Air energy equipment detection method and equipment based on vibration signals and medium

Abstract

The embodiment of the invention discloses an air energy equipment detection method based on a vibration signal, equipment and a medium, belongs to the technical field of equipment fault detection, and solves the problems that a traditional air energy equipment detection method depends on manual analysis of vibration data, time and labor are consumed, and certain errors and delay exist. Comprising the following steps: acquiring vibration signals corresponding to different to-be-tested air energy equipment; carrying out noise reduction processing on the vibration signal through fusion of an adaptive filter and wavelet transform; based on a multi-layer noise reduction auto-encoder, carrying out layer-by-layer encoding and decoding on the vibration signals after noise reduction to obtain feature data corresponding to different air energy devices to be detected; performing associated information analysis on the feature data through a preset multi-head self-attention mechanism model, and determining different operation states corresponding to the to-be-detected air energy equipment; and on the basis of the air energy equipment state, constructing an equipment association diagram, so as to determine cross-equipment maintenance information corresponding to the plurality of air energy equipment to be detected on the basis of the equipment association diagram.

Description

Technical Field

[0001] The present application relates to the technical field of equipment fault detection, and particularly to a detection method, device and medium for air energy equipment based on vibration signals. Background Art

[0002] In the current field of industrial equipment maintenance and management, as an important energy conversion equipment, the stability and efficiency of the operation state of air energy equipment directly affect the performance and energy consumption of the entire system. In order to timely detect potential faults of the equipment and perform maintenance, it is particularly important to monitor and diagnose the operation state of air energy equipment in real time.

[0003] Vibration data, as an important information source reflecting the operation state of equipment, the accuracy and timeliness of its processing and analysis directly affect the accuracy and timeliness of fault diagnosis. Most traditional detection methods for air energy equipment rely on manual analysis of vibration data. The process of manual analysis of vibration data usually includes steps such as data acquisition, signal processing, and feature extraction. These steps mostly rely on the experience and skill level of operators, are easily affected by subjective factors, and this method is time-consuming and laborious, with certain errors and delays. Therefore, with the increase in the number of equipment and the improvement of monitoring requirements, the efficiency and accuracy of the manual analysis method are difficult to meet the actual needs. Summary of the Invention

[0004] Embodiments of the present application provide a detection method, device and medium for air energy equipment based on vibration signals, which are used to solve the following technical problems: Most traditional detection methods for air energy equipment rely on manual analysis of vibration data. This method is time-consuming and laborious, with certain errors and delays. With the increase in the number of equipment and the improvement of monitoring requirements, the efficiency and accuracy of the manual analysis method are difficult to meet the actual needs.

[0005] Embodiments of the present application adopt the following technical solutions:

[0006] Embodiments of the present application provide a detection method for air energy equipment based on vibration signals. The method includes: obtaining vibration signals respectively corresponding to different air energy equipment to be measured through vibration sensors arranged at different positions of the air energy equipment to be measured; performing noise reduction processing on the vibration signals through the fusion of an adaptive filter and wavelet transform; performing layer-by-layer encoding and decoding on the noise-reduced vibration signals based on a multi-layer noise reduction autoencoder to obtain feature data respectively corresponding to different air energy equipment to be measured; analyzing the associated information of the feature data through a pre-set multi-head self-attention mechanism model to determine the probability values of the air energy equipment to be measured corresponding to the feature data being in different operation states, so as to determine the air energy equipment states respectively corresponding to different air energy equipment to be measured based on the probability values; constructing an equipment association graph based on the air energy equipment states, so as to determine cross-equipment maintenance information corresponding to multiple air energy equipment to be measured based on the equipment association graph.

[0007] In the embodiment of the present application, the adaptive filter dynamically adjusts the filtering parameters according to the real-time characteristics of the signal, effectively suppressing noise interference; wavelet transform can perform multi-resolution analysis on the signal, finely removing noise in different frequency bands while retaining the key features of the signal. Secondly, in the embodiment of the present application, the multi-layer denoising autoencoder encodes and decodes the denoised vibration signal layer by layer, and can automatically learn the complex features contained in the vibration signal, effectively extracting the feature data that can accurately reflect the operating state of the device. Combined with the preset multi-head self-attention mechanism model, the probability values of the air energy device to be measured in different operating states are accurately determined. In addition, in the embodiment of the present application, an equipment association graph is constructed based on the equipment state, clearly showing the relationship between different air energy devices to be measured, which helps maintenance personnel to grasp the overall operating situation of the equipment, quickly determine the cross-equipment maintenance information corresponding to multiple air energy devices to be measured, improve the scientificity of maintenance decisions and the efficiency of maintenance work, reduce equipment downtime, and reduce maintenance costs.

[0008] In an implementation manner of the present application, the vibration signal is denoised by fusing the adaptive filter and wavelet transform, which specifically includes: dynamically adjusting the step factor based on the initial step size, the maximum step size, and the minimum step size; constructing a weight update function of the adaptive filtering algorithm based on the dynamically adjusted step factor, the filter coefficient vector, the error, and the vibration signal to iteratively update the filter coefficients; constructing an adaptive filter based on the recurrent neural network through the iteratively updated filter coefficients, and performing adaptive filtering on the vibration signal based on the vibration signal at the current moment and the output corresponding to the adaptive filter at the previous moment; performing wavelet transform on the filtered vibration signal, determining the error value between the wavelet-transformed signal and the current output signal of the adaptive filter, and updating and adjusting the coefficients of the adaptive filter according to the error magnitude and error direction; sparsely representing the wavelet-transformed vibration signal on an overcomplete dictionary; screening the sparse representation coefficients based on the sparse representation coefficients and the preset coefficient threshold; and reconstructing the vibration signal through the screened coefficients and the overcomplete dictionary to obtain the denoised vibration signal.

