Equipment health monitoring system and method based on edge computing

Abstract

The invention relates to the technical field of equipment monitoring analysis, in particular to an equipment health monitoring system and method based on edge computing, and the method comprises the steps: collecting multi-dimensional data of target equipment in real time, and carrying out the data preprocessing of the multi-dimensional data through an edge computing node, and obtaining the processing data of the target equipment; performing data analysis on the processed data, including performing deep analysis on the processed data in combination with a deep convolutional neural network and long and short term data features, constructing a risk assessment model, and identifying a data risk value of data of each dimension in the target equipment based on the risk assessment model; and pushing risk monitoring information to operation and maintenance personnel based on the data risk value of the target equipment. According to the method, the deep convolutional neural network and long and short term data features are integrated, and a multi-scale feature extraction and weighted fusion strategy is adopted, so that the defect of spatial-temporal feature splitting in the prior art is overcome, and the robustness under complex working conditions is enhanced while the real-time performance is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment monitoring and analysis, and particularly to an equipment health monitoring system and method based on edge computing. Background Art

[0002] Equipment health monitoring is an important item in equipment management. Equipment health monitoring based on edge computing can significantly reduce data transmission latency and avoid bandwidth pressure caused by directly uploading massive raw signals to the cloud. By completing preprocessing such as time-frequency transformation and denoising and preliminary feature extraction at the proximal end of the equipment, the real-time nature of data analysis is ensured.

[0003] The Chinese invention patent with the publication number CN118396446A discloses a method for monitoring the health status of power equipment. The method includes obtaining state parameters inside the power equipment through sensors arranged inside the power equipment; preprocessing the state parameters, and extracting state features from the preprocessed state parameters; establishing a state monitoring model according to the state features, performing real-time state monitoring on the power equipment according to the state monitoring model, and generating a state monitoring score; performing threshold determination according to the state monitoring score, and when it is greater than a preset threshold, generating an alarm message for alarm.

[0004] However, it is difficult to extract the spatial features of data from the technical content of the above patent document, and it is difficult to analyze the relationship between data and data based on time-series data. It is difficult to combine short-term transient changes and long-term regular changes to monitor and analyze the equipment, and there are certain limitations in the analysis of the spatial features and long-short-term changes of data. Summary of the Invention

[0005] The object of the present invention is to propose an equipment health monitoring system and method based on edge computing for the problems existing in the background art.

[0006] The technical solution of the present invention: An equipment health monitoring method based on edge computing includes the following steps: Real-time collect multi-dimensional data of the target equipment, and use an edge computing node to perform data preprocessing on the multi-dimensional data to obtain processed data of the target equipment; the multi-dimensional data includes equipment vibration signals, temperature time series, and pressure time series; Perform data analysis on the processed data, including deeply analyzing the processed data by combining a deep convolutional neural network and long-short-term data features, constructing a risk assessment model, and identifying the data risk value of each dimension data in the target equipment based on the risk assessment model; Push risk monitoring information to the operation and maintenance personnel based on the data risk value of the target equipment.

[0007] Preferably, the method for performing data preprocessing on the multi-dimensional data includes: The vibration signal is transformed into a time-frequency diagram through continuous wavelet transform, non-local means denoising is applied to eliminate environmental interference, and adaptive histogram equalization is used to enhance the feature contrast; The DTW algorithm is used to align the sampling frequency differences of sensors with different time series, the Encouraged Forest algorithm is used to detect outliers, the outliers are removed, and interpolation is performed based on cubic splines for repair.

[0008] Preferably, the method for data analysis of the processed data includes the following steps: Extract short-term local features from the time-frequency diagram to obtain the first convolutional feature map; Extract long-range periodic features from the time-frequency diagram to obtain the second convolutional feature map; Based on the first convolutional feature map and the second convolutional feature map, determine the n-dimensional convolutional feature vector; Perform sequence analysis on the temperature time series and the pressure time series respectively to determine the corresponding time series feature vectors; Construct a risk assessment model based on the n-dimensional convolutional feature vector and the n-dimensional time series feature vector.

