Vibration and temperature multi-parameter fusion early warning device for special equipment for fuel maintenance

Abstract

The invention relates to the technical field of fault early warning, in particular to a vibration and temperature multi-parameter fusion early warning device for fuel maintenance special equipment. The method comprises the steps that a signal collecting and processing unit collects and processes a vibration signal and a temperature signal in real time in the equipment operation process; the feature extraction and fusion unit extracts vibration features and temperature change features based on the processed vibration signals and temperature signals, and forms uniform feature vectors through an attention mechanism; the anomaly recognition unit recognizes whether equipment operation is abnormal or not by using a support vector machine model based on the unified feature vector; and the early warning decision-making unit generates a risk level according to the abnormal recognition result and outputs an early warning signal. According to the method, the feature fusion model based on the attention mechanism is constructed, and bidirectional weight distribution is performed on the vibration signal and the temperature signal on the basis of time alignment, so that the system can dynamically pay attention to modal features which are more discriminative for the current operation state.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault early warning, in particular to a vibration and temperature multi-parameter fusion early warning device for fuel service special equipment. BACKGROUND

[0002] In the field of operation and maintenance of fuel service special equipment (such as oil pumps, compressors, engines, etc.), the reliability and safety of the equipment are of great importance. The traditional state monitoring and fault early warning method mainly relies on independent monitoring and threshold alarm of a single physical parameter (such as vibration or temperature). However, this method has significant limitations: it is difficult to effectively identify slowly developing thermal-related faults caused by poor lubrication, cooling failure, etc. by analyzing only vibration signals, and the response to sudden mechanical faults such as bearing pitting and abnormal gear engagement is sluggish when monitoring temperature alone, leading to incomplete extraction of fault features and easy occurrence of false negatives or false positives. Although some systems attempt to collect vibration and temperature data simultaneously, they mostly use simple "and / or" logic rules for fusion judgment, failing to deeply explore the internal correlation and mutual influence of the two types of signals in terms of time dynamics and physical mechanisms. This static and shallow fusion method has poor adaptability under complex and variable working conditions and is not sensitive to early and weak compound fault signs. Most notably, the normal startup, loading, and variable load processes of the equipment itself can cause significant fluctuations in vibration and temperature, and the traditional method has difficulty in effectively distinguishing between these "benign" fluctuations caused by changes in working conditions and "malignant" abnormalities caused by potential faults, leading to frequent false alarms and severely reducing the credibility and practicality of the early warning system. In addition, since serious equipment failures are rare events, the fault samples available for training intelligent diagnostic models are extremely scarce, leading to insufficient generalization ability of models based on standard machine learning algorithms due to sample imbalance, and low recognition rate for rare faults. At the same time, the characteristic signals of early equipment failures are often very weak and easily masked by environmental noise and normal fluctuations, and the traditional method lacks effective means for feature enhancement and accurate capture. Therefore, a vibration and temperature multi-parameter fusion early warning device for fuel service special equipment is designed. SUMMARY

[0003] The present application aims to provide a vibration and temperature multi-parameter fusion early warning device for fuel service special equipment to solve the problems of single parameter monitoring limitations, shallow multi-parameter fusion, susceptibility to false alarms due to working condition interference, poor model generalization ability due to scarcity of fault samples, and difficulty in capturing early fault features in the traditional fuel service equipment monitoring method presented in the background.

[0004] To achieve the above-mentioned purpose, the present application aims to provide a vibration and temperature multi-parameter fusion early warning device for fuel service special equipment, comprising: A signal acquisition and processing unit acquires and processes vibration signals and temperature signals in real time during operation of the equipment; A feature extraction and fusion unit extracts vibration features and temperature change features based on the processed vibration signals and temperature signals, and forms a unified feature vector through an attention mechanism; An anomaly identification unit establishes and trains a support vector machine model based on the unified feature vector, enhances abnormal samples in the training through a negative sample generation strategy during training of the support vector machine model, and identifies whether there is an anomaly in the operation of the equipment using the trained support vector machine model; An early warning decision unit generates a risk level according to the anomaly identification result and outputs an early warning signal.

[0005] As a further improvement of the technical solution, the signal acquisition and processing unit comprises a signal acquisition module and a signal processing module; The signal acquisition module acquires vibration signals and temperature signals; The signal processing module pre-processes the acquired vibration signals and temperature signals.

[0006] As a further improvement of the technical solution, the feature extraction and fusion unit extracts vibration features and temperature change features based on the processed vibration signals and temperature signals, comprising the following steps: S1.1, based on the processed vibration signals, calculate time domain feature parameters and frequency domain feature parameters to obtain a vibration feature set representing the smoothness of the equipment operation and the mechanical impact situation; S1.2, based on the processed temperature signals, extract their trend features to form a temperature change feature set reflecting the thermal load state and cooling effect of the equipment; S1.3, align the vibration features and temperature change features in the time dimension; S1.4, fuse the aligned vibration features and temperature change features using an attention mechanism to generate a unified feature vector.

