacks such as high hardware costs, installation costs, and subsequent maintenance costs.

[0004] To save on the aforementioned costs, a scheme for identifying rotational speed from vibration data using algorithms has been developed in this field. Specifically, this involves performing a Fourier transform on the acquired raw vibration waveform data to obtain spectral data, and then employing a local peak detection strategy to find the target rotational frequency component representing the rotational speed. However, when the operating conditions of rotating machinery are complex, such as with speed fluctuations, load changes, or significant environmental interference, existing schemes for identifying rotational speed from vibration data using these conventional algorithms have low accuracy. They struggle to accurately identify the target rotational frequency component and real-time rotational speed of the rotating machinery, thus affecting the reliability of subsequent fault diagnosis.

[0005] Therefore, how to accurately and effectively identify the rotational speed of rotating machinery with complex operating conditions without increasing additional hardware costs has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] The purpose of this application is to provide a method, system, and electronic device for identifying the rotational speed of rotating machinery, which can accurately and effectively identify the rotational speed of rotating machinery with complex operating conditions without increasing additional hardware costs.

[0007] This application is implemented as follows: In a first aspect, this application provides a method for identifying the rotational speed of rotating machinery, comprising the following steps: acquiring raw vibration waveform data of the rotating machinery; performing a discrete Fourier transform on the raw vibration waveform data to obtain corresponding spectrum data, the spectrum data containing multiple frequency components arranged in frequency order; assigning voting weights to each of the multiple frequency components, including itself, as a voter, wherein the sum of the voting weights cast by each voter is a fixed value, and the sum of the weights obtained by each frequency component as a voted-for object is a non-fixed value, the voting weights representing the correlation strength between the voter and the voted-for object; identifying a target rotational frequency component from the multiple frequency components based on the sum of the voting weights obtained by each frequency component as a voted-for object; and determining the real-time rotational speed of the rotating machinery based on the target rotational frequency component.

[0008] Secondly, this application provides a system for identifying the rotational speed of rotating machinery, comprising: a data acquisition module for acquiring raw vibration waveform data of the rotating machinery; a spectrum transformation module for performing a discrete Fourier transform on the raw vibration waveform data to obtain corresponding spectrum data, the spectrum data containing multiple frequency components arranged in frequency order; a voting allocation module for assigning voting weights to all frequency components, including itself, by treating each of the multiple frequency components as a voter, wherein the sum of the voting weights cast by each voter is a fixed value, and the sum of the weights obtained by each frequency component as a voted-for object is a non-fixed value, the voting weights representing the correlation strength between the voter and the voted-for object; a rotational frequency identification module for identifying a target rotational frequency component from the multiple frequency components based on the sum of the voting weights obtained by each frequency component as a voted-for object; and a rotational speed determination module for determining the real-time rotational speed of the rotating machinery based on the target rotational frequency component.

[0009] Thirdly, this application provides an electronic device including a memory for storing one or more programs; a processor; and, when the one or more programs are executed by the processor, implementing the method as described in any one of the first aspects above.

[0010] Compared with the prior art, this application has at least the following advantages or beneficial effects: This application proposes a method for identifying the rotational speed of rotating machinery. Each frequency component is assigned a dual role: both a voter and a voted-for component. The target rotational frequency component is determined through a voting process involving all frequency components. Specifically, the sum of the votes cast by each voter is constrained to a fixed value, while the sum of the weights received by each frequency component as a voted-for component is non-fixed. This design allows frequency components strongly correlated with the target rotational frequency component to receive higher vote weights from other frequency components, thus significantly enhancing their performance in the global voting process. Conversely, noise and other interfering frequency components are effectively suppressed due to their lower vote weights. Furthermore, since this scheme achieves rotational speed identification entirely based on data collected by vibration sensors, it eliminates the need for dedicated rotational speed sensors. Therefore, it achieves high-accuracy rotational speed identification for rotating machinery operating under complex conditions without increasing additional hardware costs, providing a reliable speed basis for subsequent fault diagnosis. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart of an embodiment of a method for identifying the rotational speed of rotating machinery according to this application; Figure 2 This is a schematic diagram of the frequency conversion identification model in one embodiment of this application; Figure 3 This is a structural block diagram of an embodiment of a system for identifying the rotational speed of rotating machinery according to this application; Figure 4 This is a structural block diagram of an electronic device provided in an embodiment of this application.

[0013] Icons: 101, Data Acquisition Module; 102, Spectrum Transformation Module; 103, Voting Allocation Module; 104, Frequency Recognition Module; 105, Rotation Speed ​​Determination Module; 201, Processor; 202, Memory; 203, Communication Interface. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0015] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0016] In developing this application, the inventors discovered that while adding a dedicated speed sensor can accurately acquire the rotational speed, it suffers from high hardware, installation, and maintenance costs. While identifying the rotational speed from vibration data using conventional algorithms can save these costs, existing algorithms struggle to accurately identify the target rotational frequency component and real-time rotational speed under complex operating conditions of rotating machinery, such as speed fluctuations, load changes, or significant environmental interference. The inventors found that existing algorithms typically employ a local peak detection strategy, focusing on high-energy local features in the spectral data while neglecting the role of other frequency components in frequency identification. This leads to identification failure when the true rotational frequency component is obscured by interference.