[0009] Preferably, the short-term local feature extraction includes performing a boundary padding operation on the original time-frequency diagram and marking the padded original time-frequency diagram as the first padded time-frequency diagram; Perform a convolutional sliding operation on the first padded time-frequency diagram using two-dimensional discrete convolution, including configuring the first convolutional kernel and constraining the convolutional output size to be the same as the original time-frequency diagram input size; Slide the first convolutional kernel on the first padded time-frequency diagram, calculate the local dot product of the first padded time-frequency diagram, and when the first convolutional kernel finishes the calculation, activate the first time-frequency diagram using the Gaussian error linear unit and perform a max pooling operation to output the first convolutional feature map.

[0010] Preferably, the long-range periodic feature extraction includes performing an equidistant padding operation on the original time-frequency diagram and marking the padded original time-frequency diagram as the second padded time-frequency diagram, and performing a dilated convolution operation on the second padded time-frequency diagram; The dilated convolution operation includes configuring the second convolutional kernel, performing interval sampling on the second padded time-frequency diagram by the second convolutional kernel at the dilation rate D, activating the second padded time-frequency diagram using the same Gaussian error linear unit as above and performing an adaptive pooling operation to output the second convolutional feature map; Constrain the channel dimensions of the first convolutional kernel and the second convolutional kernel so that the number of channels of the first convolutional kernel is the same as the number of channels of the second convolutional kernel.

[0011] Preferably, based on the first convolutional feature map and the second convolutional feature map, the method for determining the n-dimensional convolutional feature vector includes: Perform bilinear interpolation downsampling on the second convolutional feature map to adapt to the resolution of the first convolutional feature map. Weightedly splice the first convolutional feature map and the second convolutional feature map along the channel dimension to obtain an n-channel feature map. Perform global average pooling on the n-channel feature map to obtain an n-dimensional convolutional feature vector; n is the number of channels.

[0012] Preferably, perform sequence analysis on the temperature time series and the pressure time series respectively to determine the corresponding n-dimensional time series feature vectors. The method includes: Take the temperature time series and the pressure time series as the target time series in turn, and perform the following operations on the target time series: Based on the target time series, construct a target time subsequence with the current moment as the end moment. The number of sequence elements in the target time subsequence is n; For any sequence element in the target time subsequence, obtain the first N sequence elements of this sequence element in the target time series, obtain the variance, mean, and median of the N sequence elements, and use the variance, mean, and median as analysis values to construct a feature analysis set of this element composed of different analysis values. Calculate the feature analysis set to determine the feature value of the corresponding sequence element; Calculate the feature value of the sequence element through the following formula: ; In the formula, , and represent variance, mean, and median respectively; represents the weight of the analysis value; i is the analysis value number, and i is a positive integer; Based on the feature values of each sequence element in the target time subsequence, construct an n-dimensional time series feature vector of the target time subsequence.

[0013] Preferably, construct a risk assessment model based on the n-dimensional convolutional feature vector and the n-dimensional time series feature vector. The method includes: Normalize the n-dimensional convolutional feature vector and the n-dimensional time series feature vector to obtain an n-dimensional convolutional processed vector and an n-dimensional time series processed vector; Vector splice the n-dimensional convolutional processed vector and different n-dimensional time series processed vectors to obtain an n-dimensional spliced vector and use it as the standard vector; calculate the cosine similarity between the n-dimensional convolutional processed vector and different n-dimensional time series processed vectors and the standard vector respectively to obtain the convolutional processed similarity and the time series processed similarity.

[0014] Preferably, the expression of the risk assessment model is as follows: ; In the formula; S is the processed similarity; The similarity threshold corresponding to the processing similarity is obtained based on big data testing of the target device in the normal operating state; D is the risk value of the processing similarity; Mark the risk values of the convolutional processing similarity and the temporal processing similarity as D0 and Dj respectively; j is the temporal processing similarity number, j = (1, 2); Determine the vibration risk value according to D0, determine the temperature risk value according to D1, and determine the pressure risk value according to D2.