[0007] As a further improvement of the technical solution, in S1.2, based on the temperature signals, the trend features are extracted, comprising the following steps: S1.21, calculate the statistics of the temperature signals within a predetermined time window, associate the working condition state with the temperature signals, and construct a working condition-temperature change rate mapping table to determine whether the current temperature change is caused by a known working condition change; S1.22, first-order difference calculation is performed on the temperature signal to obtain the temperature change rate, the temperature rise rate and the temperature drop rate are extracted as dynamic characteristics, and the temperature change rate is corrected based on the working condition-temperature change rate mapping table; S1.23, second-order difference analysis is performed on the temperature signal, and the maximum temperature gradient is extracted; S1.24, a sliding window detection method is used to identify the stable interval of the temperature, and the duration and average value of the stable interval are calculated to reflect the heat dissipation and cooling effect of the equipment under steady-state operation; S1.25, the statistical quantity, the corrected temperature change rate, the maximum temperature gradient, and the duration and average value of the stable interval are combined to form a temperature change feature set.

[0008] As a further improvement of the technical solution, in S1.21, by associating the working condition state with the temperature signal and constructing a working condition-temperature change rate mapping table, the following steps are included: Receive the working condition state signal from the device control system, synchronize the working condition state signal with the temperature signal in time, and assign the corresponding working condition label to the temperature data stream; Construct a working condition-temperature change rate mapping table to record the normal range of temperature change rate under different working conditions.

[0009] As a further improvement of the technical solution, in S1.22, the change rate is corrected based on the working condition-temperature change rate mapping table, specifically: combining the working condition label at the current time point, querying the expected normal change rate reference value under the corresponding working condition in the working condition-temperature change rate mapping table, comparing and correcting the deviation between the real-time extracted dynamic characteristics and the expected normal change rate reference value, to obtain the corrected temperature change rate.

[0010] As a further improvement of the technical solution, in S1.4, the aligned vibration features and temperature features are fused using an attention mechanism to generate a unified feature vector, including the following steps: S1.41, input the vibration features and temperature features into the attention mechanism module; S1.42, respectively, small feedforward neural network mapping is performed on the vibration features and temperature features, and the vibration features and temperature features are projected into the same feature space; S1.43, calculate the attention weight of the vibration features to the temperature features and the attention weight of the temperature features to the vibration features; S1.44, use the attention weight to weight the sum of the corresponding feature value vectors to obtain the fused context representation and form the fused feature vector; S1.45, the fused feature vector is compressed in dimension and nonlinearly mapped through a fully connected layer to generate a unified feature vector.

[0011] As a further improvement of the technical solution, the abnormality identification unit uses the trained support vector machine model to identify whether the device operation is abnormal, comprising the following steps: S2.1, normalizing the unified feature vector; S2.2, collecting multiple sets of operation data of the device in the historical health state as a normal sample set, collecting operation data of the device in the known failure stage at the same time and marking it as an abnormal sample set, and using a negative sample generation strategy to enhance the abnormal sample set to generate an enhanced abnormal sample set; S2.3, based on the normal sample set and the enhanced abnormal sample set, using a weighted support vector machine algorithm to train the support vector machine model; S2.4, inputting the normalized unified feature vector into the trained support vector machine model, and using a decision function to calculate the category of the unified feature vector; S2.5, outputting the classification label corresponding to the unified feature vector, if it is determined to be normal, it means that the device operation state has no abnormal trend, if it is determined to be abnormal, it means that there is a risk of failure.

[0012] As a further improvement of the technical solution, in S2.2, the negative sample generation strategy is used to enhance the abnormal sample set, comprising the following steps: S2.21, performing unsupervised clustering analysis on the normal sample set, and dividing the operation data in the health state into K different health clusters; S2.22, calculating the centroid and boundary of each health cluster; S2.23, for each health cluster, generating a synthetic data point at the edge of its data space through a clustering centroid-based interpolation algorithm; S2.24, marking the synthetic data point as abnormal to form a synthetic abnormal sample set; S2.25, combining the original abnormal sample set and the synthetic abnormal sample set to form an enhanced abnormal sample set.

[0013] As a further improvement of the technical solution, the warning decision unit receives the classification label output by the abnormality identification unit, calculates the abnormality degree index using the distance of the unified feature vector to the support vector machine classification hyperplane, compares the abnormality degree index with the preset threshold interval, judges the risk level, generates the corresponding warning signal according to the risk level, and outputs the warning signal to the device control system.

