Detection method and detection sensor for temperature vibration data of dynamic equipment

Abstract

The invention relates to the technical field of mechanical equipment state monitoring, in particular to a method and sensor for detecting temperature and vibration data of dynamic equipment, and the method comprises the steps: collecting a temperature signal and a vibration signal of the dynamic equipment, and carrying out the fusion processing, thereby obtaining a fusion feature set; establishing a feature distribution baseline based on a Gaussian mixture model, calculating a relative entropy of the fusion feature set and the feature distribution baseline, and generating a dynamic threshold sequence; constructing a detection model according to the fusion feature set and the dynamic threshold sequence, and outputting to obtain an abnormal mode label; performing time serialization processing on the abnormal mode label, predicting a fault probability in a future time period according to the abnormal mode label subjected to time serialization, and generating a fault prediction result; and performing priority ranking on the fault types in the fault prediction result, and generating a maintenance report according to a priority ranking result. The reliability and practicability of temperature vibration data detection of the dynamic equipment are comprehensively improved, and an efficient solution is provided for health management of the dynamic equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical equipment condition monitoring, and in particular to a method and a detection sensor for detecting temperature and vibration data of a moving equipment. Background Art

[0002] Dynamic equipment refers to equipment that performs mechanical motion or energy conversion during operation, typically involving rotation, reciprocating, or vibration. Dynamic equipment uses moving parts (such as rotors, bearings, gears, and pistons) to achieve energy transfer or mechanical operation. Examples include pumps, fans, compressors, motors, turbines, and internal combustion engines. These devices generate temperature changes and vibration signals during operation, which are important indicators of their health.

[0003] In existing technologies, high-precision sensors are typically used to collect temperature and vibration data from moving equipment, and this data is analyzed using signal processing techniques. For example, vibration signals are often analyzed using fast Fourier transforms (FFTs) or time-domain analysis to extract features such as amplitude and frequency, while temperature signals are characterized by calculating statistical properties such as mean and standard deviation. These features are then fed into classification algorithms (such as decision trees or naive Bayesian classifiers) to determine whether the equipment is experiencing an anomaly.

[0004] Despite this, existing technologies generally have the technical defect of being unable to fully utilize the dynamic correlation between the two signals when processing temperature and vibration data. Existing methods usually process temperature and vibration signals separately as independent signal sources, lacking a comprehensive analysis of the mutual influence between the two during equipment operation, which in turn affects the accuracy of fault prediction and the effectiveness of maintenance decisions. Summary of the Invention

[0005] The main purpose of the present invention is to provide a method and a detection sensor for detecting temperature and vibration data of a moving device, aiming to overcome the defect in the prior art that the dynamic correlation between temperature and vibration signals cannot be fully utilized.

[0006] In order to achieve the above-mentioned problem, the present invention proposes a method for detecting temperature and vibration data of a moving device, the method comprising: collecting temperature signals and vibration signals of the moving equipment, and synchronously processing the temperature signals and vibration signals to obtain a multi-dimensional data matrix; extracting a first change feature of the vibration signal and a second change feature of the temperature signal from the multidimensional data matrix, and fusing the first change feature and the second change feature to obtain a fused feature set; Establishing a feature distribution baseline based on a Gaussian mixture model, calculating the relative entropy between the fused feature set and the feature distribution baseline, and generating a dynamic threshold sequence; Building a detection model based on the fusion feature set and the dynamic threshold sequence, calculating the joint probability of temperature anomaly and vibration anomaly through the detection model, and outputting an abnormal pattern label; Performing time serialization processing on the abnormal pattern labels, predicting the failure probability in a future time period based on the time serialized abnormal pattern labels, and generating a failure prediction result; The fault types in the fault prediction results are prioritized, and a maintenance report is generated according to the priority ranking results.

[0007] Furthermore, the step of collecting the temperature signal and the vibration signal of the moving device and synchronously processing the temperature signal and the vibration signal to obtain a multidimensional data matrix includes: Based on multi-channel data acquisition, the temperature signal and vibration signal of the moving equipment are collected, and the acquisition timestamp of each signal is recorded at the same time to generate the original signal set containing the temperature and vibration signal amplitude; According to the timestamp information in the original signal set, the temperature and vibration signals are time-aligned based on a linear interpolation algorithm to form a time-synchronized aligned signal set; According to the amplitude range of the alignment signal set, the signal values ​​in the alignment signal set are mapped to a fixed interval for normalization to obtain a normalized signal set; The time series, amplitude and phase information of the standardized signal set are subjected to multi-dimensional reconstruction processing to obtain a multi-dimensional data matrix.