[0015] The present invention also discloses a device health monitoring system based on edge computing, which applies the above-mentioned device health monitoring method based on edge computing, and specifically includes: A data acquisition module, which is used to collect multi-dimensional data of the target device in real time, and use the edge computing node to perform data preprocessing on the multi-dimensional data to obtain the processed data of the target device; A data analysis module, which is used to perform data analysis on the processed data, including performing in-depth analysis on the processed data by combining a deep convolutional neural network and long-term and short-term data features, constructing a risk assessment model, and identifying the data risk value of each dimension data in the target device based on the risk assessment model; A monitoring and warning module, which is used to push risk monitoring information to the operation and maintenance personnel based on the data risk value of the target device.

[0016] Compared with the prior art, the above technical solution of the present invention has the following beneficial technical effects: By integrating a deep convolutional neural network and long-term and short-term data features, adopting a multi-scale feature extraction and weighted fusion strategy, effectively combining the time-frequency characteristics of vibration signals with the temporal evolution law of temperature / pressure, solving the defect of the traditional method of spatio-temporal feature fragmentation, enhancing the robustness under complex working conditions while ensuring real-time performance, and providing an end-to-end intelligent monitoring framework for industrial equipment health management. Brief Description of the Drawings

[0017] Figure 1 It is a method block diagram of Embodiment 1 proposed by the present invention. Detailed Embodiments