[0014] Compared with the prior art, the present application has the following advantages: 1. In the fuel service special equipment vibration and temperature multi-parameter fusion early warning device, a feature fusion model based on an attention mechanism is constructed, vibration signals and temperature signals are time-aligned, and bidirectional weight distribution is performed, so that the system can dynamically pay attention to the modal features (such as abnormal vibration or abnormal temperature rise) that are more discriminative for the current operating state, and effectively overcome the problems of fixed weight and poor adaptability in traditional weighted fusion methods. At the same time, by introducing a working condition-temperature change rate mapping table to correct the temperature change rate, the temperature fluctuations caused by normal working condition switching such as load and speed are eliminated, and the fault sensitivity of the temperature feature is significantly improved. Combined with the complementary information of vibration and temperature, the precise perception of potential faults such as mechanical wear, friction aggravation and cooling failure is realized, and the stability and reliability of the early warning system under complex variable working condition conditions are enhanced.

[0015] 2. In the fuel service special equipment vibration and temperature multi-parameter fusion early warning device, in view of the problem that fault data is rare in the actual operation of fuel service equipment and it is difficult to train an effective classification model, the device adopts a negative sample generation + weighted SVM strategy for sample enhancement: representative synthetic abnormal samples are generated on the periphery of each healthy cluster boundary through cluster analysis of healthy data, simulating the evolution path of the equipment from normal to abnormal state, and effectively expanding the abnormal sample set. Combined with the weighted support vector machine, a higher penalty coefficient is given to the rare abnormal class, and the discrimination ability of the classification hyperplane in the healthy and abnormal boundary area is strengthened. This method can build a high-sensitivity abnormal recognition model without relying on a large amount of real fault data, significantly improving the system's ability to capture early weak fault signs and meeting the actual needs of industrial equipment predictive maintenance. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The overall flowchart of the present application is shown in the figure; The meanings of the various numbers in the figure are as follows: 1. Signal acquisition and processing unit; 2. Feature extraction and fusion unit; 3. Abnormal recognition unit; 4. Early warning decision unit. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0018] Embodiment: Please refer to Figure 1As shown, a fuel service special equipment vibration and temperature multi-parameter fusion early warning device is provided, comprising a signal acquisition and processing unit 1, which acquires and processes vibration signals and temperature signals in real time during equipment operation; In this embodiment, the signal acquisition and processing unit 1 comprises a signal acquisition module and a signal processing module; The signal acquisition module acquires vibration signals and temperature signals, wherein the vibration sensors are arranged at positions such as bearings and pump bodies that are prone to mechanical impact and wear, for acquiring vibration acceleration or speed signals; the temperature sensors are arranged at heat-sensitive positions such as pipelines and machine body surfaces, for acquiring surface temperatures and temperature changes during equipment operation, so as to ensure that the obtained data cover the main stress and heat receiving areas of the equipment; The signal processing module pre-processes the acquired vibration signals and temperature signals, including filtering, denoising, normalization and time synchronization processing, to eliminate the influence of environmental interference and sensor noise on signal quality, and to ensure the consistency of vibration and temperature signals in time scale and amplitude scale through data format conversion and feature alignment, so as to provide clean and comparable input data for subsequent feature extraction and fusion.