[0008] Furthermore, the step of extracting a first change feature of the vibration signal and a second change feature of the temperature signal from the multidimensional data matrix, and fusing the first change feature and the second change feature to obtain a fused feature set includes: Decomposing the vibration signal in the multidimensional data matrix into a plurality of frequency band sub-signals to generate a first variation feature including a time-frequency energy feature; Analyzing the temperature signal in the multidimensional data matrix based on the time domain information of the time-frequency energy feature to generate a second change feature including a gradient statistical feature vector; Discretize the first and second change features based on Shannon entropy to generate a preferred feature subset; Constructing a graphical model based on the preferred feature subset, setting each feature dimension in the preferred feature subset as a node in the graphical model, and obtaining a temperature-vibration coupling feature; An operation change state of the moving device is obtained, and weighted processing is performed on the temperature-vibration coupling feature based on the operation change state to generate a fusion feature set.

[0009] Furthermore, the step of analyzing the temperature signal in the multidimensional data matrix based on the time domain information of the time-frequency energy feature to generate a second change feature including a gradient statistical feature vector further includes: Constructing a sliding window with a preset number of bits, traversing the temperature signal in the multidimensional data matrix with the sliding window, and calculating a gradient sequence of temperature values ​​in each window; Extracting first-order statistics and second-order statistics of the gradient sequence, and adjusting the step size of the window in combination with the time domain information of the time-frequency energy feature to generate statistics; The statistics corresponding to multiple windows are concatenated to generate a gradient statistical feature vector.

[0010] Furthermore, the step of discretizing the first change feature and the second change feature based on Shannon entropy to generate a preferred feature subset further includes: Discretize each feature dimension of the first change feature and the second change feature to generate a probability distribution of the feature; Calculating the entropy value of each feature dimension using the Shannon entropy formula according to the probability distribution to obtain an entropy value set of the first change feature and the second change feature; The information gain of the first change feature and the second change feature is calculated based on the entropy value set, an information gain ranking is generated, and feature dimensions below a preset information gain threshold are removed from the first change feature and the second change feature to form the preferred feature subset.

[0011] Furthermore, the steps of establishing a feature distribution baseline based on a Gaussian mixture model, calculating the relative entropy between the fused feature set and the feature distribution baseline, and generating a dynamic threshold sequence include: Acquire historical temperature signals and vibration signals of the moving equipment, perform Gaussian mixture model clustering processing on the fused feature set according to the historical temperature signals and vibration signals, and obtain a feature distribution baseline; Performing probability density estimation on the fused feature set according to the feature distribution baseline to obtain a feature probability distribution; Calculating the relative entropy of the probability density of the feature probability distribution of the fused feature set and the feature distribution baseline to generate an initial relative entropy sequence reflecting the difference in feature distribution; A time series dependency relationship in the initial relative entropy sequence is obtained, and a threshold sequence related to the abnormal state of the device is predicted based on the time series dependency relationship to generate a dynamic threshold sequence.

[0012] Furthermore, the step of constructing a detection model based on the fusion feature set and the dynamic threshold sequence, calculating the joint probability of temperature anomaly and vibration anomaly by the detection model, and outputting an abnormal pattern label includes: Performing nonlinear feature mapping on the fused feature set to obtain a deep feature representation; Adaptively calibrating the dynamic threshold sequence according to the deep feature representation to obtain a calibrated threshold sequence; Using the calibration threshold sequence as a constraint, analyzing the interaction between temperature and vibration features in the deep feature representation to generate a joint probability distribution of temperature anomalies and vibration anomalies; Performing variational Bayesian clustering on the abnormal patterns according to the joint probability distribution to obtain an initial abnormal pattern set; The clustering results in the initial abnormal pattern set are mapped to a predefined label set, a probability-based confidence value is assigned to each of the label sets, and an abnormal pattern label including an abnormal pattern label and a confidence score is output.