[0018] Embodiment 1, as Figure 1 shown, the device health monitoring method based on edge computing proposed by the present invention includes the following steps: Collect multi-dimensional data of the target device in real time, and use the edge computing node to perform data preprocessing on the multi-dimensional data to obtain the processed data of the target device; the multi-dimensional data includes device vibration signals, temperature time series, and pressure time series; it should be noted that the vibration signals are collected by vibration sensors, and the temperature time series and pressure time series are obtained by corresponding time series sensors; The method for performing data preprocessing on the multi-dimensional data includes: The vibration signal is converted into a time-frequency diagram through continuous wavelet transform, non-local means denoising is applied to eliminate environmental interference, and adaptive histogram equalization is used to enhance the feature contrast; The DTW algorithm is used to align the sampling frequency differences of sensors with different time series. The encouraging forest algorithm is used to detect abnormal points, the abnormal points are removed, and interpolation is performed based on cubic splines; among them, the DTW algorithm includes: selecting the sensor with the highest time coverage as the reference, and defining the corresponding time series as the reference sequence, and the time series corresponding to other sensors as the sequence to be aligned. Each sequence is segmented by a sliding window. For the sequence data in each window, the optimal warping path between the reference sequence and the sequence to be aligned is calculated, and the timestamp remapping relationship of the sequence to be aligned to the reference sequence is established based on the optimal warping path; It should be noted that continuous wavelet transform, non-local means denoising, adaptive histogram equalization, DTW algorithm, encouraging forest algorithm, and spline interpolation algorithm are all existing technologies and will not be elaborated here one by one; Data analysis is performed on the processed data, including in-depth analysis of the processed data by combining deep convolutional neural network and long-term and short-term data features, constructing a risk assessment model, and identifying the data risk values of each dimension data in the target device based on the risk assessment model; The method for performing data analysis on the processed data includes: Short-term local feature extraction is performed on the time-frequency diagram to obtain the first convolutional feature map; the first convolutional feature map is used to capture the local transient of the vibration signal of the target device; Long-range periodic feature extraction is performed on the time-frequency diagram to obtain the second convolutional feature map; the first convolutional feature map is used to capture the periodic pattern of the target device; Short-term local feature extraction includes performing boundary padding on the original time-frequency diagram and marking the padded original time-frequency diagram as the first padded time-frequency diagram; Exemplarily, if the original edge pixels are [a, b, c, d], then after padding, it is [d, c, b, a, d, c]); another example is that the size of the original input time-frequency diagram is: 128×128×32, and the size of the output padded time-frequency diagram is extended to 132×132×32; A two-dimensional discrete convolution is used to perform a convolution sliding operation on the first padded time-frequency diagram, including configuring the first convolution kernel, and obtaining the output result after passing through the standard convolution kernel through the following formula: ; In the formula, is the size of the padded time-frequency diagram; is the convolution output size; P is the preset padding value of the first convolution kernel; K is the size of the first convolution kernel; S is the step size; Constrain the convolutional output size to be the same as the input size of the original time-frequency map; Slide the first convolutional kernel on the first padded time-frequency map to calculate the local dot product of the first padded time-frequency map. After the first convolutional kernel finishes the calculation, activate the first time-frequency map using the Gaussian error linear unit and perform a max pooling operation to output the first convolutional feature map; The long-term periodic feature extraction includes performing an equidistant padding operation on the original time-frequency map and marking the padded original time-frequency map as the second padded time-frequency map, and performing a dilated convolution operation on the second padded time-frequency map; The equidistant padding logic is based on the following formula: ; where, is the dilation size; ; where, D is the dilation rate, determined based on the amount of pixels for equidistant padding; The dilated convolution operation includes configuring the second convolutional kernel, performing interval sampling on the second padded time-frequency map by the second convolutional kernel at the dilation rate D, activating the second padded time-frequency map using the same Gaussian error linear unit as above and performing an adaptive pooling operation to output the second convolutional feature map; Constrain the channel dimensions of the first convolutional kernel and the second convolutional kernel to make the number of channels of the first convolutional kernel the same as the number of channels of the second convolutional kernel; It should be noted that the size of the second convolutional kernel is smaller than the size of the first convolutional kernel; Based on the first convolutional feature map and the second convolutional feature map, determine the n-dimensional convolutional feature vector. The method includes: Perform bilinear interpolation downsampling on the second convolutional feature map to adapt to the resolution of the first convolutional feature map, perform weighted concatenation of the first convolutional feature map and the second convolutional feature map along the channel dimension to obtain an n-channel feature map, and perform global average pooling on the n-channel feature map to obtain the n-dimensional convolutional feature vector; n is the number of channels; It should be noted that the weighted concatenation is to perform weight assignment on the first convolutional feature map obtained by short-term local feature extraction and the second convolutional feature map obtained by long-range periodic feature extraction respectively, and perform a concatenation operation on the first convolutional feature map and the second convolutional feature map based on the weight assignment; the weight assignment is determined based on the proportion of the influence of local transient anomalies and long-range periodic anomalies of the target device; Perform sequence analysis on the temperature time series and the pressure time series respectively to determine the corresponding n-dimensional time series feature vectors. The method includes: Take the temperature time series and the pressure time series as the target time series in turn, and perform the following operations on the target time series: Based on the target time series, construct a target time subsequence with the current moment as the end moment, and the number of sequence elements of the target time subsequence is n; For any sequence element in the target time subsequence, obtain the first N sequence elements of this sequence element in the target time series, obtain the variance, mean, and median of the N sequence elements, and use the variance, mean, and median as analysis values to construct a characteristic analysis set of this element composed of different analysis values, and calculate the characteristic analysis set to determine the characteristic value of the corresponding sequence element; Calculate the characteristic value of the sequence element through the following formula: ; In the formula, , and respectively represent variance, mean, and median; represents the weight of the analysis value; i is the analysis value number, and i is a positive integer; Based on the characteristic values of each sequence element in the target time subsequence, construct an n-dimensional time series feature vector of the target time subsequence; Construct a risk assessment model based on the n-dimensional convolutional feature vector and the n-dimensional time series feature vector. The method includes: Normalize the n-dimensional convolutional feature vector and the n-dimensional time series feature vector to obtain an n-dimensional convolutional processing vector and an n-dimensional time series processing vector; it should be noted that the normalization method used here is Min-Max normalization processing, which maps the value of each dimension to the interval [0, 1]. The formula is: ; In the formula, is the processed value of the value x of each dimension after normalization processing; and are the dimension minimum value and the dimension maximum value respectively; Concatenate the n-dimensional convolutional processing vector and different n-dimensional time series processing vectors to obtain an n-dimensional concatenated vector and use it as the standard vector; calculate the cosine similarity between the n-dimensional convolutional processing vector and different n-dimensional time series processing vectors and the standard vector respectively to obtain the convolutional processing similarity and the time series processing similarity; The expression for constructing the risk assessment model is as follows: ; In the formula; S is the processing similarity; is the similarity threshold corresponding to the processing similarity, obtained based on the big data test of the target device in the normal operation state; D is the risk value of the processing similarity; Mark the risk values of the convolutional processing similarity and the time series processing similarity as D0 and Dj respectively; j is the time series processing similarity number, j = (1, 2); it should be noted that D1 is the risk value corresponding to the temperature data, and D2 is the risk value corresponding to the pressure data; Determine the vibration risk value according to D0, determine the temperature risk value according to D1, and determine the pressure risk value according to D2; Push risk monitoring information to the operation and maintenance personnel based on the data risk value of the target device.