[0019] The feature extraction and fusion unit 2 extracts vibration features and temperature change features based on the processed vibration signals and temperature signals, and forms a unified feature vector through an attention mechanism, including the following steps: S1.1, based on the processed vibration signals, calculate time domain feature parameters (such as root mean square value, kurtosis, skewness, peak factor) and frequency domain feature parameters (such as frequency band energy, main frequency component, envelope spectrum feature), to obtain a vibration feature set representing the smoothness of equipment operation and mechanical impact, specifically: in the time domain, calculate the root mean square value of the vibration signal point by point to represent the overall energy level, calculate the kurtosis to reflect the impact component, calculate the skewness to measure the symmetry of the signal distribution, and calculate the peak factor to reveal the pulse component and extreme value feature; further convert the vibration signal to the frequency domain through fast Fourier transform, analyze the power spectral density, extract the energy distribution in each frequency band, identify the main frequency component corresponding to the maximum amplitude, and extract the envelope spectrum feature through envelope demodulation technology to capture the impact signal caused by bearing failure, gear meshing abnormality, etc., so as to obtain a vibration feature parameter set that can fully represent the equipment operation state; S1.2, based on the processed temperature signal, extract its change trend feature to form a temperature change feature set reflecting the thermal load state and cooling effect of the equipment, including the following steps: S1.21, within a preset time window, calculate the statistics of the temperature signal, including the average temperature, maximum value, minimum value and variance, to characterize the overall thermal load level of the device, by associating the working condition state with the temperature signal and constructing a working condition-temperature change rate mapping table, to determine whether the current temperature change is caused by a known working condition change; Further, the temperature change of fuel service equipment (such as an aero-engine, fuel pump, etc.) is not always caused by a fault, and its working temperature is strongly dependent on the current working condition (such as speed, load, throttle state). For example, when the engine switches from idle to maximum thrust, the normal rising rate of temperature will be very fast, and if only one fixed threshold is used for judgment, this normal and intense working condition change will be misreported as an overheating anomaly. The prior art cannot distinguish between temperature fluctuations caused by normal working condition changes and abnormal temperature rises caused by potential faults, resulting in a very high false alarm rate of the early warning system, which cannot be used; the present invention introduces a working condition-aware dynamic correction mechanism (i.e. a working condition-temperature change rate mapping table), which does not statically judge the temperature level or change rate, but first identifies the current running state (working condition) of the device, then queries the expected normal change range under this state as a reference, and finally compares the real-time data with this dynamic and working condition-compliant reference. This greatly reduces false alarms caused by normal operation, enabling the system to more accurately capture abnormal thermal phenomena that deviate from the expected behavior, significantly improving the accuracy and reliability of the early warning; By associating the working condition state with the temperature signal and constructing a working condition-temperature change rate mapping table, including the following steps: Receive working condition state signals (including speed setpoint, power gear, load current) from the device control system, time-synchronize the working condition state signals with the temperature signals, and assign corresponding working condition labels (including steady state, acceleration, high load, deceleration) to the temperature data stream; Construct a working condition-temperature change rate mapping table to record the normal range of temperature change rate (including temperature rise rate) under different working conditions (or working condition switching) (for example, from idle to cruising power, the normal temperature rise rate range is 0.5-1.0°C / s); The working condition-temperature change rate mapping table is a multi-dimensional lookup table constructed based on historical health operation data, and is used to store the normal temperature change behavior benchmark values of the device under different working conditions. The construction is as follows: the working condition state vector is used as an index, the vector is composed of working condition parameters such as speed, load rate, flow setting value, valve opening, environment temperature and working mode code, and is used to uniquely represent the device operation state; the temperature change characteristic benchmark value is used as the storage content, the benchmark value includes the expected temperature rise rate range, the expected temperature drop rate range, the steady state temperature expectation value and the typical time to reach the steady state, and is used to reflect the temperature evolution law of the device in the healthy state; and the metadata such as data source, sample number and last update time are attached, which is used to ensure the reliability and updateability of the mapping table. During operation, the system can query the corresponding temperature benchmark characteristics in the mapping table based on the real-time collected working condition vector, compare with the real-time temperature change characteristics, so as to realize the correction and discrimination of abnormal temperature behavior, and can dynamically update the mapping table through the online learning mechanism during the long-term health operation of the device, so as to adapt to the aging of the device or the drift of the sensor, and realize the continuous tracking and optimization of the working condition-temperature characteristics; S1.22, first-order difference or slope calculation is performed on the temperature signal to obtain the temperature change rate (i.e. the rate of change of temperature with time, specifically: in a preset time window, a sequence of discrete sampling points of the temperature signal is obtained, for two adjacent sampling points, the temperature difference value is calculated and divided by the corresponding time interval to obtain the temperature change rate of the interval; the whole sequence is calculated in turn to form a temperature change rate sequence), the temperature rise rate and the temperature drop rate are extracted as dynamic characteristics, and the temperature change rate is corrected based on the working condition-temperature change rate mapping table, this step eliminates the temperature change caused by normal working conditions, so that the corrected change rate more directly reflects the abnormal temperature change caused by faults (such as increased friction and cooling failure); Further, the change rate is corrected based on the working condition-temperature change rate mapping table, specifically: the working condition label of the current time point is combined to query the expected normal change rate benchmark value under the corresponding working condition in the working condition-temperature change rate mapping table, the real-time extracted dynamic characteristics are compared with the expected normal change rate benchmark value and deviation correction (if the real-time change rate is within the benchmark range (for example: under a certain load and speed, the benchmark temperature rise rate range is 0.5-1.0°C / s), it is considered as normal fluctuation; if it deviates from the benchmark range, the deviation between it and the benchmark mean value is calculated, and the corrected change rate is obtained through correction factor or difference compensation, so as to eliminate the normal temperature fluctuation caused by different working conditions), to obtain the corrected temperature change rate; S1.23, second-order difference or curvature analysis is performed on the temperature signal to extract the maximum temperature gradient for identifying rapid temperature rise or cooling anomalies (this step is calculated based on the corrected temperature change rate sequence), specifically: on the basis of the pre-processed temperature sequence of temperature change rate obtained by first-order difference, the difference or slope change of temperature change rate of adjacent sampling points is calculated, that is, the second-order difference value, to form the temperature curvature sequence; by analyzing the curvature sequence, local extreme points are identified, that is, points of significant change in temperature change rate, which are used to identify the time when the temperature change direction changes significantly, and the temperature change rate corresponding to the point with the maximum absolute curvature value is extracted as the maximum temperature gradient; S1.24, a sliding window detection method is used to identify the stable interval of the temperature, and the duration and average value of the stable interval are calculated (specifically: on the pre-processed temperature signal sequence, a fixed length sliding window is set, and the window is moved along the time axis; in each window, the variance or change amplitude of the temperature signal is calculated, and when the variance or change amplitude is lower than the pre-set threshold, the window is determined as a temperature stable section; complete stable intervals are formed by merging the windows that continuously meet the stable condition, and the start and end times of the intervals are recorded to calculate the duration of the stable interval, and the average temperature value in the interval is calculated to obtain the average temperature feature), to reflect the cooling effect of the device under stable operation; S1.25, the statistical quantity, the corrected temperature change rate, the maximum temperature gradient, and the duration and average value of the stable interval are combined to form a temperature change feature set; S1.3, align the vibration features and temperature change features in the time dimension, and ensure the consistency of the two in the time scale by linear interpolation method (first determine the sampling frequency and time stamp of the vibration signal and the temperature signal, and select the signal with higher frequency as the reference time axis; then linearly interpolate the signal with low frequency or non-uniform sampling on the reference time axis, and estimate the signal value corresponding to the reference time point according to the known values of adjacent sampling points; in this process, the time axis is corrected synchronously to ensure that the time of all feature points is consistent, thereby generating the vibration feature sequence and the temperature feature sequence after time alignment, and realizing the unification of the two in the time scale); S1.4, the aligned vibration features and temperature change features are fused by using attention mechanism to generate a unified feature vector; Further, the aligned vibration features and temperature features are fused by using attention mechanism to generate a unified feature vector, including the following steps: S1.41, the vibration features and temperature features are input into the attention mechanism module to ensure that the sampling time points of the two types of features are consistent for feature interaction at each time; S1.42, respectively, a small feedforward neural network mapping is performed on the vibration feature and the temperature feature, and the vibration feature and the temperature feature are projected into the same feature space, wherein the small feedforward neural network mapping input is a vibration feature vector or a temperature feature vector, which first passes through one to two fully connected hidden layers, the number of nodes in each layer can be appropriately set according to the dimension of the feature, for example, the vibration feature is projected into the same dimension as the temperature feature, a nonlinear activation function (such as ReLU) is connected after each hidden layer to enhance the nonlinear expression ability, and a batch normalization or Dropout layer can be added to improve the training stability and prevent overfitting; the hidden layer output is mapped to the target feature space through a fully connected layer, and a projection vector with the same dimension as the other type of feature is obtained as the output, which is used for subsequent attention mechanism calculation, so as to realize efficient mapping and information interaction of the vibration feature and the temperature feature in the unified feature space; S1.43, the attention weight of the vibration feature to the temperature feature and the attention weight of the temperature feature to the vibration feature are calculated; Wherein, the attention weight of the vibration feature to the temperature feature is: ; The attention weight of the temperature feature to the vibration feature is: ; In the formula, is the query vector of the temperature feature, is the key vector of the vibration feature, is the value vector of the vibration feature, is the query vector of the vibration feature, is the key vector of the temperature feature, is the value vector of the temperature feature, is a normalization function, is a vector dimension, used for scaling, to keep the balance of attention distribution, represents the transposition operation; S1.44, the corresponding feature value vectors are weighted and summed using the attention weights to obtain the fused context representation, forming a fused feature vector; Wherein, the fused context representation is respectively , ; The fused feature vector is: ; In the formula, is the context representation of the vibration feature under the guidance of the temperature feature, a context representation of the temperature feature under guidance of the vibration feature, for a stitching operation; S1.45, the fused feature vector is subjected to dimension compression and nonlinear mapping through a fully connected layer to generate a unified feature vector, specifically: the fused feature vector is input into the fully connected layer, the number of nodes of the fully connected layer is set to the dimension of the target unified feature vector, the input vector is mapped to a low-dimensional feature space through matrix multiplication, and a nonlinear activation function (ReLU function) is added to enhance the feature expression capability; batch normalization or Dropout can be selected during the mapping process to improve network stability and prevent overfitting; after the fully connected layer processing, the output vector is the unified feature vector, which not only compresses the dimension, but also fuses the interaction information of the vibration feature and the temperature feature, providing clean and efficient input features for subsequent anomaly identification.