[0009] In an implementation manner of the present application, the denoised vibration signal is encoded and decoded layer by layer based on a multi-layer denoising autoencoder to obtain characteristic data corresponding to different air energy devices to be measured, specifically including: determining an energy low-frequency band and a noise high-frequency band based on the energy distribution and noise level corresponding to the vibration signal sample; increasing the noise intensity in the energy low-frequency band and injecting a noise type with the same noise characteristics in the noise high-frequency band; discriminating the characteristics after encoding the original signal and the encoded characteristics after being contaminated by noise through a preset discriminator, so as to adjust the parameters of the multi-layer denoising autoencoder based on the discrimination result; inputting the vibration signal into the multi-layer denoising autoencoder with adjusted parameters for decoding, determining a first correlation between the encoded characteristics and the decoded characteristics through a global attention mechanism, and determining a second correlation between each frequency band in the encoded characteristics through a local attention mechanism; assigning weights to each encoded characteristic based on the first correlation and the second correlation to obtain a reconstructed vibration signal, so as to obtain characteristic data corresponding to different air energy devices to be measured based on the reconstructed vibration signal.

[0010] In an implementation manner of the present application, weights are assigned to each encoded characteristic based on the first correlation and the second correlation to obtain a reconstructed vibration signal, specifically including: through the function:

[0011]

[0012] assigning weights to each encoded characteristic based on the first correlation and the second correlation; where α is an adjustment parameter; β is an adjustment parameter; W is a weight; G is a global attention score vector; L is a local attention score matrix; L ij is the local correlation between feature i and feature j; E is an encoded feature vector; n is the dimension; based on the function:

[0013] E i ' = W i ·Ei;

[0014] obtaining the reconstructed vibration signal; where E' is a weighted feature vector; W is a weight; E is an encoded feature vector.

[0015] In an implementation of the present application, by presetting a multi-head self-attention mechanism model, the associated information of the feature data is analyzed to determine the probability values of the air energy device to be measured corresponding to the feature data in different operating states, specifically including: determining the number of self-attention heads based on the dimension and data complexity corresponding to the feature data; in the constructed multi-head self-attention mechanism stacked model structure, processing the output of the next layer based on the output of the previous layer, and through a preset feed-forward neural network, performing non-linear transformation and feature integration on the feature representation output by the multi-head self-attention mechanism stacked model; by presetting a Softmax activation function, converting the output of the multi-head self-attention mechanism stacked model into the probability values of the air energy device to be measured in different operating states.

[0016] In an implementation of the present application, by presetting a Softmax activation function, the output of the multi-head self-attention mechanism stacked model is converted into the probability values of the air energy device to be measured in different operating states, specifically including: based on the function:

[0017]

[0018] where

[0019] the probability values of the air energy device to be measured in different operating states are determined; where P j is the probability value; z j is the output of the multi-head self-attention mechanism stacked model, where n is the number of categories of the operating states of the air energy device; the air energy device state transition weight matrix is W = [w ij n×n ; u is the feature uncertainty coefficient; ρ j is the correction factor; i is the air energy device state category i; j is the air energy device state category j; γ is the first adjustment parameter; is the second adjustment parameter; ACF i (τ) is the autocorrelation function; τ is the delay time; (h i -l i ) is the model uncertainty interval.

[0020] ​In an implementation manner of the present application, based on the states of air energy devices, a device association graph is constructed to determine cross-device maintenance information corresponding to multiple air energy devices to be tested based on the device association graph. Specifically, it includes: constructing a device association graph based on the positions between different air energy devices to be tested, the similarity between different air energy devices to be tested, and the production process relationship between different air energy devices to be tested; marking air energy devices with the same state in the device association graph and determining a reference air energy device in a faulty state; determining the fault propagation paths respectively corresponding to each reference air energy device in the device association graph, and based on the fault propagation paths, determining associated reference air energy devices; performing the same marking on the associated reference air energy devices caused by the same fault reason; grouping the reference air energy devices and the associated reference air energy devices based on the markings of the air energy devices, so as to construct corresponding maintenance information for different groups of air energy devices.

[0021] In an implementation manner of the present application, determining the fault propagation paths respectively corresponding to each reference air energy device in the device association graph, and based on the fault propagation paths, determining associated reference air energy devices, specifically includes: in the historical air energy device fault database, determining the combination of historical faulty air energy devices corresponding to the reference air energy device, so as to determine the historical data of air energy device faults based on the combination of historical faulty air energy devices; constructing a Bayesian network model according to the device association graph and the historical data of air energy device faults; wherein, in the Bayesian network model, the state of the air energy device is used as a node, the association relationship between air energy devices is used as an edge, and the weight of the edge is used as a conditional probability; using the reference air energy device as the starting point of the faulty air energy device, and determining the propagation probability of the fault between air energy devices and the reference propagation path through Bayesian network reasoning; sorting the probabilities of different propagation paths, and determining the fault propagation path and the associated reference air energy device based on the sorting result.

[0022] An embodiment of the present application provides a detection device for air energy equipment based on vibration signals, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to: obtain vibration signals respectively corresponding to different air energy equipment to be measured through vibration sensors arranged at different positions of the air energy equipment to be measured; perform noise reduction processing on the vibration signals through the fusion of an adaptive filter and wavelet transform; perform layer-by-layer encoding and decoding on the noise-reduced vibration signals based on a multi-layer noise reduction autoencoder to obtain characteristic data respectively corresponding to different air energy equipment to be measured; perform associated information analysis on the characteristic data through a preset multi-head self-attention mechanism model to determine probability values of the air energy equipment to be measured corresponding to the characteristic data being in different operating states, so as to determine the air energy equipment states respectively corresponding to different air energy equipment to be measured based on the probability values; construct a device association graph based on the air energy equipment states, so as to determine cross-device maintenance information corresponding to multiple air energy equipment to be measured based on the device association graph.