[0017] Based on this, this application proposes a method for identifying the rotational speed of rotating machinery. It employs a unique global voting mechanism that allows frequency components strongly correlated with the target rotational frequency to receive higher voting weights from other frequency components, thus significantly enhancing their performance, while suppressing interfering frequency components such as noise. Therefore, this application can accurately and effectively identify the real-time rotational speed of rotating machinery under complex operating conditions without increasing hardware costs.

[0018] After introducing the basic principles of this application, various non-limiting embodiments of this application will be described in detail below with reference to the accompanying drawings. Unless otherwise specified, the various embodiments and features described below can be combined with each other.

[0019] Please see Figure 1 The method for identifying the rotational speed of this rotating machinery includes the following steps: Step S101: Obtain the original vibration waveform data of the rotating machinery.

[0020] It should be noted that the raw vibration waveform data can be collected by vibration sensors already installed on key components of rotating machinery. These sensors acquire vibration signals generated during equipment operation, which are then converted into corresponding electrical signals and converted from analog to digital to obtain a discrete digital sequence, thus forming the raw vibration waveform data. This raw vibration waveform data completely records the continuous process of the equipment's vibration amplitude changing over time. Since there is no need to install an additional dedicated speed sensor, the data acquisition can be completed using existing vibration sensors in the existing equipment condition monitoring device / system. Therefore, the raw data required for speed identification can be obtained without increasing hardware costs.

[0021] Step S102: Perform a discrete Fourier transform on the original vibration waveform data to obtain the corresponding spectrum data, which contains multiple frequency components arranged in frequency order.

[0022] Step S102 performs a Discrete Fourier Transform on the original vibration waveform data obtained in step S101, thereby converting the waveform signal in the time domain to the frequency domain and obtaining spectral data. This spectral data consists of multiple frequency components arranged in frequency order, each frequency component corresponding to a specific frequency value and its energy amplitude, including the target rotational frequency component representing the equipment's rotational speed. In other words, through the Discrete Fourier Transform, the time-domain waveform, which is difficult to directly observe rotational speed information, can be converted into a frequency-domain spectrum that clearly presents the energy distribution of each frequency component, providing a processable data format for subsequent identification of the target rotational frequency component.

[0023] For example, when performing a Discrete Fourier Transform (DFT) on the original vibration waveform data to obtain the corresponding spectrum data, a further DFT can be performed first, followed by modulo operation to obtain the corresponding spectrum data. This is because modulo operation not only converts the complex form of the transform result into a spectral energy sequence, which directly reflects the vibration energy magnitude of each frequency component—energy being the core physical basis for judging the importance of a frequency component—but also eliminates phase information that is not very helpful for frequency conversion identification. This is because phase is random and unstable, easily affected by external factors such as sensor installation location; removing it results in more stable and reliable data. Furthermore, modulo operation simplifies the subsequent learning task performed using the frequency conversion identification model. After modulo operation, the model does not need to handle complex real-imaginary part relationships; it only needs to focus on learning the energy correlation between frequency components, thus more accurately distinguishing the target frequency conversion component from noise interference.

[0024] Step S103: Each of the multiple frequency components is treated as a voter, and voting weights are assigned to all frequency components, including itself. The sum of the voting weights cast by each voter is a fixed value, and the sum of the weights obtained by each frequency component as a voted object is a non-fixed value. The voting weights characterize the correlation strength between the voter and the voted object.

[0025] Step S103 assigns a dual role to each frequency component in the spectral data obtained in step S102—both a voter and a voted-for. As a voter, each frequency component needs to allocate voting weights to all frequency components, including itself. The magnitude of these voting weights characterizes the strength of the association between the voter and the voted-for. The sum of all voting weights cast by each voter is constrained to a fixed value (e.g., 1), meaning each frequency component participates in the voting with a fixed total number of votes. However, the sum of voting weights obtained by each frequency component as a voted-for is unconstrained and can be any non-fixed value. This design allows frequency components strongly correlated with the target frequency conversion component to obtain higher voting weights from other frequency components, thus being significantly enhanced in this global voting mechanism; while interference frequency components unrelated to frequency conversion, such as noise, are effectively suppressed due to their lower voting weights.

[0026] In other words, step S103 above, by having each of the multiple frequency components act as a voter and assign voting weights to all frequency components, including itself, breaks through the limitation of traditional local peak detection algorithms that only focus on a few high-energy frequency components. It enables all frequency components to participate in the decision-making process and, through its unique design of "fixed output + variable reception," achieves automatic enhancement of the target frequency conversion component and automatic suppression of interference components, laying the foundation for accurate identification of the target frequency conversion component in the future.