[0013] Furthermore, the step of performing time serialization processing on the abnormal pattern labels, predicting the failure probability in a future time period based on the time serialized abnormal pattern labels, and generating a failure prediction result includes: Mapping the abnormal pattern label and the timestamp association to a high-dimensional continuous vector space to generate a time series vector; Inputting the time series vector into a long short-term memory network for training to obtain a fault mapping feature; Performing variational autoencoding processing on the fault mapping features to obtain fault probability distribution parameters; generating an acceleration factor based on the first change feature and the second change feature of the fused feature set, and performing correction processing on the fault probability distribution parameter according to the acceleration factor to obtain a corrected probability distribution; Obtaining a potential spatial variable in the corrected probability distribution, generating a fault occurrence time series based on the spatial variable, and obtaining an estimated value of the remaining useful life; The remaining useful life estimate and the failure probability distribution parameters are subjected to multi-objective fusion processing to obtain a failure prediction result.

[0014] Furthermore, the step of prioritizing the fault types in the fault prediction results and generating a maintenance report according to the priority sorting results includes: Quantify the impact of each fault type in the fault prediction results to obtain a fault impact weight set; Dynamically prioritize the fault types according to the fault impact weight set to obtain a priority sequence; Optimize the allocation of maintenance resources according to the priority sequence to obtain a resource allocation plan; The resource allocation plan is matched with a preset maintenance strategy to generate the maintenance report.

[0015] The present invention also provides a detection sensor, comprising: An acquisition module is used to acquire temperature signals and vibration signals of the moving equipment, and synchronously process the temperature signals and vibration signals to output a multi-dimensional data matrix; a feature extraction module, configured to receive the multidimensional data matrix output by the acquisition module, extract a first change feature of the vibration signal and a second change feature of the temperature signal from the multidimensional data matrix, fuse the first change feature and the second change feature, and output a fused feature set; A threshold generation module is used to establish a feature distribution baseline based on a Gaussian mixture model, receive the fused feature set output by the feature extraction module, calculate the relative entropy between the fused feature set and the feature distribution baseline, and generate a dynamic threshold sequence; an anomaly detection module, configured to receive the fused feature set output by the feature extraction module and the dynamic threshold sequence output by the threshold generation module, construct a detection model, calculate the joint probability of temperature anomaly and vibration anomaly, and output an anomaly pattern label; A fault prediction module is configured to receive the abnormal pattern labels output by the abnormality detection module, perform time serialization processing on the abnormal pattern labels, predict the probability of failure in a future time period based on the time serialized abnormal pattern labels, and generate a fault prediction result; The central processing module is used to receive the fault prediction results output by the fault prediction module, prioritize the fault types in the fault prediction results, generate a maintenance report based on the priority sorting results, and coordinate the work of the acquisition module, feature extraction module, threshold generation module, anomaly detection module and fault prediction module.

[0016] Compared with the prior art, this application has the following beneficial effects: The present application proposes a method and a detection sensor for detecting temperature and vibration data of moving equipment. By synchronously processing the temperature signal and vibration signal of the moving equipment and constructing a multidimensional data matrix, the dynamic correlation between the two signals in the time dimension is effectively captured. By extracting the first change feature of the vibration signal and the second change feature of the temperature signal, and fusing them to generate a fusion feature set, the method fully utilizes the complementarity and inherent correlation between the two signals. Compared with the analysis of a single signal feature in traditional methods, it can more accurately characterize the complex changes in the operating status of the equipment and improve the sensitivity and accuracy of anomaly detection. Furthermore, a feature distribution baseline is established based on a Gaussian mixture model, and a dynamic threshold sequence is generated by calculating the relative entropy of the fusion feature set and the baseline, thereby realizing adaptive monitoring of the equipment status and refined classification of abnormal equipment status, providing a scientific basis and priority guidance for maintenance decisions, optimizing the allocation of maintenance resources, and reducing downtime costs and economic losses caused by sudden failures.

[0017] In summary, the present invention comprehensively improves the accuracy, reliability and practicality of temperature and vibration data detection of dynamic equipment through a series of steps such as signal synchronization, feature fusion, dynamic threshold, joint probability modeling, time series prediction and priority sorting, and provides an efficient and intelligent solution for the health management of dynamic equipment. It is particularly suitable for the operation and maintenance of key equipment such as pumps, fans, compressors, and motors, and has significant technical advantages and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0019] The structures, proportions, sizes, etc. depicted in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not intended to limit the conditions under which this application can be implemented, and therefore have no substantive technical significance. Any structural modifications, changes in proportional relationships, or adjustments in size, without affecting the efficacy and objectives that can be achieved by this application, should still fall within the scope of the technical contents disclosed in this application.

[0020] Figure 1 This is a schematic diagram of the steps of a method for detecting temperature and vibration data of a moving device in one embodiment of the present invention; Figure 2 The figure is a schematic block diagram of the structure of a detection sensor according to an embodiment of the present invention.