[0023] A non-volatile computer storage medium provided by an embodiment of the present application stores computer-executable instructions, and the computer-executable instructions are set to: obtain vibration signals respectively corresponding to different air energy equipment to be measured through vibration sensors arranged at different positions of the air energy equipment to be measured; perform noise reduction processing on the vibration signals through the fusion of an adaptive filter and wavelet transform; perform layer-by-layer encoding and decoding on the noise-reduced vibration signals based on a multi-layer noise reduction autoencoder to obtain characteristic data respectively corresponding to different air energy equipment to be measured; perform associated information analysis on the characteristic data through a preset multi-head self-attention mechanism model to determine probability values of the air energy equipment to be measured corresponding to the characteristic data being in different operating states, so as to determine the air energy equipment states respectively corresponding to different air energy equipment to be measured based on the probability values; construct a device association graph based on the air energy equipment states, so as to determine cross-device maintenance information corresponding to multiple air energy equipment to be measured based on the device association graph.

[0024] The above at least one technical solution adopted in the embodiments of the present application can achieve the following beneficial effects: In the embodiments of the present application, the adaptive filter dynamically adjusts the filtering parameters according to the real-time characteristics of the signal, effectively suppressing noise interference; wavelet transform can perform multi-resolution analysis on the signal, finely removing noise in different frequency bands while retaining the key features of the signal. Secondly, in the embodiments of the present application, the multi-layer denoising autoencoder encodes and decodes the denoised vibration signal layer by layer, and can automatically learn the complex features contained in the vibration signal, effectively extracting the feature data that can accurately reflect the operating state of the device. Combined with the preset multi-head self-attention mechanism model, the probability values of the air energy device to be measured in different operating states are accurately determined. In addition, in the embodiments of the present application, an equipment association graph is constructed based on the equipment state, clearly showing the relationships between different air energy devices to be measured, which helps maintenance personnel to grasp the overall equipment operation situation, quickly determine the cross-equipment maintenance information corresponding to multiple air energy devices to be measured, improve the scientificity of maintenance decisions and the efficiency of maintenance work, reduce equipment downtime, and reduce maintenance costs. Description of the Drawings

[0025] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments recorded in the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings. In the drawings:

[0026] Figure 1 It is a flowchart of a method for detecting an air energy device based on vibration signals provided by an embodiment of the present application;

[0027] Figure 2 It is a schematic structural diagram of a device for detecting an air energy device based on vibration signals provided by an embodiment of the present application.

[0028] Reference Signs:

[0029] 200: Device for detecting an air energy device based on vibration signals, 201: Processor, 202: Memory. Detailed Embodiments

[0030] The embodiments of the present application provide a method, device and medium for detecting an air energy device based on vibration signals.

[0031] To enable those skilled in the art to better understand the technical solutions in this application, the following will clearly and completely describe the technical solutions in the embodiments of this application in conjunction with the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all the embodiments. Based on the embodiments of this specification, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of this application.

[0032] The technical solutions proposed in the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0033] Figure 1 It is a flowchart of a method for detecting an air energy device based on vibration signals provided in an embodiment of this application. As Figure 1 shown, the method for detecting an air energy device based on vibration signals includes the following steps:

[0034] Step 101: Obtain the vibration signals corresponding to different air energy devices to be measured through vibration sensors arranged at different positions of the air energy device to be measured.

[0035] In one implementation manner of this application, multiple vibration sensors are installed at different positions of the air energy device to be measured. These positions may include key components, vibration-sensitive areas, or parts to be monitored of the device. For example, vibration sensors can be arranged on key components such as compressors, fans, condensers, and water pumps respectively. After the vibration signals collected by the vibration sensors are converted into electrical signals, they can be transmitted to the data analysis and processing module.

[0036] Step 102: Denoise the vibration signals through the fusion of an adaptive filter and wavelet transform.

[0037] In one implementation manner of this application, the step size factor is dynamically adjusted based on the initial step size, the maximum step size, and the minimum step size. Based on the dynamically adjusted step size factor, the filter coefficient vector, the error, and the vibration signal, a weight update function of the adaptive filtering algorithm is constructed to iteratively update the filter coefficients. An adaptive filter based on a recurrent neural network is constructed through the iteratively updated filter coefficients. Based on the vibration signal at the current moment and the output corresponding to the adaptive filter at the previous moment, the vibration signal is adaptively filtered. The wavelet transform is performed on the filtered vibration signal, and the error value between the signal after wavelet transform and the current output signal of the adaptive filter is determined. According to the magnitude and direction of the error, the coefficients of the adaptive filter are updated and adjusted. The vibration signal after wavelet transform is sparsely represented on an overcomplete dictionary. Based on the sparse representation coefficients and a preset coefficient threshold, the sparse representation coefficients are screened. The vibration signal is reconstructed through the screened coefficients and the overcomplete dictionary to obtain the denoised vibration signal.

[0038] Specifically, the step-size factor plays a crucial role in the adaptive filtering algorithm, determining the update speed and stability of the filter coefficients. Let the initial step size be μ 0 The maximum step size is μ max and the minimum step size is μ min . The step-size factor is changed according to the variance of the signal or the change of the error. For example, when the error is large, in order to accelerate the convergence speed, the step-size factor can be increased; when the error gradually decreases and tends to be stable, in order to reduce the steady-state error, the step-size factor is decreased. Among them, the dynamic adjustment formula can be expressed as:

[0039]

[0040] where μ(n) is the step-size factor at the nth moment, e(n) is the error at the nth moment, and N is the number of samples within a certain time window before the current moment. According to the proportional relationship between the current error and the maximum error within the time window, the value of the step-size factor between the minimum value and the maximum value is dynamically adjusted.

[0041] Let the filter coefficient vector be w(n), the input vibration signal be x(n), and the desired response be d(n), then the output of the filter is y(n) = w T (n)·x(n), and the error e(n) = d(n) - y(n). The weight update function based on the least mean square algorithm is:

[0042] w(n + 1) = w(n) + 2μ(n)e(n)x(n);

[0043] where 2μ(n)e(n)x(n) is used to adjust the direction and amplitude of the filter coefficients according to the current step-size factor, error, and input signal. By continuously iterating this update formula, the filter coefficients will gradually converge to the optimal value, enabling the filter to better filter the vibration signal.