[0027] Step S104: Identify the target frequency switching component from the plurality of frequency components based on the sum of the voting weights obtained by each frequency component as the voter.

[0028] Since the sum of voting weights reflects the degree of attention each frequency component receives in the global voting mechanism in step S103, the frequency component with the highest sum of voting weights is the component most strongly correlated with all other frequency components, which is the target rotational frequency component representing the equipment speed. Therefore, in step S104, the target rotational frequency component can be accurately identified from multiple frequency components by directly comparing the sum of the voting weights of each frequency component.

[0029] It should be noted that, since the sum of voting weights in step S104 has already enhanced the target frequency component and suppressed the interference component through step S103, the identification result in step S104 has high accuracy and high robustness. Even under complex working conditions with large speed fluctuations, load changes or large environmental interference, the true target frequency component can be reliably identified.

[0030] Step S105: Determine the real-time rotational speed of the rotating machinery based on the target rotational frequency component.

[0031] Step S105 obtains the corresponding frequency value of the target rotational frequency component identified in step S104. This frequency value is the rotational frequency of the rotating machinery. Since there is a definite reciprocal relationship between rotational frequency and rotational speed, specifically, rotational speed (in revolutions per minute) equals rotational frequency (in Hertz) multiplied by 60, the real-time rotational speed of the rotating machinery can be obtained through simple calculation.

[0032] In summary, this application assigns a dual role to each frequency component in the spectral data—both a voter and a voted-for—determining the target frequency switching component through a voting process involving all frequency components. Specifically, the sum of the votes cast by each voter is constrained to a fixed value, while the sum of the weights received by each frequency component as a voted-for is non-fixed. This design allows frequency components strongly correlated with the target frequency switching component to receive higher vote weights from other frequency components, thus significantly enhancing their performance in the global voting process, while interfering frequency components such as noise are effectively suppressed due to their lower vote weights.

[0033] Therefore, this application can achieve high-accuracy rotational speed identification through this voting mechanism without increasing hardware costs. Furthermore, since all frequency components participate in the voting, components strongly correlated with the rotational frequency receive higher weights from other components and are thus enhanced, while noise and other interfering components are suppressed. This effectively solves the problem in existing technologies where the true rotational frequency is obscured due to rotational speed fluctuations or environmental interference, leading to identification failure.

[0034] Based on the aforementioned scheme, in some implementations of this application, the step of assigning voting weights to all frequency components, including itself, by treating each of the multiple frequency components as a voter includes: generating a spectral energy sequence, a spectral position sequence, and a position encoding sequence corresponding to each frequency component based on the spectral data; determining the corresponding spectral vector representation using a fully connected neural network based on the spectral energy sequence and the spectral position sequence, and converting the position encoding sequence into the corresponding position encoding vector representation using an embedding layer vector matrix; fusing the spectral vector representation and the position encoding vector representation to obtain an input vector sequence; performing a linear transformation on the input vector sequence to obtain the corresponding query matrix, key matrix, and value matrix; and determining the voting weight assigned by each vector element in the input vector sequence as a voter to the value vector of all vector elements based on the degree of matching between each query vector in the query matrix and all key vectors in the query matrix.

[0035] Understandably, this implementation converts the raw spectral data into a vector form that can be processed by deep learning models, and introduces a query matrix, a key matrix, and a value matrix. This makes it easier to assign voting weights to all frequency components, including itself, by treating each of the multiple frequency components as a voter.

[0036] It should be noted that the spectral energy sequence and spectral position sequence correspond to the vertical and horizontal axes of the spectrum, respectively. The query matrix is ​​used to express the query intent of each vector element as a voter, the key matrix is ​​used to express the features of each vector element as a voted object, and the value matrix carries the actual information content of each vector element.

[0037] In summary, this implementation transforms the raw spectral data into a structured vector sequence, assigning each frequency component a query, key, and value role, thus enabling subsequent voting attention calculations. The introduction of position-encoded sequences allows the model to perceive the sequential relationships of frequency components within the spectrum, leading to a better understanding of the spectrum's physical meaning. Furthermore, through trainable linear transformations, the query matrix, key matrix, and value matrix are continuously optimized during model training, ultimately learning the most suitable representation for the frequency transition recognition task, supporting accurate identification of the target frequency transition component.

[0038] Based on the aforementioned scheme, in some implementations of this application, the step of determining the voting weight assigned by each vector element in the input vector sequence as a voter to the value vector of all vector elements based on the matching degree between each query vector in the query matrix and all key vectors in the query matrix includes: calculating the formula... The attention score matrix is ​​calculated. ,in, For querying the matrix, The key matrix, The superscript represents the dimension of the spectrum vector and the position encoding vector. Indicates transpose; according to the calculation formula For the attention score matrix Each row is normalized to obtain the voting weight matrix. ,in, This is the normalization function; according to the calculation formula... Calculate the first The voting fusion vector corresponding to each voter ,in, This represents the number of spectral components contained in the spectral data. For the first The voter was assigned to the first The voting weight of each voter Value matrix The A vector of values.