[0021] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0023] Those skilled in the art will appreciate that, unless expressly stated otherwise, the singular forms "a", "an", "above", and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of features, integers, steps, operations, elements, modules, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, modules, components, and / or groups thereof. It should be understood that when an element is said to be "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. In addition, "connected" or "coupled" as used herein may include wireless connections or wireless couplings. The term "and / or" used herein includes all or any module and all combinations of one or more associated listed items.

[0024] Those skilled in the art will understand that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those skilled in the art in the art to which this invention belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art and, unless specifically defined as such, will not be interpreted in an idealized or overly formal sense.

[0025] Reference Figure 1 , an embodiment of the present invention provides a method for detecting temperature and vibration data of a moving device, comprising the following steps: S1: collecting temperature signals and vibration signals of the moving equipment, and synchronously processing the temperature signals and vibration signals to obtain a multidimensional data matrix; In step S1, dynamic equipment, such as industrial motors, turbines, or pumps, typically generates significant temperature and vibration signatures during operation. Sensors such as thermocouples or infrared thermometers are used to collect temperature signals, and accelerometers or laser vibrometers are used to collect vibration signals. Each collected data point is assigned a time stamp, aligning the temperature and vibration signals on the time axis. After acquiring the raw signals, wavelet denoising is used to remove high-frequency noise while preserving low-frequency trends and mid-frequency signatures. By decomposing the signals into components of varying frequencies, high-frequency noise is identified and filtered out, while retaining low-frequency trends (such as slow temperature changes) and mid-frequency signatures (such as periodic vibration fluctuations) that are meaningful for equipment status analysis. Because the temperature signal (measured in degrees Celsius) and the vibration signal (measured in acceleration g or displacement mm) have different physical dimensions and numerical ranges, normalization is performed to normalize the two signals to a uniform dimension. Z-score normalization or min-max normalization can be used to improve data versatility. After the above processing, the resulting multidimensional data matrix contains time series, amplitude, and phase information. A matrix is ​​a two-dimensional or higher-dimensional array, where each row represents a time point and each column corresponds to a signal feature. For example, for an acquisition process with a sampling period of 1 second and a sampling rate of 1000 Hz, the matrix contains 1000 rows (corresponding to 1000 time points within 1 second), each row including multiple feature values ​​such as temperature amplitude, vibration amplitude, and vibration phase. The phase of the vibration signal can reflect the periodic characteristics of the equipment's operation. For example, rotor imbalance or bearing wear may cause specific phase offsets.

[0026] S2: extracting a first change feature of the vibration signal and a second change feature of the temperature signal from the multidimensional data matrix, and fusing the first change feature and the second change feature to obtain a fused feature set; In step S2, the first variation feature of the vibration signal is extracted and time-frequency analysis is performed on the vibration signal to capture its dynamic characteristics. Vibration signals are non-stationary, with their frequency and amplitude varying over time. Therefore, time-frequency analysis techniques (such as short-time Fourier transform or wavelet transform) can effectively decompose the instantaneous frequency and amplitude variations of the signal. For example, the wavelet transform can be used to decompose the vibration signal into subbands of varying frequency scales, and the instantaneous frequency and amplitude of each subband are extracted as the first variation feature. Assuming the sampling frequency of the vibration signal is 1000 Hz, the wavelet transform decomposition yields low-frequency and high-frequency subbands. The low-frequency subband may reflect the overall vibration trend of the equipment, while the high-frequency subband may capture transient impact characteristics. These characteristics can characterize the vibration behavior of the equipment under different operating conditions. For example, bearing wear may cause an increase in high-frequency amplitude. Extracting the second variation feature of the temperature signal requires analyzing its dynamic variation trend. Temperature signals typically change slowly, but their gradient characteristics (i.e., the rate of change of temperature over time) can reflect subtle changes in the equipment's operating state. For example, by calculating the difference or gradient of the temperature signal, the temperature change rate can be determined. If the equipment is experiencing localized overheating due to friction or overload, the gradient feature will exhibit a significant upward trend. Assuming that the temperature signal fluctuates by 0.1°C per minute during normal operation, but may reach 0.5°C under abnormal conditions, this variation can be quantified by extracting the gradient feature. To further improve feature robustness, the gradient feature can be smoothed (e.g., using a sliding window average) to reduce noise interference. After extracting the first and second change features, they are fused. The goal of this fusion process is to integrate the vibration and temperature features and explore the coupling relationship between them, thereby generating a more representative feature set. Specifically, the instantaneous frequency and amplitude features of the vibration signal and the gradient features of the temperature signal are combined into a high-dimensional feature vector and input into the PCA model. PCA uses a linear transformation to map the high-dimensional features into a low-dimensional space, preserving the main variance of the data. Furthermore, the mutual information coefficient (MIC) between the two can be calculated. Mutual information is a nonlinear correlation measure that quantifies the dependency between vibration and temperature features. For example, when equipment is overloaded, the vibration amplitude may increase, and the temperature gradient may also increase significantly, resulting in a higher mutual information coefficient between the two. By calculating the mutual information coefficient, a temperature-vibration coupling feature can be constructed as part of a fused feature set. Machine learning models (such as random forests or gradient boosting trees) can then be used to evaluate each feature's contribution to the equipment status classification, generate a feature importance score, and adjust the weight accordingly. The resulting fused feature set consists of time-frequency features (the instantaneous frequency and amplitude of the vibration signal), gradient features (the rate of change of the temperature signal), coupling features (the temperature-vibration mutual information coefficient), and corresponding weight coefficients. This feature set can comprehensively characterize the equipment's operating status.