[0044] Furthermore, the recurrent neural network has a memory function and can handle long-term dependencies in time series data. In the adaptive filter of the embodiment of the present application, in addition to inputting the vibration signal x(n) at the current moment, the output y(n - 1) of the adaptive filter at the previous moment is also considered. The hidden layer of the recurrent neural network processes these input information and outputs the predicted value of the signal at the current moment. The error between the predicted value and the desired response is used to further adjust the filter coefficients, thereby achieving adaptive filtering of the vibration signal.

[0045] Further, perform wavelet transform on the vibration signal after adaptive filtering, and then adjust the coefficients of the adaptive filter by comparing the error between the signal after wavelet transform and the current output signal of the adaptive filter. Specifically, wavelet transform can decompose the signal into sub-band signals of different frequency bands, analyze the signal at different time and frequency resolutions, determine the signal after wavelet transform and the current output signal of the adaptive filter, calculate the error between the two, and adjust the coefficients of the adaptive filter using a weight update function similar to the aforementioned one according to the magnitude and direction of the error. For example, if the error is large and positive, it indicates that the output of the adaptive filter is small, and the relevant coefficient needs to be increased to improve the output; conversely, if the error is negative and large, the relevant coefficient needs to be decreased. In this way, the performance of the adaptive filter is continuously optimized.

[0046] Further, sparsely represent the vibration signal after wavelet transform on an over-complete dictionary, then screen according to the sparse representation coefficients and a preset coefficient threshold, and finally reconstruct to obtain the vibration signal after noise reduction. Specifically, on the over-complete dictionary, the vibration signal can be represented as a linear combination of a few atoms, and the sparse representation of the signal can be obtained by solving an optimization problem. For example, the orthogonal matching pursuit algorithm is used to solve the sparse representation coefficients. Set a preset coefficient threshold, screen the sparse representation coefficient vector, retain the coefficients whose absolute values are greater than the preset coefficient threshold, and set the remaining coefficients to zero to obtain the screened coefficient vector. Then, reconstruct the vibration signal using the screened coefficients and the over-complete dictionary. In this way, the noise and redundant components in the signal are removed, and the vibration signal after noise reduction is obtained.

[0047] Step 103: Perform layer-by-layer encoding and decoding on the vibration signal after noise reduction based on a multi-layer denoising autoencoder to obtain the characteristic data corresponding to different air energy devices to be measured.

[0048] In an implementation manner of the present application, based on the energy distribution and noise level of the vibration signal sample, the low-energy frequency band and the high-noise frequency band are determined. Increase the noise intensity in the low-energy frequency band, and inject the same type of noise as the noise characteristics in the high-noise frequency band. Discriminate between the features after encoding the original signal and the encoded features after noise pollution through a preset discriminator, so as to adjust the parameters of the multi-layer denoising autoencoder based on the discrimination result. Input the vibration signal into the multi-layer denoising autoencoder with adjusted parameters for decoding, determine the first correlation between the encoded feature and the decoded feature through the global attention mechanism, and determine the second correlation between each frequency band in the encoded feature through the local attention mechanism. Assign weights to each encoded feature based on the first correlation and the second correlation to obtain the reconstructed vibration signal, so as to obtain the characteristic data corresponding to different air energy devices to be measured based on the reconstructed vibration signal.

[0049] Specifically, by analyzing the vibration signal samples and calculating the energy distribution of the signal in different frequency intervals, the low-frequency band where the energy is mainly concentrated can be determined. For example, perform Fourier transform on the vibration signal of the air energy device collected, convert it to the frequency domain, and calculate the power spectral density values at different frequency points. Observe the power spectral density curve. In the low-frequency part, the frequency interval where the power value is high and relatively concentrated is the low-frequency energy band; in the high-frequency part, the frequency interval where the amplitude fluctuates greatly and exceeds the normal signal range can be determined as the high-frequency noise band.

[0050] Furthermore, in the determined low-frequency energy band, to enhance the adaptability of the multi-layer denoising autoencoder to low-frequency noise, appropriately increase the noise intensity. It can be achieved by generating a noise signal similar to the original low-frequency band noise characteristics (such as Gaussian white noise), adjusting its amplitude according to needs, and then superimposing it on the low-frequency energy band part of the original vibration signal. In the high-frequency noise band, inject the same type of noise as the noise characteristics of this band. If the noise in the high-frequency noise band is known to be high-frequency electromagnetic noise, generate a noise signal with similar spectral characteristics by simulating the generation mechanism of high-frequency electromagnetic noise, and then inject it into the high-frequency noise band of the original vibration signal.

[0051] Furthermore, the preset discriminator is used to distinguish the features after encoding the original signal from the encoded features after being contaminated by noise. The discriminator is usually constructed based on a neural network, such as a convolutional neural network or a recurrent neural network. It performs feature extraction and classification on the input encoded features to determine whether they come from the original signal or the signal contaminated by noise. Based on the discrimination result, the backpropagation algorithm is used to adjust the parameters of the multi-layer denoising autoencoder. If the discriminator wrongly discriminates the encoded features after being contaminated by noise as the encoded features of the original signal, it indicates that the multi-layer denoising autoencoder has insufficient noise suppression ability and its parameters need to be adjusted to enhance the denoising effect.

[0052] Furthermore, input the vibration signal into the multi-layer denoising autoencoder with adjusted parameters for decoding. When the global attention mechanism calculates the first correlation between the encoded features and the features generated at each moment during the decoding process, it will traverse all the encoded features and the feature representations generated at each moment during the decoding process. Through dot product operation or cosine similarity calculation, evaluate the similarity degree of the encoded feature vector and the decoded feature vector in different dimensions. The local attention mechanism focuses on the relationship between different frequency bands in the encoded features. Since the vibration signal has different frequency components, the features of different frequency bands may be correlated. The local attention mechanism calculates the second correlation by analyzing the relationship between the feature vectors of adjacent frequency bands or frequency bands with a specific frequency interval.