[0020] The anomaly identification unit 3 establishes and trains a support vector machine model based on the unified feature vector, and in the training process of the support vector machine model, the abnormal samples in the training are enhanced through a negative sample generation strategy, and the trained support vector machine model is used to identify whether the device is running abnormally; In this embodiment, the anomaly identification unit 3 uses the trained support vector machine model to identify whether the device is running abnormally, including the following steps: S2.1, the unified feature vector is subjected to normalization processing, and each feature value is linearly mapped to a preset range (such as 0-1) or subjected to standardization (subtracting the mean and then dividing by the standard deviation) to eliminate the dimensional and amplitude differences between different feature dimensions; S2.2, a plurality of sets of running data of the device in the historical health state are collected as a normal sample set, and running data of the device in the known fault stage are collected and marked as an abnormal sample set (real and rare fault data), and the negative sample generation strategy is used to enhance the abnormal sample set to generate an enhanced abnormal sample set, wherein, to solve the problem of insufficient abnormal samples, the negative sample generation strategy is introduced to expand the negative class samples (the negative sample generation strategy is used to generate artificial abnormal samples (i.e. negative samples) at the data level, specifically: clustering the normal samples, generating synthetic abnormal points outside the boundary of each healthy cluster, obtaining an enhanced and more representative abnormal sample set (original abnormal + synthetic abnormal)), and a weighted support vector machine algorithm is used to improve the sensitivity of the model to rare abnormalities (the weighted support vector machine algorithm is based on the enhanced abnormal sample set and the original normal sample set, and different weights are assigned to the normal and abnormal classes (the abnormal weight is greater than the normal weight) when training the SVM model); considering that the device fault is essentially a deviation of the running parameters from the health state space, the generated synthetic abnormal samples aim to construct the decision boundary between normal and abnormal, and focus on depicting the peripheral contour of the health state rather than reproducing the detailed features of a specific fault type; The negative sample generation strategy is mainly aimed at the sample imbalance problem caused by the lack of abnormal samples and the difficulty in obtaining fault data in the actual operation of fuel service special equipment. In the equipment maintenance scene, the number of normal operation samples is much larger than that of fault samples. The traditional support vector machine model will produce classification bias due to the lack of abnormal samples, resulting in reduced sensitivity to early faults and increased false positive rate. Compared with the traditional simple oversampling or artificially synthesized abnormal data method, the strategy generates synthesized abnormal samples close to the true decision boundary through unsupervised clustering analysis of the multi-modal distribution characteristics of healthy data and interpolation based on cluster centroids, not only enhancing the sample diversity, but also significantly improving the classification boundary accuracy of the support vector machine under the condition of sparse abnormal samples. This intelligent sample enhancement strategy based on data distribution characteristics can more accurately identify the early abnormal trend of the equipment, reduce the false negative rate, and at the same time maintain the discrimination accuracy of the normal operation state; The negative sample generation strategy is mainly aimed at the sample imbalance problem caused by the lack of abnormal samples and the difficulty in obtaining fault data in the actual operation of fuel service special equipment. In the equipment maintenance scene, the number of normal operation samples is much larger than that of fault samples. The traditional support vector machine model will produce classification bias due to the lack of abnormal samples, resulting in reduced sensitivity to early faults and increased false positive rate. Compared with the traditional simple oversampling or artificially synthesized abnormal data method, the strategy generates synthesized abnormal samples close to the true decision boundary through unsupervised clustering analysis of the multi-modal distribution characteristics of healthy data and interpolation based on cluster centroids, not only enhancing the sample diversity, but also significantly improving the classification boundary accuracy of the support vector machine under the condition of sparse abnormal samples. This intelligent sample enhancement strategy based on data distribution characteristics can more accurately identify the early abnormal trend of the equipment, reduce the false negative rate, and at the same time maintain the discrimination accuracy of the normal operation state; S2.21, unsupervised clustering analysis (K-Means) is performed on the normal sample set, and the running data in the healthy state are divided into K different healthy clusters; these clusters represent different working conditions, load operation modes (for example: high load steady operation, low load operation, start-stop transition period, etc.) of the equipment in the healthy state, and the number of clusters K is determined by the silhouette coefficient; S2.22, calculate the centroid and boundary of each healthy cluster, after clustering analysis of the normal sample set, calculate the mean of each dimension feature value as the centroid of the cluster, which represents the center position of the normal samples in the cluster; at the same time, according to the distance distribution of the samples in the cluster to the centroid, the boundary of the cluster is determined, for example, the maximum distance or a certain percentage (such as 95% percentile) can be selected as the boundary radius to describe the coverage of the cluster; S2.23, for each healthy cluster, a series of synthetic data points are generated at the edge of its data space (i.e. away from the global normal data centroid) through interpolation algorithm based on cluster centroid, the generation process is: selecting the centroid vector of a certain healthy cluster and the reference point or abnormal direction vector outside the boundary of the cluster as the interpolation end points; then introducing an interpolation coefficient ( ) between them, obtaining a new synthetic data point between the centroid and the point outside the boundary through linear interpolation formula , in the formula, is the centroid of the cluster, is the point outside the boundary; by selecting different boundary reference points and interpolation coefficients multiple times, synthetic data points distributed outside the healthy cluster and covering different degrees of abnormal deviation can be generated, thereby expanding the abnormal sample set and improving the discrimination ability of the classification model in the boundary region; S2.24, label the synthetic data points as abnormal, forming a synthetic abnormal sample set; S2.25, combine the original abnormal sample set and the synthetic abnormal sample set (i.e., put each record (data row) in the original abnormal sample set and each record in the generated synthetic abnormal sample set into the same training set), forming an enhanced abnormal sample set; S2.3, based on the normal sample set and the enhanced abnormal sample set, train a support vector machine (SVM) model using a weighted support vector machine algorithm to determine the optimal classification hyperplane and support vectors, forming a classification model that can distinguish between normal and abnormal operating states, wherein the input is the normalized unified feature vector, the SVM model maps the input features to a high-dimensional feature space through a kernel function (linear kernel) to enhance the separability of the data; in the high-dimensional space, the model determines the optimal classification hyperplane through training, which is composed of support vectors, weight coefficients, and bias terms, and is used to maximize the separation between the two classes of samples (normal and abnormal); the output is the predicted label of the class to which the input feature vector belongs, or the distance from the hyperplane calculated by the decision function, which gives a continuous abnormality index for subsequent risk level determination and early warning signal generation; the SVM model itself does not have a traditional hidden layer, its ability mainly depends on kernel mapping and support vector selection, and it realizes direct mapping and classification from input to output; S2.4, input the normalized unified feature vector into the trained support vector machine model, and use the decision function to calculate the class to which the unified feature vector belongs, according to the distance from the support vector machine classification hyperplane, to determine which class (normal or abnormal) it belongs to; S2.5, output the classification label corresponding to the unified feature vector, if it is determined to be normal, it indicates that the device operating state has no abnormal trend, if it is determined to be abnormal, it indicates that there is a potential risk of failure.