[0039] Understandably, this implementation concretizes the abstract voting mechanism through mathematical calculations, calculates the attention score by using the matching degree between the query matrix and the key matrix, obtains the voting weight by normalization, and finally generates a voting fusion vector by weighted fusion, thereby realizing the core computational logic of the global voting mechanism unique to this application.

[0040] It should be noted that in the above implementation, the original attention score can be obtained by calculating the matching degree between each query vector and all key vectors. Then, by normalizing each row of the attention score matrix, a voting weight matrix is ​​obtained, which ensures that the sum of the vote weights cast by each voter is a fixed value, that is, the sum of the weights that each frequency component, as a voter, assigns to all the voted objects is 1. The calculation formula is equivalent to This is because it is essentially an operation of "using the weight of a certain column to sum all rows in a weighted manner," which is precisely the operation that can aggregate the contributions of all voters to the same candidate. That is, the calculation formula merges the weighted contributions that each candidate receives from all voters to obtain an enhanced representation of that candidate.

[0041] It's also worth noting that in conventional attention mechanisms, the rows and columns of the attention weight matrix are symmetrical. After Softmax normalization, a probability-weighted average is formed, with the weights of each row and column summing to 1. The implementation described above breaks this traditional calculation method. By transposing the attention weight matrix, each frequency component scores all frequency components in the spectrum, ensuring that the sum of the scores given by each frequency component is 1, but the total score received by each frequency component is not fixed. This operation treats every frequency component in the spectrum equally, allowing all frequency components to participate in the voting decision, while simultaneously allowing for differences in the strength of the influence of different frequency components on the target frequency. Components with strong correlation to the frequency shift receive higher scores and are amplified, while irrelevant components such as noise are suppressed due to lower scores.

[0042] In summary, this implementation transforms the allocation and fusion of voting weights into mathematically implementable matrix operations through the aforementioned calculation process. Since the sum of the weights cast by each voter is constrained to a fixed value, while the sum of the weights received by each voted-for component is non-fixed, frequency components strongly correlated with the target frequency transition component can obtain higher weights from other frequency components, thus being significantly enhanced in the voting fusion vector. Meanwhile, noise and other interfering components are suppressed, providing high-quality input features for accurate identification of the target frequency transition component.

[0043] Based on the aforementioned scheme, in some implementations of this application, the step of identifying the target frequency transition component from the plurality of frequency components based on the sum of the voting weights obtained by each frequency component as a voter includes: generating an attention output sequence based on the sum of the voting weights obtained by each frequency component as a voter; inputting the attention output sequence into a classification network to obtain the probability value of the frequency component corresponding to each position in the attention output sequence as a frequency transition; and determining the frequency component with the highest probability value as the target frequency transition component.

[0044] Understandably, this implementation transforms the frequency transition identification problem into a classification problem by using a classification network to predict the probability of the attention output sequence, thereby accurately selecting the target frequency transition component from multiple frequency components. Specifically, an attention output sequence is first generated by summing the voting weights of each frequency component as a voter. Each element in this attention output sequence corresponds to a frequency component and has undergone enhancement processing through a global voting mechanism, with frequency components strongly correlated with the frequency transition receiving higher weights. The attention output sequence is then input into the classification network, which predicts the probability of each position in the sequence and outputs the probability value of the frequency component corresponding to that position. Finally, the frequency component with the highest probability value is determined as the target frequency transition component.

[0045] Therefore, the processing method described in this application makes the frequency conversion recognition process clearly interpretable and trainable. It should be noted that, because the target frequency conversion component in the attention output sequence has been enhanced and the interference component has been suppressed, the classification network will be able to more easily learn the features that distinguish between frequency conversion and non-frequency conversion, thus maintaining high accuracy in recognition results even under complex conditions. This effectively solves the problem of frequency conversion recognition failure caused by interference frequencies in existing technologies.

[0046] Based on the aforementioned scheme, in some implementations of this application, each of the multiple frequency components is treated as a voter, and voting weights are assigned to all frequency components, including itself; and the step of identifying the target frequency switching component from the multiple frequency components based on the sum of the voting weights obtained by each frequency component as a voter is performed by a pre-trained frequency switching identification model. The frequency switching identification model includes an input representation layer, a voting attention layer, and a classification layer connected in sequence. The input representation layer is used to generate an input vector sequence based on the spectral data; the voting attention layer is used to treat each vector element in the input vector sequence as a voter, assign voting weights to all voters based on the correlation strength between each voter and all other voters, and generate an attention output sequence; the classification layer is used to determine the target frequency switching component based on the attention output sequence.