[0053] Further, for each encoded feature, its weight is comprehensively determined according to the first correlation with the decoded feature and the second correlation between frequency bands. First, based on the first correlation, the similarity values between each encoded feature and all decoded features are weighted and summed to obtain the weight based on the first correlation. Then, the weight is adjusted according to the second correlation. If the frequency band where the encoded feature is located has a strong correlation with other important frequency bands, its weight is appropriately increased; otherwise, the weight is appropriately decreased. After determining the weight of each encoded feature, each encoded feature is multiplied by its corresponding weight through the decoding process, and then a reconstructed signal is generated through the operations of the decoding layer. The reconstructed vibration signal contains key information about the operating state of the device. For different air source heat pump devices to be measured, by performing further feature extraction operations on the reconstructed signal, such as time-domain analysis, calculating the mean, variance, peak index, etc.; frequency-domain analysis, extracting frequency components, power spectral density, etc.; or joint time-frequency domain analysis, such as wavelet transform coefficients, etc., the corresponding feature data is obtained.

[0054] Specifically, when assigning weights to each encoded feature based on the first correlation and the second correlation to obtain the reconstructed vibration signal, the embodiment of the present application uses the function:

[0055]

[0056] to assign weights to each encoded feature based on the first correlation and the second correlation; where α is an adjustment parameter; β is an adjustment parameter; W is the weight; G is the global attention score vector; L is the local attention score matrix; L ij is the local correlation between feature i and feature j; E is the encoded feature vector; n is the dimension;

[0057] Based on the function:

[0058] E i ' = W i ·Ei;

[0059] the reconstructed vibration signal is obtained; where E' is the weighted feature vector; W is the weight; E is the encoded feature vector.

[0060] Step 104: Through the preset multi-head self-attention mechanism model, perform associated information analysis on the feature data to determine the probability values of the air source heat pump devices to be measured corresponding to different operating states, so as to determine the air source heat pump device states corresponding to different air source heat pump devices to be measured based on the probability values.

[0061] In an implementation of the present application, the number of self-attention heads is determined based on the dimension and data complexity corresponding to the feature data. In the constructed stacked model structure of the multi-head self-attention mechanism, the output of the next layer is processed based on the output of the previous layer, and the feature representation output by the stacked model of the multi-head self-attention mechanism is non-linearly transformed and feature-integrated through a pre-set feed-forward neural network. By pre-setting the Softmax activation function, the output of the stacked model of the multi-head self-attention mechanism is converted into the probability values of the air energy device to be measured in different operating states.

[0062] Specifically, the dimension and complexity of the feature data are the key factors determining the number of self-attention heads. When the dimension of the feature data is high, it means that the data contains rich information, and more self-attention heads are needed to capture this information from different perspectives. For example, if the feature data covers information in multiple dimensions such as the vibration frequency, temperature, and pressure of the air energy device, the high dimension makes it difficult for a single self-attention head to comprehensively capture the relationships between all dimensions. At this time, increasing the number of self-attention heads, with each head focusing on the feature associations of different dimensions or dimension combinations, can analyze the data more comprehensively.

[0063] Regarding the data complexity, if the feature data presents a complex distribution pattern or there are multiple potential feature interaction relationships, more self-attention heads are also needed to mine these complex relationships. For example, when the air energy device is in a complex working condition, the characteristics of its vibration signal may interact with multiple operating parameters, and this influence relationship is not simply linear. As the data complexity increases, more self-attention heads can analyze the data from different subspaces through parallel computing, so as to better handle this complex situation.

[0064] Furthermore, in the constructed stacked model structure of the multi-head self-attention mechanism, each layer is processed based on the output of the previous layer. The feature representation output by the previous layer contains the preliminary analysis results of the input data, and the next layer further extracts features and mines relationships from these feature representations through the multi-head self-attention mechanism. Each self-attention head performs weighted summation on the input features in different subspaces, focusing on the feature associations in different aspects. For example, one self-attention head may pay more attention to the long-distance dependence relationships between the vibration frequency features, while another head focuses on the association between the vibration frequency and the operating time of the device.

[0065] Furthermore, through multi-layer stacking, the model can gradually and deeply explore the complex feature relationships in the data. The output of each layer further abstracts and integrates features based on the previous layer, enabling the model to understand the data more comprehensively and deeply. For example, the bottom layer may mainly capture the local features and simple associations of the data. As the number of layers increases, the upper layers gradually focus on the global features and complex cross-dimensional associations of the data. For example, taking a 3-layer stacked model of the multi-head self-attention mechanism as an example, the first layer receives the preprocessed feature data of the air energy device, processes the data through 15 self-attention heads, and outputs a set of feature representations, which initially capture the associations between different frequency bands of vibration signals in the feature data. The second layer takes the output of the first layer as input and processes it again through 15 self-attention heads. At this time, the model may capture the association between the vibration signal features and the device energy consumption features. The third layer further explores the potential connections between these features and the device failure history based on the output of the second layer, and outputs more advanced feature representations, providing more valuable information for subsequent analysis.

[0066] Furthermore, in a feed-forward neural network, there are usually multiple fully connected layers. Although the feature representations output by the stacked model of the multi-head self-attention mechanism already contain rich feature association information, further non-linear transformation and feature integration are still required. The feed-forward neural network performs non-linear transformation on the input feature representations through fully connected layers. For example, using the ReLU activation function to increase the non-linear expression ability of the model, enabling the model to learn more complex functional relationships. At the same time, the connection weights between different neurons perform weighted summation on the features to achieve feature integration. In this way, the feature information extracted by the multi-head self-attention mechanism is mapped to a new space, making the feature representations more discriminative and providing more effective data for subsequent determination of the device operating state probability value.