[0021] The early warning decision unit 4 generates a risk level according to the abnormality recognition result and outputs a warning signal; In the embodiment, the early warning decision unit 4 receives the classification label output by the anomaly identification unit 3, calculates an anomaly degree index using the distance (the absolute value of the distance reflects the degree of deviation of the feature vector from the center of the health cluster) of the unified feature vector to the support vector machine classification hyperplane, and uses the anomaly degree index to characterize the magnitude of the deviation of the device from the normal operating state. The anomaly degree index is compared with a preset threshold interval to determine the risk level (the risk level includes normal, slight anomaly, general anomaly, and serious anomaly. Below the conventional threshold a is normal, between the conventional threshold a and the slight threshold d is slight anomaly, between the slight threshold d and the general threshold b is general anomaly, and above the serious threshold c is serious anomaly). According to the risk level, a corresponding early warning signal (the early warning signal includes a low-risk prompt, a medium-risk warning, and a high-risk alarm. The slight anomaly triggers the low-risk prompt, which is manifested as a system interface prompt or a slight sound and light alarm. The general anomaly triggers the medium-risk warning, which is accompanied by a continuous sound and light warning or an SMS / email notification. The serious anomaly triggers the high-risk alarm, which triggers an emergency shutdown instruction, a forced sound and light alarm, and the sending of high-priority alarm information to the control system) is generated and output to the device control system.