[0047] Understandably, this implementation encapsulates the process of identifying target frequency components from spectral data into an end-to-end neural network processing flow by constructing a frequency conversion recognition model consisting of an input representation layer, a voting attention layer, and a classification layer. Therefore, this allows the entire recognition process to complete parameter optimization during offline training, and in online applications, only forward computation is needed to quickly and accurately obtain the recognition result. The three layers of the frequency conversion recognition model—input representation layer, voting attention layer, and classification layer—can fully learn the complex relationships between frequency components in the spectral data, thus maintaining high accuracy in identifying the target frequency components even under complex operating conditions with fluctuating rotational speeds, varying loads, or significant environmental interference.

[0048] Specifically, in practical applications, the spectral data is first input to the input representation layer of the frequency transition recognition model. This layer is responsible for converting the raw spectral data into an input vector sequence that can be processed later. The input vector sequence is then passed to the voting attention layer, which treats each vector element as a voter and assigns voting weights to all voters based on the correlation strength between each voter and all other voters, generating an attention output sequence. During this process, the sum of the voting weights cast by each voter is constrained to a fixed value, while the sum of the weights received by each vector element as the voted-for component is not fixed. This allows vector elements with a strong correlation to the target frequency transition component to receive higher weights and be strengthened. Finally, the attention output sequence is input to the classification layer, which determines the target frequency transition component based on the attention output sequence.

[0049] For example, during the training of the frequency conversion recognition model, the loss function can be the cross-entropy loss function, that is, the calculation formula of the target loss function L for training is: ,in, f c Let y be the c-th element in the spectral position sequence. c p is the c-th element in the spectral energy sequence. c It is the c-th element in the position-coded sequence.

[0050] Based on the aforementioned scheme, in some implementations of this application, the frequency transfer recognition model is a neural network model based on an attention mechanism, which is obtained through supervised learning training. The training steps include: obtaining a training dataset, which contains multiple training samples, each training sample including sample spectrum data and labeled frequency transfer components corresponding to the sample spectrum data; taking the sample spectrum data as input and the labeled frequency transfer components corresponding to the sample spectrum data as target output, iteratively training the initial neural network model until the loss function converges; and determining the converged neural network model as the frequency transfer recognition model.

[0051] Understandably, this implementation uses supervised learning to enable the frequency conversion recognition model to automatically learn the complex relationships between frequency components in the spectrum data and the correspondence between these relationships and the actual frequency conversion from a large amount of real-world operating data. This allows the trained model to have stronger adaptability and generalization ability, maintaining high accuracy in frequency conversion recognition even under complex operating conditions such as speed fluctuations, load changes, or environmental interference.

[0052] Specifically, in this implementation, a training dataset is first obtained. This dataset contains multiple training samples, each consisting of sample spectral data and the corresponding labeled transconversion frequency (TRF) components. During training, the sample spectral data is input to an initial neural network model. The model performs forward computation and outputs a prediction result. Simultaneously, the corresponding labeled TRF components are used as the target output and compared with the model's prediction result to calculate the loss function value. The backpropagation algorithm is used to continuously adjust the model's internal parameters, gradually reducing the loss function value. This process is repeated iteratively to train the model until the loss function converges. At this point, the model has learned the ability to accurately identify TRF components from the sample spectral data. Finally, the converged neural network model is determined as the final TRF recognition model, which can be used for online TRF recognition tasks.

[0053] Based on the aforementioned scheme, in some implementations of this application, the input representation layer is further configured to: extract the spectral energy sequence Y={y1, y2, ..., y...} based on the spectral data. m} and the spectral position sequence F={f1, f2, ..., f m The algorithm generates a position encoding sequence P = {1, 2, ..., m} of the same length as the spectral data, where m is the number of frequency components contained in the spectral data. It then maps the spectral energy sequence Y and the spectral position sequence F to a spectral vector representation Ef, with the dimension of Ef being m×d, using a fully connected neural network. Finally, it maps the position encoding sequence P to a position encoding vector representation Ep, with the dimension of Ep also being m×d, wherein the parameters of the embedding layer can be adjusted during model training. The spectral vector representation Ef and the position encoding vector representation Ep are then added element-wise to obtain the input vector sequence E = Ef + Ep.

[0054] It should be noted that the spectral vector representation Ef carries the energy and frequency information of each frequency component, while the positional encoding vector representation Ep provides the model with information about the order of these frequency components within the spectrum. The resulting input vector sequence E, formed by fusing the two, contains both the physical features of each frequency component and preserves their relative positions within the spectrum, providing high-quality input for the subsequent voting attention layer to accurately calculate the correlation strength between frequency components. Furthermore, the trainable nature of the embedding layer parameters allows the model to learn the positional representation best suited to the task requirements during training.