[0067] Furthermore, the role of the Softmax activation function is to convert the output of the stacked model of the multi-head self-attention mechanism after being processed by the feed-forward neural network into the probability values of the air energy device to be tested in different operating states. The Softmax function converts each element in the input vector into a probability value, and the sum of all probability values is 1. In this way, the model output is converted into a probability distribution, and each probability value represents the likelihood of the device being in the corresponding operating state.

[0068] Specifically, converting the output of the stacked model of the multi-head self-attention mechanism into the probability values of the air energy device to be tested in different operating states through the preset Softmax activation function can be:

[0069] Based on the function:

[0070]

[0071] Among them,

[0072] Determine the probability values of the air energy device to be measured in different operating states;

[0073] Among them, P j is the probability value; zj is the output of the stacked multi-head self-attention mechanism model, where n is the number of categories of the operating states of the air energy device; the air energy device state transition weight matrix is W = [w ij n×n ; u is the feature uncertainty coefficient; ρ j is the correction factor; i is the air energy device state category i; j is the air energy device state category j; γ is the first adjustment parameter; is the second adjustment parameter; ACF i (τ) is the autocorrelation function; τ is the delay time; (h i -l i ) is the model uncertainty interval.

[0074] Step 105: Construct a device association graph based on the air energy device state, so as to determine the cross-device maintenance information corresponding to multiple air energy devices to be measured based on the device association graph.

[0075] In an implementation manner of the present application, a device association graph is constructed based on the positions between different air energy devices to be measured, the similarity between different air energy devices to be measured, and the production process relationship between different air energy devices to be measured. In the device association graph, air energy devices in the same state are marked, and a reference air energy device in the fault state is determined. In the device association graph, the fault propagation paths corresponding to each reference air energy device are determined, and based on the fault propagation paths, associated reference air energy devices are determined. The associated reference air energy devices caused by the same fault cause are marked the same. Based on the air energy device markings, the reference air energy devices and the associated reference air energy devices are grouped, so as to construct corresponding maintenance information for different groups of air energy devices.

[0076] ​Specifically, obtain the location coordinates of different air energy devices to be tested. Evaluate the similarity from multiple dimensions of the devices, including device models, operating parameters, fault history, etc. For device models, devices of the same model have similarities in structure and performance, and the types and probabilities of failures may also be similar. In terms of operating parameters, compare parameters such as refrigeration / heating power, compressor speed, and refrigerant pressure of the devices, and determine the similarity by calculating the cosine similarity of the parameter vectors. Also, analyze the roles and upstream and downstream relationships of air energy devices in the production process. For example, in some industrial production scenarios, air energy devices may be used to provide hot water or cold water at a specific temperature for the production process. If the hot water produced by device A is directly supplied to device B for the production link, then there is a strong association between device A and device B. By constructing a directed graph of the production process, clarify the directions of material flow and energy flow between devices, and determine the association relationship and weight. For devices at key production process nodes, their association weights with upstream and downstream devices are higher.

[0077] Furthermore, monitor the operating status of air energy devices in real time, and classify them into states such as normal operation, minor faults, and serious faults. For devices in the same state, use nodes of the same color or shape in the device association graph to represent them. When a fault occurs in an air energy device, screen out devices that are highly similar to the faulty device in terms of model, operating parameters, fault history, etc. and are in a fault state from the device association graph as reference devices.

[0078] Furthermore, in the device association graph, determine the fault propagation path. According to the determined fault propagation path, find other faulty devices or devices with potential fault risks on the path as associated reference devices. Label the associated reference devices caused by the same fault reason with the same label, and the label content includes information such as the fault reason and fault type. For example, if multiple air energy devices have refrigerant leakage faults due to aging of the refrigerant pipeline, these devices are uniformly labeled as the "refrigerant pipeline aging-induced leakage fault group", and the fault reason and possible influence range are detailed in the label, which is convenient for maintenance personnel to quickly understand the commonalities of device faults. Based on the device labels, group the reference devices and the associated reference devices. They can be grouped according to the fault type, such as grouping all devices with compressor faults into one group and grouping devices with electrical faults into another group; they can also be grouped according to the fault reason, such as grouping devices with faults caused by environmental factors into one group. Consider factors such as the location and similarity of the devices during grouping, and try to group devices that are convenient for centralized maintenance into the same group. For different groups of air energy devices, formulate corresponding maintenance information. For the compressor fault group, the maintenance information includes compressor model, replacement process, list of maintenance tools, estimated maintenance time, etc.; for the electrical fault group, the maintenance information covers electrical circuit detection methods, fault component replacement guides, safety precautions, etc. At the same time, reasonably allocate maintenance personnel and resources according to the device grouping situation to improve the maintenance efficiency.

[0079] In an implementation manner of the present application, in the historical air energy device failure database, determine the historical failed air energy device combination corresponding to the reference air energy device, so as to determine the historical data of the air energy device failure based on the historical failed air energy device combination. According to the device association diagram and the historical data of the air energy device failure, construct a Bayesian network model; wherein, in the Bayesian network model, the state of the air energy device is used as a node, the association relationship between air energy devices is used as an edge, and the weight of the edge is used as a conditional probability. Use the reference air energy device as the starting point of the failed air energy device, and determine the propagation probability and the reference propagation path of the failure among the air energy devices through Bayesian network reasoning. Sort the probabilities of different propagation paths, and determine the failure propagation path and the associated reference air energy device based on the sorting result.

[0080] Specifically, in the historical air energy device failure database, for the determined reference air energy device, find the historical failed air energy devices similar to it in terms of failure characteristics by querying fields such as device model, failure type, and failure occurrence time, and form a historical failed air energy device combination. For the formed historical failed air energy device combination, extract its detailed failure historical data. These data include the operating parameters of the device before the failure, such as refrigeration / heating power, refrigerant pressure, compressor speed, etc., the environmental parameters at the time of the failure, such as machine room temperature, humidity, etc., the frequency of the failure occurrence, and the failure repair records, such as repair time, repair measures, replaced parts, etc.