[0019] Embodiment 2, the device health monitoring system based on edge computing proposed by the present invention is applied to the device health monitoring method based on edge computing proposed in Embodiment 1, and specifically includes: A data acquisition module, configured to collect multi-dimensional data of the target device in real time, and use an edge computing node to perform data preprocessing on the multi-dimensional data to obtain the processed data of the target device; A data analysis module, configured to perform data analysis on the processed data, including performing in-depth analysis on the processed data by combining a deep convolutional neural network and long-term and short-term data features, constructing a risk assessment model, and identifying the data risk value of each dimension data in the target device based on the risk assessment model; A monitoring and warning module, configured to push risk monitoring information to the operation and maintenance personnel based on the data risk value of the target device.

[0020] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto, and various changes can be made without departing from the spirit of the present invention within the scope of knowledge possessed by those skilled in the art.

Claims

1. A device health monitoring method based on edge computing, characterized in that It includes the following steps: Collect multi-dimensional data of the target device in real time, and use the edge computing node to perform data preprocessing on the multi-dimensional data to obtain the processed data of the target device; the multi-dimensional data includes device vibration signals, temperature time series, and pressure time series; Perform data analysis on the processed data, including performing in-depth analysis on the processed data by combining a deep convolutional neural network and long-term and short-term data features, constructing a risk assessment model, and identifying the data risk values of each dimension of data in the target device based on the risk assessment model; Push risk monitoring information to the operation and maintenance personnel based on the data risk value of the target device.

2. The device health monitoring method based on edge computing according to claim 1, wherein The method for performing data preprocessing on the multi-dimensional data includes: Convert the vibration signal into a time-frequency diagram through continuous wavelet transform, apply non-local means denoising to eliminate environmental interference, and use adaptive histogram equalization to enhance feature contrast; Use the DTW algorithm to align the sampling frequency differences of different time-series sensors, use the boosting forest algorithm to detect outliers, remove the outliers, and perform repair based on cubic spline interpolation.

3. The device health monitoring method based on edge computing according to claim 1, characterized in that The method for performing data analysis on the processed data includes: Extract short-term local features from the time-frequency diagram to obtain the first convolutional feature map; Extract long-term periodic features from the time-frequency diagram to obtain the second convolutional feature map; Based on the first convolutional feature map and the second convolutional feature map, determine the n-dimensional convolutional feature vector; Perform sequence analysis on the temperature time series and the pressure time series respectively to determine the corresponding time-series feature vectors; Construct a risk assessment model based on the n-dimensional convolutional feature vector and the n-dimensional time-series feature vector.

4. The device health monitoring method based on edge computing according to claim 3, characterized in that The short-term local feature extraction includes performing boundary padding on the original time-frequency diagram and marking the padded original time-frequency diagram as the first padded time-frequency diagram; Perform a convolutional sliding operation on the first padded time-frequency diagram using two-dimensional discrete convolution, including configuring the first convolutional kernel and constraining the convolutional output size to be the same as the original time-frequency diagram input size; Use the first convolutional kernel to slide on the first padded time-frequency diagram, calculate the local dot product of the first padded time-frequency diagram, and when the first convolutional kernel finishes the calculation, use the Gaussian error linear unit to activate the first time-frequency diagram and perform a max pooling operation to output the first convolutional feature map.