[0055] Based on the aforementioned scheme, in some implementations of this application, the classification layer is a multilayer perceptron network, which is further used to: flatten the attention output sequence into a one-dimensional feature vector; process the one-dimensional feature vector sequentially through multiple fully connected layers and nonlinear activation functions to obtain a probability distribution corresponding to each frequency component; and output the probability distribution, wherein each value in the probability distribution represents the probability that the corresponding frequency component is a frequency transition.

[0056] Understandably, this implementation classifies the attention output sequence through a multilayer perceptron network, transforming the question of whether each frequency component is a frequency transition into a probability prediction problem.

[0057] Specifically, the classification layer first flattens the attention output sequence into a one-dimensional feature vector. This attention output sequence might originally be a multi-dimensional structure; the flattening operation transforms it into a long vector for subsequent fully connected layers to process. This one-dimensional feature vector is then sequentially input into multiple fully connected layers, each followed by a non-linear activation function. Through this multi-layered non-linear transformation, the classification layer can extract high-level features related to frequency transition recognition from the attention output sequence. Finally, after a series of transformations, the classification layer outputs a probability distribution corresponding to each frequency component, where each value in the probability distribution represents the probability that the corresponding frequency component is a frequency transition.

[0058] It should be noted that this implementation, through the combination of multiple fully connected layers and nonlinear activation functions, can capture the complex nonlinear relationships in the attention output sequence, thereby more accurately distinguishing between transition frequency components and non-transition frequency components. Furthermore, since the final output is a probability distribution, the target transition frequency component can be directly determined based on the probability values, making the decision-making process intuitive and clear. Simultaneously, the parameters of the multilayer perceptron network, serving as the classification layer, can be optimized end-to-end during model training, together with the preceding input representation layer and voting attention layer, to achieve optimal performance for the entire transition frequency recognition model.

[0059] To enable those skilled in the art to more intuitively understand this application, a specific example will be provided below. This example is an exemplary demonstration combining the overall technical paradigm of this application with some optional implementation details. It should be noted that the following demonstration is intended to aid understanding and does not constitute an exhaustive list of all embodiments of this application, nor does it imply that this application must include all the details described below in its specific implementation.

[0060] In this example, a pre-trained frequency recognition model is used to determine the frequency and rotational speed of rotating machinery by voting on the influence of each frequency component in the spectral data on the frequency.

[0061] The following provides a detailed explanation of the implementation process.

[0062] The first step is to collect waveform data using a vibration sensor. Vibration sensors are typically installed on critical components of rotating machinery that need to be monitored. The equipment's status is obtained by collecting vibration acceleration data. The collected vibration acceleration data is the raw vibration waveform data, denoted as... ,in N This represents the number of sampling points.

[0063] The original vibration waveform data is then subjected to a Discrete Fourier Transform (using any existing Discrete Fourier Transform, which will not be elaborated here) and a modulus operation is performed to obtain the spectrum data.

[0064] Next, the signal transponder recognition model will be used for processing. This model is a deep learning model that involves two stages: training and inference. Its structural diagram is shown below. Figure 2 As shown. In actual online real-time equipment status monitoring and fault diagnosis, only the inference stage of the frequency conversion identification model is involved; the training stage is completed offline.

[0065] The input to the frequency switching identification model includes a spectral energy sequence, a spectral position sequence, and a position coding sequence. The spectral energy sequence is denoted as Y = {y1, y2, ..., y...}. m The spectral position sequence is denoted as F = {f1, f2, ..., f}. m Both originate from spectral data and represent the vertical and horizontal axes of the spectrum, respectively. The position-coded sequence is a sequence of natural numbers of the same length as the spectral data, denoted as P = {1, 2, ..., m}.

[0066] The first layer (input representation layer) of the frequency conversion recognition model consists of two parts: a fully connected neural network and a position encoding layer. The fully connected neural network converts the spectral data into a dense vector representation Ef, as follows: Ef = FFN(Y, F), where FFN represents the fully connected neural network. Simultaneously, the position encoding layer converts the discrete sequence of natural numbers into a dense vector representation Ep through the embedding layer vector matrix, as follows: Ep = Embed(P), where Embed represents the embedding layer vector matrix, the parameters of which can be adjusted during model training. Then, the spectral vector representation Ef and the position encoding vector representation Ep are summed to obtain the input vector sequence E = Ef + Ep.

[0067] The second part of the frequency conversion recognition model is the voting attention layer, which takes the input vector sequence E as input. The calculation formula for the voting attention layer is as follows: Q=W q E, Q=W k E, Q=W v E

[0068]

[0069] ,Right now,

[0070] Among them, W q W k W v The parameters are trainable; the definitions of the remaining parameters are provided above and will not be repeated here. The voting attention described above exhibits a "flow non-conservation" characteristic, meaning that each frequency component votes for all frequency components (including itself), and the sum of the votes cast by each voter is 1. However, the sum of the votes received by each frequency component as the object of voting is not a fixed value. For frequency components with high attention, the sum of the votes received by that frequency component is higher; for frequency components with low attention, such as noise, almost no frequency components vote for them, and therefore they are suppressed.