[0081] Furthermore, construct a Bayesian network model based on the device association diagram. Use the states of the air energy devices, such as normal operation, minor failure, serious failure, etc., as nodes. The association relationships between devices, such as the connection relationships established based on proximity in location, similarity in model, upstream and downstream of the production process, etc., are used as edges. The weight of the edge is determined according to the method determined when constructing the device association diagram before. For example, the connection weight of devices close in location may be calculated based on the Euclidean distance, and the connection weight of devices with the same model may be evaluated based on the similarity of structure and performance. These weights will be used as the conditional probabilities in the Bayesian network. Use the determined historical data of the air energy device failure to calculate the conditional probabilities between nodes through statistical methods. For example, count the number of times that when device A is in a minor failure state, the associated device B is also in a minor failure state in the historical data, and divide it by the total number of times that device A is in a minor failure state to obtain the conditional probability that device B is in a minor failure state when device A is in a minor failure state. By calculating and setting the conditional probabilities between all node pairs, construct a complete Bayesian network model.

[0082] Further, taking the reference air energy device as the starting point of the faulty air energy device, the inference algorithm of the Bayesian network, such as the joint tree algorithm, is used for inference in the constructed Bayesian network model. The algorithm calculates the propagation probability of the fault among the air energy devices according to the fault state of the starting device and the conditional probabilities between the nodes. The probabilities of different propagation paths are sorted in descending order. Based on the sorting results, the path with the highest probability is selected as the most likely fault propagation path. At the same time, other devices on the path (such as device B) and other devices associated with the starting device through this path are the associated reference air energy devices.

[0083] Figure 2 FIG. is a schematic structural diagram of an air energy device detection device based on vibration signals provided by an embodiment of the present application. As Figure 2 shown, the air energy device detection device 200 based on vibration signals includes: at least one processor 201; and a memory 202 communicatively connected to the at least one processor 201; wherein, the memory 202 stores instructions executable by the at least one processor 201, and the instructions are executed by the at least one processor 201 to enable the at least one processor 201 to: obtain vibration signals respectively corresponding to different air energy devices to be measured through vibration sensors arranged at different positions of the air energy devices to be measured; perform noise reduction processing on the vibration signals through the fusion of an adaptive filter and wavelet transform; perform layer-by-layer encoding and decoding on the noise-reduced vibration signals based on a multi-layer noise reduction autoencoder to obtain feature data respectively corresponding to different air energy devices to be measured; perform associated information analysis on the feature data through a preset multi-head self-attention mechanism model to determine the probability values of the air energy devices to be measured corresponding to the feature data in different operating states, so as to determine the air energy device states respectively corresponding to different air energy devices to be measured based on the probability values; construct a device association graph based on the air energy device states, so as to determine cross-device maintenance information corresponding to multiple air energy devices to be measured based on the device association graph.

[0084] A non-volatile computer storage medium provided by an embodiment of the present application stores computer-executable instructions, and the computer-executable instructions are set to: obtain vibration signals respectively corresponding to different air energy devices to be measured through vibration sensors arranged at different positions of the air energy devices to be measured; perform noise reduction processing on the vibration signals through the fusion of an adaptive filter and wavelet transform; perform layer-by-layer encoding and decoding on the noise-reduced vibration signals based on a multi-layer noise reduction autoencoder to obtain feature data respectively corresponding to different air energy devices to be measured; perform associated information analysis on the feature data through a preset multi-head self-attention mechanism model to determine probability values corresponding to different operating states of the air energy devices to be measured corresponding to the feature data, so as to determine the air energy device states respectively corresponding to different air energy devices to be measured based on the probability values; construct a device association graph based on the air energy device states, so as to determine cross-device maintenance information corresponding to multiple air energy devices to be measured based on the device association graph.

[0085] The embodiments in the present application are all described in a progressive manner. For the same or similar parts among the embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the device, air energy device, and non-volatile computer storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can refer to the partial description of the method embodiments.

[0086] The above are only the embodiments of the present application and are not used to limit the present application. For those skilled in the art, various changes and modifications can be made to the embodiments of the present application. These modifications or substitutions do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for detecting air energy equipment based on vibration signals, characterized in that: The method comprises: Vibration sensors are arranged at different positions of the air energy equipment to be tested to obtain vibration signals corresponding to different air energy equipment to be tested; The vibration signal is subjected to noise reduction processing by fusing an adaptive filter with a wavelet transform; Based on a multi-layer noise reduction autoencoder, the vibration signal after noise reduction is encoded and decoded layer by layer to obtain characteristic data corresponding to the different air energy devices to be tested; By presetting a multi-head self-attention mechanism model, the feature data is analyzed for correlation information, and the probability values ​​of the air energy equipment to be tested corresponding to the feature data being in different operating states are determined, so as to determine the air energy equipment states corresponding to the different air energy equipment to be tested respectively based on the probability values; Based on the state of the air energy equipment, a device association diagram is constructed to determine cross-device maintenance information corresponding to a plurality of air energy equipment to be tested based on the device association diagram.

2. The air energy equipment detection method based on vibration signal according to claim 1, characterized in that: The denoising process of the vibration signal is performed by fusing the adaptive filter with the wavelet transform, specifically comprising: Dynamically adjust the step size factor based on the initial step size, the maximum step size, and the minimum step size; Based on the dynamically adjusted step size factor, the filter coefficient vector, the error and the vibration signal, a weight update function of the adaptive filtering algorithm is constructed to iteratively update the filter coefficient; An adaptive filter based on a recursive neural network is constructed by iteratively updating the filter coefficients, and the vibration signal is adaptively filtered based on the vibration signal at the current moment and the output corresponding to the adaptive filter at the previous moment; Performing wavelet transformation on the filtered vibration signal, determining an error value between the wavelet transformed signal and the current output signal of the adaptive filter, and updating and adjusting the coefficients of the adaptive filter according to the error size and error direction; Sparsely representing the vibration signal after wavelet transformation on an overcomplete dictionary; Based on the sparse representation coefficient and a preset coefficient threshold, the sparse representation coefficient is screened; The vibration signal is reconstructed by using the filtered coefficients and the overcomplete dictionary to obtain a vibration signal after noise reduction.