[0027] S3: Establishing a feature distribution baseline based on a Gaussian mixture model, calculating the relative entropy between the fused feature set and the feature distribution baseline, and generating a dynamic threshold sequence; In step S3, the Gaussian mixture model is a probabilistic model that assumes data is generated by a weighted combination of multiple Gaussian distributions. In the context of dynamic equipment temperature and vibration data monitoring, the fused feature set includes primary variation characteristics of the vibration signal (such as amplitude and frequency components) and secondary variation characteristics of the temperature signal (such as temperature fluctuation rate and mean shift). These characteristics exhibit non-uniform distributions. For example, equipment may exhibit multiple typical patterns under different loads or operating conditions. Therefore, by fitting historical health data, the Gaussian mixture model can capture the characteristics of these multimodal distributions, thereby establishing a reliable feature distribution baseline. Specifically, historical health data is processed in steps S1 and S2 to generate a fused feature set based on temperature and vibration data collected from the equipment during normal operation. This data is then fed into the Gaussian mixture model, and the expectation-maximization (EM) algorithm is used to iteratively optimize the model parameters, including the mean, covariance matrix, and mixture weights of each Gaussian component. Ultimately, a baseline model is obtained that describes the feature distribution under normal conditions. After establishing the feature distribution baseline, the relative entropy (KL divergence) between the fused feature set and this baseline is calculated to quantify the degree to which the current feature distribution deviates from the normal state. KL divergence is a metric that measures the difference between two probability distributions. In implementation, the real-time fused feature set is input into a Gaussian mixture model, its probability density function is calculated, and then compared with the baseline model's probability density function to obtain the KL divergence value. A larger KL divergence value indicates a greater deviation from the normal state of the current feature distribution, potentially indicating an anomaly. A dynamic adjustment mechanism is then used to generate a dynamic threshold sequence. By setting a fixed time window (e.g., the last 24 hours of data), the parameters of the Gaussian mixture model are regularly updated to adapt to short-term changes in device operating conditions. For example, assuming a sliding window of 100 data sets, the model parameters are updated every 10 data sets. The mean, covariance, and weights of the Gaussian mixture model are reestimated based on the latest fused feature set within the window, thereby updating the feature distribution baseline. Based on the updated baseline, the KL divergence is recalculated, and an initial threshold is generated based on the statistical properties of the recent KL divergence (e.g., the mean plus twice the standard deviation). This initial threshold is further adaptively adjusted based on changes in the feature distribution within the sliding window, forming a dynamic threshold sequence. Each item in the dynamic threshold sequence corresponds to the anomaly detection threshold at a certain point in time, which can reflect the real-time changes in the operating status of the equipment, thereby improving the accuracy and robustness of anomaly detection.