[0022] The basic principles, main features and advantages of the present application are shown and described above. It should be understood by those skilled in the art that the present application is not limited by the above embodiments, and the above embodiments and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the present application.

Claims

1. A vibration and temperature multi-parameter fusion early warning device for fuel service special equipment, characterized in that, The application relates to a device operation state abnormality identification method and device. The device operation state abnormality identification method comprises the following steps: A signal acquisition and processing unit (1) acquires and processes vibration signals and temperature signals in real time during device operation; A feature extraction and fusion unit (2) extracts vibration features and temperature change features based on the processed vibration signals and temperature signals, and forms a unified feature vector through an attention mechanism; An abnormality identification unit (3) establishes and trains a support vector machine model based on the unified feature vector, enhances abnormal samples in the training process through a negative sample generation strategy, and identifies whether device operation is abnormal by using the trained support vector machine model; 2. The vibration and temperature multi-parameter fusion early warning device for fuel service special equipment according to claim 1, characterized in that: An early warning decision unit (4) generates a risk level according to the abnormality identification result and outputs an early warning signal. The signal acquisition and processing unit (1) comprises a signal acquisition module and a signal processing module; The signal acquisition module acquires vibration signals and temperature signals; 3. The vibration and temperature multi-parameter fusion early warning device for fuel service special equipment according to claim 1, characterized in that: The signal processing module pre-processes the acquired vibration signals and temperature signals. The feature extraction and fusion unit (2) extracts vibration features and temperature change features based on the processed vibration signals and temperature signals, and comprises the following steps: S1.

1. Based on the processed vibration signals, time domain feature parameters and frequency domain feature parameters are calculated to obtain a vibration feature set representing device operation stability and mechanical impact; S1.

2. Based on the processed temperature signals, a change trend feature is extracted to form a temperature change feature set reflecting device thermal load state and cooling effect; S1.

3. The vibration features and the temperature change features are aligned in the time dimension; 4. The vibration and temperature multi-parameter fusion early warning device for fuel service special equipment according to claim 3, characterized in that: S1.

4. The aligned vibration features and the temperature change features are fused by using an attention mechanism to generate a unified feature vector. In S1.2, based on the temperature signals, the change trend feature is extracted, and comprises the following steps: S1.