3. The air energy equipment detection method based on vibration signal according to claim 1, characterized in that: The multi-layer noise reduction autoencoder is used to encode and decode the noise-reduced vibration signal layer by layer to obtain characteristic data corresponding to the different air energy devices to be tested, specifically including: Based on the energy distribution and noise level corresponding to the vibration signal samples, the low-frequency band of energy and the high-frequency band of noise are determined; Increase the noise intensity in the energy low frequency band, and inject a noise type with the same noise characteristics as the noise in the noise high frequency band; The features after encoding the original signal and the features after being polluted by noise are discriminated by a preset discriminator, so as to adjust the parameters of the multi-layer denoising autoencoder based on the discriminant result; Decoding the multi-layer denoising autoencoder after adjusting the vibration signal input parameters, determining a first correlation between the encoding feature and the decoding feature through a global attention mechanism, and determining a second correlation between frequency bands in the encoding feature through a local attention mechanism; Based on the first correlation and the second correlation, weights are assigned to the coding features to obtain a reconstructed vibration signal, so as to obtain characteristic data corresponding to the different air energy devices to be tested based on the reconstructed vibration signal.

4. The air energy equipment detection method based on vibration signal according to claim 1 is characterized in that: The step of assigning weights to the coding features based on the first correlation and the second correlation to obtain a reconstructed vibration signal specifically includes: Through the function: The first correlation and the second correlation are weighted for each of the encoding features; wherein α is an adjustment parameter; β is an adjustment parameter; W is a weight; G is a global attention score vector; L is a local attention score matrix; L ij is the local correlation between feature i and feature j; E is the encoded feature vector; n is the dimension; Function-based: HAVE BEEN i '=W i ·Oi; The reconstructed vibration signal is obtained; wherein, E' is the weighted feature vector; W is the weight; and E is the encoded feature vector.

5. The air energy equipment detection method based on vibration signal according to claim 1, characterized in that: The presetting of the multi-head self-attention mechanism model to analyze the associated information of the feature data to determine the probability value of the air energy equipment to be tested corresponding to the feature data being in different operating states specifically includes: Determine the number of attention heads based on the dimension and data complexity corresponding to the feature data; In the constructed multi-head self-attention mechanism stacking model structure, the output of the next layer is processed based on the output of the previous layer, and the feature representation of the output of the multi-head self-attention mechanism stacking model is nonlinearly transformed and integrated through a preset feedforward neural network; By presetting a Softmax activation function, the output of the multi-head self-attention mechanism stacking model is converted into a probability value that the air energy device to be tested is in different operating states.

6. The air energy equipment detection method based on vibration signal according to claim 5, characterized in that: The method of converting the output of the multi-head self-attention mechanism stacking model into a probability value of the air energy device to be tested being in different operating states by presetting a Softmax activation function specifically includes: Function-based: in, Determining the probability values ​​of the air energy equipment to be tested being in different operating states; Among them, P j is the probability value; zj is the output of the multi-head self-attention mechanism stacking model, where n is the number of categories of air energy equipment operation status; the air energy equipment state transfer weight matrix is ​​W = [w ij ] n×n ; u is the characteristic uncertainty coefficient; ρ j is the correction factor; i is the air energy equipment state category i; j is the air energy equipment state category j; γ is the first adjustment parameter; is the second adjustment parameter; ACF i (τ) is the autocorrelation function; τ is the delay time; (h i -l i ) is the model uncertainty interval.

7. The air energy equipment detection method based on vibration signal according to claim 1, characterized in that: The step of constructing a device association diagram based on the state of the air energy device to determine cross-device maintenance information corresponding to a plurality of air energy devices to be tested based on the device association diagram specifically includes: Building a device association diagram based on the positions of different air energy devices to be tested, the similarities between the different air energy devices to be tested, and the production process relationships between the different air energy devices to be tested; In the device association diagram, air energy devices in the same state are marked, and reference air energy devices in a fault state are determined; Determine the fault propagation paths corresponding to the reference air energy devices in the device association diagram, and determine the associated reference air energy devices based on the fault propagation paths; The same marking shall be applied to the associated reference air energy equipment caused by the same fault cause; The reference air energy devices and the associated reference air energy devices are grouped based on the air energy device labels, so as to construct corresponding maintenance information for the air energy devices in different groups.

8. The air energy equipment detection method based on vibration signal according to claim 7 is characterized in that: The step of determining the fault propagation paths respectively corresponding to the reference air energy devices in the device association diagram, and determining the associated reference air energy devices based on the fault propagation paths specifically includes: In a historical air energy equipment failure database, determining a historical fault air energy equipment combination corresponding to the reference air energy equipment, and determining air energy equipment failure history data based on the historical fault air energy equipment combination; A Bayesian network model is constructed according to the device association diagram and the air energy device fault history data; wherein, in the Bayesian network model, the air energy device status is used as a node, the association relationship between the air energy devices is used as an edge, and the weight of the edge is used as a conditional probability; Taking the reference air energy device as the starting point of the faulty air energy device, the propagation probability of the fault between the air energy devices and the reference propagation path are determined through Bayesian network reasoning; The probabilities of different propagation paths are sorted, and the fault propagation path and the associated reference air energy device are determined based on the sorting result.

9. An air energy equipment detection device based on vibration signals, characterized in that: The device comprises a memory for storing computer program instructions and a processor for executing the program instructions, wherein when the computer program instructions are executed by the processor, the device is triggered to execute the method according to any one of claims 1 to 8.

10. A non-volatile computer storage medium storing computer executable instructions, characterized in that: The computer executable instructions can execute the method according to any one of claims 1 to 8.

Images