[0028] S4: constructing a detection model based on the fusion feature set and the dynamic threshold sequence, calculating the joint probability of temperature anomaly and vibration anomaly through the detection model, and outputting an abnormal pattern label; In step S4, when constructing the detection model, a deep belief network (DBN) is used to perform nonlinear mapping on the fused feature set to identify potential abnormal patterns. A DBN is a deep learning model based on a multi-layer restricted Boltzmann machine (RBM). Specifically, the fused feature set is input into the DBN. The network's input layer receives various dimensions of the fused feature set, such as the frequency components of the vibration signal and the fluctuation rate of the temperature signal. Through layer-by-layer training of the multi-layer RBM, the network learns a deep feature representation of the data. These deep features can capture the complex coupling relationship between temperature and vibration signals, such as abnormal vibration frequencies caused by high temperatures or localized temperature increases caused by intense vibration. During training, a dynamic threshold sequence is used as auxiliary input to help the network determine which feature combinations deviate from the normal range. For example, if the relative entropy of a feature set exceeds a dynamic threshold, the network will flag it as a potential abnormal pattern. After completing the nonlinear mapping, the detection model uses Bayesian inference to calculate the joint probability of temperature and vibration anomalies to determine the presence of a temperature-vibration coupling fault. The core of Bayesian reasoning is to use conditional probability models to infer the probability of an abnormal event. Specifically, the features in the fused feature set can be divided into two categories: one related to temperature anomalies (such as excessive temperature fluctuations) and the other related to vibration anomalies (such as sudden changes in vibration amplitude). Using deep features extracted by a deep belief network, the model can identify the correlation between these features. Bayesian reasoning further calculates the joint probability of temperature and vibration anomalies, expressed as P(T, V|F) = P(T|F)·P(V|T, F), where T represents the temperature anomaly, V represents the vibration anomaly, and F represents the fused feature set. This approach allows the model to not only determine the type of a single anomaly but also identify the interactive effects of temperature and vibration anomalies. The joint probability is then matched to predefined anomaly categories, which can include high-temperature vibration anomalies, low-temperature vibration anomalies, and single high-temperature anomalies. Each category is assigned a specific anomaly pattern label. Abnormal pattern labels include an abnormality category and a confidence score. The confidence score can be quantified using a joint probability value. For example, if P(T, V|F) = 0.85, the model believes with 85% confidence that the current state belongs to a specific abnormal pattern and outputs a structured label, such as {abnormality category: high temperature vibration abnormality, confidence level 0.85}. These labels provide key input for subsequent fault prediction and maintenance report generation.

[0029] S5: performing time serialization processing on the abnormal pattern labels, predicting the failure probability in a future time period based on the time serialized abnormal pattern labels, and generating a failure prediction result; In step S5, the abnormal pattern labels are organized chronologically into a continuous time series dataset. A long short-term memory (LSTM) network is used to learn the mapping relationship between historical abnormal patterns and faults. The remaining service life estimate is generated by combining temperature and vibration characteristics with acceleration factor correction. The LSTM network can capture the long-term dependencies between abnormal pattern labels in the temporal dimension. Through its gating mechanism (input gate, forget gate, and output gate), the LSTM model selectively remembers or forgets historical information, thereby learning the mapping relationship between abnormal patterns and faults. Specifically, the LSTM model input is a time window of sequence data, such as the abnormal pattern labels and the corresponding fused feature set for the past 60 minutes. The output is the failure probability for a future time period (e.g., the next 24 hours). Training the LSTM model requires historical data containing known abnormal pattern label sequences and their corresponding actual fault records. For example, if historical data indicates that a certain piece of equipment experienced a bearing failure after three consecutive days of high temperature and low vibration abnormal patterns, the LSTM model will learn the association between this pattern and the bearing failure. During training, the model optimizes parameters by minimizing the error between predicted failure probabilities and actual failures, and calibrates the model using an acceleration factor. The acceleration factor is related to extreme values ​​of temperature and vibration. For example, high temperature accelerates material aging, and high vibration exacerbates mechanical wear. By introducing the acceleration factor into the LSTM model, the estimated failure probability can be corrected, making the prediction more consistent with the device's actual operating conditions. The model receives abnormal pattern labels and a fused feature set from the current time point and a historical time window, and outputs a probability value representing the probability of the device failing within a certain time period (such as 24 hours or 7 days). For example, if the current time is t and the model inputs a sequence of abnormal pattern labels (e.g., normal-high temperature, low vibration-normal) from the past 60 minutes and the corresponding fused feature set, the LSTM model will calculate a probability of failure within the next 24 hours of 0.75. Assuming that the temperature feature indicates that the device's current operating temperature is 10°C higher than normal, and the vibration feature shows a 20% increase in amplitude, using an acceleration factor model (such as the Arrhenius model or the Eyring model), the device's remaining useful life under current operating conditions can be estimated to be 500 hours. This estimate can be combined with the probability of failure to form a comprehensive failure prediction result.

[0030] S6: Prioritize the fault types in the fault prediction results, and generate a maintenance report according to the priority ranking results.