5. The device health monitoring method based on edge computing according to claim 4, characterized in that The long-term periodic feature extraction includes performing equidistant padding on the original time-frequency diagram and marking the padded original time-frequency diagram as the second padded time-frequency diagram, and performing dilated convolution on the second padded time-frequency diagram; The dilated convolution operation includes configuring the second convolutional kernel, performing interval sampling on the second padded time-frequency diagram by the second convolutional kernel at the dilation rate D, using the same Gaussian error linear unit as above to activate the second padded time-frequency diagram and perform an adaptive pooling operation to output the second convolutional feature map; Constrain the channel dimensions of the first convolutional kernel and the second convolutional kernel so that the number of channels of the first convolutional kernel is the same as the number of channels of the second convolutional kernel.

6. The method for device health monitoring based on edge computing according to claim 5, wherein, Based on the first convolutional feature map and the second convolutional feature map, determine the n-dimensional convolutional feature vector. The method includes: Perform bilinear interpolation downsampling on the second convolutional feature map to adapt to the resolution of the first convolutional feature map, and perform weighted splicing of the first convolutional feature map and the second convolutional feature map along the channel dimension to obtain an n-channel feature map. Perform global average pooling on the n-channel feature map to obtain an n-dimensional convolutional feature vector; n is the number of channels.

7. The method for device health monitoring based on edge computing according to claim 6, characterized in that, Perform sequence analysis on the temperature time series and the pressure time series respectively to determine the corresponding n-dimensional time series feature vectors. The methods include: Take the temperature time series and the pressure time series as the target time series in turn, and perform the following operations on the target time series: Based on the target time series, construct a target time subsequence with the current moment as the end moment, and the number of sequence elements in the target time subsequence is n; For any sequence element in the target time subsequence, obtain the first N sequence elements of the sequence element in the target time series, obtain the variance, mean, and median of the N sequence elements, and use the variance, mean, and median as analysis values to construct a feature analysis set of the element composed of different analysis values. Calculate the feature analysis set to determine the feature value of the corresponding sequence element; Calculate the feature value of the sequence element through the following formula: ； Wherein, , and represent variance, mean value and median respectively; represents the weight of the analysis value; i is the analysis value number, and i is a positive integer; Based on the feature values of each sequence element in the target time subsequence, construct an n-dimensional time series feature vector of the target time subsequence.

8. The method for device health monitoring based on edge computing according to claim 7, characterized in that Construct a risk assessment model based on the n-dimensional convolutional feature vector and the n-dimensional time series feature vector. The methods include: Normalize the n-dimensional convolutional feature vector and the n-dimensional time series feature vector to obtain an n-dimensional convolutional processing vector and an n-dimensional time series processing vector; Perform vector splicing on the n-dimensional convolutional processing vector and different n-dimensional time series processing vectors to obtain an n-dimensional splicing vector and use it as the standard vector; calculate the cosine similarity between the n-dimensional convolutional processing vector and different n-dimensional time series processing vectors and the standard vector respectively to obtain the convolutional processing similarity and the time series processing similarity.

9. The method for device health monitoring based on edge computing according to claim 8, characterized in that, The expression for constructing the risk assessment model is as follows: ； wherein; S is the processing similarity; is the similarity threshold corresponding to the processing similarity, obtained based on big data testing of the target device under normal operating conditions; D is the risk value of the processing similarity; Mark the risk values of the convolutional processing similarity and the time series processing similarity as D0 and Dj respectively; j is the time series processing similarity number, j = (1, 2); Determine the vibration risk value according to D0, determine the temperature risk value according to D1, and determine the pressure risk value according to D2.

10. An edge-computing-based device health monitoring system, which is applied to the edge-computing-based device health monitoring method according to any one of claims 1 to 9 above, and is characterized in that, Specifically include: A data acquisition module for real-time collecting multi-dimensional data of the target device, and using an edge computing node to perform data preprocessing on the multi-dimensional data to obtain the processed data of the target device; A data analysis module for performing data analysis on the processed data, including performing in-depth analysis on the processed data by combining a deep convolutional neural network and long-term and short-term data features, constructing a risk assessment model, and identifying the data risk value of each dimension data in the target device based on the risk assessment model; A monitoring and warning module for pushing risk monitoring information to the operation and maintenance personnel based on the data risk value of the target device.

Images