21. In a preset time window, the statistical quantity of the temperature signals is calculated, the working condition state is associated with the temperature signals, and a working condition-temperature change rate mapping table is constructed to determine whether the current temperature change is caused by a known working condition change; S1.

22. The temperature signals are subjected to first-order difference calculation to obtain temperature change rates, the temperature rise rate and the temperature drop rate are extracted as dynamic features, and the temperature change rates are corrected based on the working condition-temperature change rate mapping table; S1.

23. The temperature signals are subjected to second-order difference analysis to extract the maximum temperature gradient; S1.

24. A sliding window detection method is used to identify the stable interval of the temperature signals, and the duration and average value of the stable interval are calculated to reflect the heat dissipation and cooling effect of the device under stable operation; 5. The vibration and temperature multi-parameter fusion early warning device for fuel service special equipment according to claim 4, characterized in that: S1.

25. The statistical quantity, the corrected temperature change rate, the maximum temperature gradient and the duration and average value of the stable interval are combined to form a temperature change feature set. In S1.21, the working condition state is associated with the temperature signals, and a working condition-temperature change rate mapping table is constructed, comprising the following steps: Receiving a working condition state signal from the equipment control system, time-synchronizing the working condition state signal with the temperature signal, and assigning a corresponding working condition label to the temperature data stream; Constructing a working condition-temperature change rate mapping table for recording the normal range of temperature change rate under different working conditions.

6. The vibration and temperature multi-parameter fusion early warning device for fuel service special equipment according to claim 4, characterized in that: In S1.22, the change rate is corrected based on the working condition-temperature change rate mapping table, specifically: combining the working condition label at the current time point, querying the expected normal change rate reference value under the corresponding working condition in the working condition-temperature change rate mapping table, comparing and correcting the deviation between the real-time extracted dynamic feature and the expected normal change rate reference value, and obtaining the corrected temperature change rate.

7. The vibration and temperature multi-parameter fusion early warning device for fuel service special equipment according to claim 3, characterized in that: In S1.4, the aligned vibration feature and temperature feature are fused using an attention mechanism to generate a unified feature vector, including the following steps: S1.41, inputting the vibration feature and the temperature feature into an attention mechanism module; S1.42, performing small feedforward neural network mapping on the vibration feature and the temperature feature respectively, and projecting the vibration feature and the temperature feature into the same feature space; S1.43, calculating the attention weight of the vibration feature to the temperature feature and the attention weight of the temperature feature to the vibration feature; S1.44, using the attention weight to weight the corresponding feature value vector, obtaining the fused context representation, and forming the fused feature vector; S1.45, performing dimension compression and nonlinear mapping on the fused feature vector through a fully connected layer to generate a unified feature vector.

8. The vibration and temperature multi-parameter fusion early warning device for fuel service special equipment according to claim 1, characterized in that: The anomaly recognition unit (3) uses the trained support vector machine model to identify whether the equipment operation is abnormal, including the following steps: S2.1, normalizing the unified feature vector; S2.2, collecting multiple sets of running data of the equipment under historical health states as normal sample sets, collecting running data of the equipment in the known fault stage as abnormal sample sets, and using a negative sample generation strategy to enhance the abnormal sample sets to generate enhanced abnormal sample sets; S2.3, based on the normal sample sets and the enhanced abnormal sample sets, using a weighted support vector machine algorithm to train the support vector machine model; S2.4, inputting the normalized unified feature vector into the trained support vector machine model, and using a decision function to calculate the category to which the unified feature vector belongs; S2.5, outputting the classification label corresponding to the unified feature vector, if it is determined to be normal, it means that the equipment operation state has no abnormal trend, if it is determined to be abnormal, it means that there is a risk of failure.

9. The vibration and temperature multi-parameter fusion early warning device for fuel service special equipment according to claim 8, characterized in that: In S2.2, the negative sample generation strategy is used to enhance the abnormal sample set, including the following steps: S2.21, performing unsupervised clustering analysis on the normal sample set, and dividing the running data under the health state into K different health clusters; S2.22, calculating the centroid and boundary of each health cluster; S2.23, for each health cluster, generating synthetic data points on the edge of its data space through a clustering centroid-based interpolation algorithm; S2.24, marking the synthetic data points as abnormal to form a synthetic abnormal sample set; S2.25, combine the original abnormal sample set with the synthetic abnormal sample set to form an enhanced abnormal sample set.

10. The vibration and temperature multi-parameter fusion early warning device for fuel service special equipment according to claim 9, characterized in that: The early warning decision unit (4) receives the classification label output by the anomaly identification unit (3), calculates an anomaly degree index using the distance from the unified feature vector to the support vector machine classification hyperplane, compares the anomaly degree index with a preset threshold interval, makes a risk level determination, generates a corresponding early warning signal according to the risk level, and outputs the early warning signal to the equipment control system.

Images