[0031] In step S6, fault types are prioritized based on the prediction results. Prioritization can be based on factors such as fault severity, probability of occurrence, potential impact, and the equipment's operating environment. An intelligent maintenance scheduling algorithm can be employed to assess the urgency and importance of each fault using a fault probability curve and lifespan estimate. In specific implementations, the intelligent maintenance scheduling algorithm can prioritize fault types using a weighted scoring mechanism, assigning weights to each fault type based on factors such as the fault's impact on equipment performance, repair cost, downtime, and safety. For example, bearing wear, which can lead to complete equipment shutdown and high repair costs and impact production efficiency, is therefore given a higher weight; whereas minor overheating, which may only require adjustments to operating parameters, is therefore given a lower weight. The algorithm combines the failure probability, lifespan estimate, and weight to calculate a comprehensive priority score for each fault. Fault types with higher scores are prioritized, forming a prioritized list. Based on the prioritized results, a maintenance report details each fault type, predicted occurrence time, priority, recommended maintenance measures, and resource requirements. Report generation requires consideration of equipment operating parameters and actual operating conditions. For example, for the highest-priority bearing wear fault, the report might recommend scheduling a shutdown for inspection, bearing replacement, and adjusting the equipment speed to slow the wear process within the next 48 hours. For the second-priority overheating fault, the report might recommend real-time temperature monitoring and load reduction, while also scheduling a deeper inspection at the next scheduled shutdown. Alternatively, by analyzing abnormal conditions, specific control instructions can be generated, such as reducing equipment speed by 10% or load by 20%, to extend component life and prevent further failure. These control instructions, along with maintenance recommendations, are incorporated into the report, forming a closed-loop detection and control system.

[0032] In one embodiment, the step of collecting the temperature signal and the vibration signal of the moving device and synchronously processing the temperature signal and the vibration signal to obtain a multidimensional data matrix includes: Based on multi-channel data acquisition, the temperature signal and vibration signal of the moving equipment are collected, and the acquisition timestamp of each signal is recorded at the same time to generate the original signal set containing the temperature and vibration signal amplitude; According to the timestamp information in the original signal set, the temperature and vibration signals are time-aligned based on a linear interpolation algorithm to form a time-synchronized aligned signal set; According to the amplitude range of the alignment signal set, the signal values ​​in the alignment signal set are mapped to a fixed interval for normalization to obtain a normalized signal set; The time series, amplitude and phase information of the standardized signal set are subjected to multi-dimensional reconstruction processing to obtain a multi-dimensional data matrix.

[0033] In the above-described embodiment, a multi-channel data acquisition system is used to accurately capture signals. Dynamic equipment (such as rotating machinery, motors, or pumps) generates temperature and vibration signals during operation. These signals reflect the equipment's operating status. The multi-channel acquisition system simultaneously acquires data from multiple sensors, ensuring signal parallelism and integrity. After acquisition, a raw signal set is generated, consisting of a time series of temperature and vibration signals, with data points in each sequence corresponding to timestamps. Linear interpolation is used to estimate the data value at the intermediate point by assuming a linear change in signal value between two known data points. By applying a similar interpolation process to all signals, the temperature and vibration signals can be aligned to a unified timeline. This aligned signal set ensures that each time point has a corresponding temperature and vibration value, forming a time-synchronized aligned signal set. Based on the amplitude range within the aligned signal set, the signal values ​​are mapped to a fixed interval and normalized to produce a standardized signal set, eliminating differences in signal dimensions and amplitude ranges. Normalized signals are integrated into a structured data representation. The time series provides timing information, the amplitude reflects the signal's intensity, and the phase information is extracted from the vibration signal via Fourier transform or Hilbert transform to characterize the signal's periodicity or frequency characteristics. For example, a vibration signal may contain multiple frequency components (such as 50 Hz and 100 Hz). The phase information of these components can be obtained via a fast Fourier transform (FFT). The reconstruction process organizes this information into a multidimensional matrix. For example, assume the aligned signal set contains N time points, each with one temperature value and three vibration values ​​(x, y, and z axes). The phase information of the vibration signal is extracted via FFT as the phase values ​​of M frequency components. The resulting multidimensional data matrix is ​​an N × (4 + M) matrix, where each row corresponds to a time point and the columns include the standardized temperature value, the x-axis vibration value, the y-axis vibration value, the z-axis vibration value, and the M phase values. This matrix structure can comprehensively characterize the operating state of the moving equipment.