[0071] The third part of the frequency transition recognition model is the classification network (i.e., the classification layer), which is mainly composed of a multilayer perceptron. The classification network directly operates on the voting attention output to obtain the output vector. Finally, an argmax operation is performed on the output sequence vector to obtain the position of the frequency transition component in the spectrum data, thereby obtaining the target frequency transition component.

[0072] Please see Figure 3 This application also provides a system for identifying the rotational speed of rotating machinery, comprising: a data acquisition module 101 for acquiring raw vibration waveform data of the rotating machinery; a spectrum transformation module 102 for performing a discrete Fourier transform on the raw vibration waveform data to obtain corresponding spectrum data, the spectrum data containing multiple frequency components arranged in frequency order; a voting allocation module 103 for assigning voting weights to all frequency components, including itself, by treating each of the multiple frequency components as a voter, wherein the sum of the voting weights cast by each voter is a fixed value, and the sum of the weights obtained by each frequency component as a voted-for is a non-fixed value, the voting weights representing the correlation strength between the voter and the voted-for; a rotational frequency identification module 104 for identifying a target rotational frequency component from the multiple frequency components based on the sum of the voting weights obtained by each frequency component as a voted-for; and a rotational speed determination module 105 for determining the real-time rotational speed of the rotating machinery based on the target rotational frequency component.

[0073] For the specific implementation process of the above system, please refer to the method for identifying the rotational speed of rotating machinery provided in the above embodiment, which will not be repeated here.

[0074] Please see Figure 4This application also provides an electronic device, which includes at least one processor 201 and at least one memory 202. The processor 201 and memory 202 are directly connected to each other, or communicate with each other through a communication interface 203, or are electrically connected through one or more communication buses or signal lines to achieve data transmission or interaction. The memory 202 stores program instructions that can be executed by the processor 201. The processor 201 can call and execute the program instructions to implement any of the methods for identifying the rotational speed of rotating machinery provided by the various implementations described above. For example, implementing: Obtain the raw vibration waveform data of the rotating machinery. Perform a Discrete Fourier Transform on the raw vibration waveform data to obtain the corresponding spectrum data, which contains multiple frequency components arranged in frequency order. Each of these frequency components is treated as a voter, and voting weights are assigned to all frequency components, including itself. The sum of the votes cast by each voter is a fixed value, while the sum of the weights received by each frequency component as a voted-for component is a non-fixed value. These voting weights characterize the correlation strength between the voter and the voted-for component. Based on the sum of the vote weights received by each frequency component as a voted-for component, identify the target rotational frequency component from the multiple frequency components. Determine the real-time rotational speed of the rotating machinery based on the target rotational frequency component.

[0075] The memory 202 may be, but is not limited to, random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), etc.

[0076] The processor 201 can be an integrated circuit chip with signal processing capabilities. The processor 201 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0077] Understandable. Figure 4 The structure shown is for illustrative purposes only; the electronic device may also include components that are more advanced than those shown. Figure 4 The more or fewer components shown, or having the same Figure 4 The different configurations shown. Figure 4 The components shown can be implemented using hardware, software, or a combination thereof.

[0078] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of this application is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within this application. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for identifying the rotational speed of rotating machinery, characterized in that, The method includes: Acquire raw vibration waveform data of rotating machinery; The original vibration waveform data is subjected to a discrete Fourier transform to obtain the corresponding spectrum data, which contains multiple frequency components arranged in frequency order. Each of the multiple frequency components is treated as a voter, and voting weights are assigned to all frequency components, including itself. The sum of the voting weights cast by each voter is a fixed value, while the sum of the weights obtained by each frequency component as a voted-for object is a non-fixed value. The voting weights characterize the correlation strength between the voter and the voted-for object. The target frequency switching component is identified from the plurality of frequency components based on the sum of the voting weights obtained by each frequency component as a voter. The real-time rotational speed of the rotating machinery is determined based on the target rotational frequency component.

2. The method according to claim 1, characterized in that, The step of assigning voting weights to all frequency components, including itself, by treating each of the plurality of frequency components as a voter includes: Generate a spectral energy sequence, a spectral position sequence, and a position coding sequence corresponding to each frequency component based on the spectral data; A fully connected neural network is used to determine the corresponding spectral vector representation based on the spectral energy sequence and the spectral position sequence, and the position encoding sequence is converted into the corresponding position encoding vector representation using the embedding layer vector matrix; The spectral vector representation and the positional encoding vector representation are fused to obtain the input vector sequence; A linear transformation is performed on the input vector sequence to obtain the corresponding query matrix, key matrix, and value matrix. Based on the degree of matching between each query vector in the query matrix and all key vectors in the query matrix, determine the voting weight assigned by each vector element in the input vector sequence as a voter to the value vector of all vector elements.

3. The method according to claim 2, characterized in that, The step of determining the voting weight assigned by each vector element in the input vector sequence as a voter to the value vector of all vector elements based on the matching degree between each query vector in the query matrix and all key vectors in the query matrix includes: According to the calculation formula The attention score matrix is ​​calculated. ,in, For querying the matrix, The key matrix, The superscript represents the dimension of the spectrum vector and the position encoding vector. Indicates transpose; According to the calculation formula For the attention score matrix Each row is normalized to obtain the voting weight matrix. ,in, This is the normalization function; According to the calculation formula Calculate the first The voting fusion vector corresponding to each voter ,in, This represents the number of spectral components contained in the spectral data. For the first The voter was assigned to the first The voting weight of each voter Value matrix The A vector of values.

4. The method according to claim 1, characterized in that, The step of identifying the target frequency switching component from the plurality of frequency components by summing the voting weights obtained by each frequency component as a voter includes: An attention output sequence is generated based on the sum of the voting weights obtained by each frequency component as the voter. The attention output sequence is input into the classification network to obtain the probability value of the frequency component corresponding to each position in the attention output sequence being a frequency transition. The frequency component with the highest probability value is determined as the target frequency conversion component.

5. The method according to claim 1, characterized in that, Each of the multiple frequency components is treated as a voter, and voting weights are assigned to all frequency components, including itself; and the steps of identifying the target frequency switching component from the multiple frequency components based on the sum of the voting weights obtained by each frequency component as a voter are all performed by a pre-trained frequency switching recognition model. The frequency conversion recognition model includes an input representation layer, a voting attention layer, and a classification layer connected in sequence. The input representation layer is used to generate an input vector sequence based on the spectrum data; The voting attention layer is used to treat each vector element in the input vector sequence as a voter, and to assign voting weights to all voters based on the correlation strength between each voter and all other voters, thereby generating an attention output sequence. The classification layer is used to determine the target frequency conversion component based on the attention output sequence.

6. The method according to claim 5, characterized in that, The frequency conversion recognition model is a neural network model based on an attention mechanism, obtained through supervised learning training. Its training steps include: Obtain a training dataset, which contains multiple training samples, each training sample including sample spectral data and labeled frequency transition components corresponding to the sample spectral data; The sample spectrum data is used as input, and the labeled frequency conversion component corresponding to the sample spectrum data is used as the target output to iteratively train the initial neural network model until the loss function converges. The neural network model that has been trained and converged is determined as the frequency conversion recognition model.

7. The method according to claim 5, characterized in that, The input representation layer is further used for: Based on the aforementioned spectral data, extract the spectral energy sequence Y = {y1, y2, ..., y...} m } and the spectral position sequence F={f1, f2, ..., f m }, and generate a position coding sequence P={1,2,…,m} of the same length as the spectrum data, where m is the number of frequency components contained in the spectrum data; The spectral energy sequence Y and the spectral position sequence F are mapped to a spectral vector representation Ef using a fully connected neural network, wherein the dimension of the spectral vector representation Ef is m×d; The positional encoding sequence P is mapped to a positional encoding vector representation Ep by an embedding layer vector matrix. The positional encoding vector representation Ep has a dimension of m×d, and the parameters of the embedding layer can be adjusted during model training. The spectral vector representation Ef is added element-wise to the positional encoding vector representation Ep to obtain the input vector sequence E=Ef+Ep.

8. The method according to claim 5, characterized in that, The classification layer is a multilayer perceptron network, further used for: Flatten the attention output sequence into a one-dimensional feature vector; The one-dimensional feature vector is processed sequentially through multiple fully connected layers and nonlinear activation functions to obtain the probability distribution corresponding to each frequency component; Output the probability distribution, where each value in the probability distribution represents the probability that the corresponding frequency component is a frequency shift.

9. A system for identifying the rotational speed of rotating machinery, characterized in that, include: The data acquisition module is used to acquire the raw vibration waveform data of rotating machinery. The spectrum transformation module is used to perform discrete Fourier transform on the original vibration waveform data to obtain the corresponding spectrum data, which contains multiple frequency components arranged in frequency order. The voting allocation module is used to assign voting weights to all frequency components, including itself, by treating each of the multiple frequency components as a voter. The sum of the voting weights cast by each voter is a fixed value, while the sum of the weights obtained by each frequency component as a voted object is a non-fixed value. The voting weights characterize the correlation strength between the voter and the voted object. The frequency conversion identification module is used to identify the target frequency conversion component from the multiple frequency components based on the sum of the voting weights obtained by each frequency component as the voter. The rotational speed determination module is used to determine the real-time rotational speed of the rotating machinery based on the target rotational frequency component.

10. An electronic device, characterized in that, include: Memory, used to store one or more programs; processor; When the one or more programs are executed by the processor, the method as described in any one of claims 1-8 is implemented.

Images