[0034] In one embodiment, the step of extracting a first change feature of a vibration signal and a second change feature of a temperature signal from the multidimensional data matrix, and fusing the first change feature and the second change feature to obtain a fused feature set includes: Decomposing the vibration signal in the multidimensional data matrix into a plurality of frequency band sub-signals to generate a first variation feature including a time-frequency energy feature; Analyzing the temperature signal in the multidimensional data matrix based on the time domain information of the time-frequency energy feature to generate a second change feature including a gradient statistical feature vector; Discretize the first and second change features based on Shannon entropy to generate a preferred feature subset; Constructing a graphical model based on the preferred feature subset, setting each feature dimension in the preferred feature subset as a node in the graphical model, and obtaining a temperature-vibration coupling feature; An operation change state of the moving device is obtained, and weighted processing is performed on the temperature-vibration coupling feature based on the operation change state to generate a fusion feature set.

[0035] In the above embodiment, a wavelet packet decomposition algorithm is applied to the vibration signal in a multidimensional data matrix. Through recursive decomposition, the signal is stratified into high-frequency and low-frequency components. The sum of squares of the subnodes in each layer is calculated to characterize the energy distribution. The energy values ​​across all frequency bands are normalized and arranged in a time series, generating a set of time-frequency energy feature vectors that reflect the dynamic changes in the vibration signal. Based on the time-domain information of the time-frequency energy feature, the temperature signal in the multidimensional data matrix is ​​analyzed to generate a second variation feature consisting of a gradient statistical feature vector. This process utilizes the time-domain characteristics of the vibration signal to guide feature extraction of the temperature signal, thereby capturing the potential coupling relationship between the temperature and vibration signals. The time-domain information of the time-frequency energy feature includes the mean, variance, or rate of change of the energy sequence. Analysis of the temperature signal can be achieved by calculating the gradient of its time series. The gradient reflects the rate of change of the temperature signal, for example, by calculating the temperature change at each time point using a difference formula. Furthermore, the gradient statistical feature vector can be generated by statistically analyzing the characteristics of the gradient sequence (such as the mean, standard deviation, maximum, and minimum values). The first and second variation features are discretized based on Shannon entropy. Shannon entropy is a measure of information uncertainty. The first variation feature (time-frequency energy matrix) and the second variation feature (gradient statistics vector) contain redundant or low-information dimensions. Discretization can reduce dimensionality while retaining key information. Specifically, discretization can be achieved by binning continuous feature values ​​into discrete intervals. For example, time-frequency energy values ​​or gradient values ​​are mapped to a fixed number of intervals and the Shannon entropy of each feature dimension is calculated. Feature dimensions with lower entropy values ​​indicate higher information content and should be retained first. In the gradient statistics vector of the second variation feature, statistics with lower entropy values ​​(such as standard deviation) are selected. A preferred feature subset includes a portion of the time-frequency energy sequence and gradient statistics, forming a reduced-dimensional feature set. A graphical model is constructed based on the preferred feature subset. A graphical model is a data structure that represents relationships between features and is suitable for capturing the interactive relationship between temperature and vibration features. Specifically, each feature dimension in the preferred feature subset (such as the mean energy value in a frequency band or the standard deviation of the temperature gradient) is defined as a node in a graph, and edges between nodes are determined by inter-feature correlations (such as the Pearson correlation coefficient). The graphical model can be represented by an adjacency matrix, where the matrix elements represent edge weights. The temperature-vibration coupling feature is a structured representation of the graphical model, including node features (eigenvalues) and edge weights (correlations). The operational change state of the dynamic equipment is obtained. The operational change state may include the load, speed, or failure mode of the equipment. This can be obtained through sensors or control systems, and weights are assigned to the nodes or edges of the graphical model. The weights are determined based on the operational state. For example, under high load conditions, the weights of nodes related to high-frequency vibrations may increase. The fused feature set is the weighted graphical model features and may be output in the form of a vector or matrix.

[0036] In one embodiment, the step of analyzing the temperature signal in the multidimensional data matrix based on the time domain information of the time-frequency energy feature to generate a second change feature including a gradient statistical feature vector further includes: Constructing a sliding window with a preset number of bits, traversing the temperature signal in the multidimensional data matrix with the sliding window, and calculating a gradient sequence of temperature values ​​in each window; Extracting first-order statistics and second-order statistics of the gradient sequence, and adjusting the step size of the window in combination with the time domain information of the time-frequency energy feature to generate statistics; The statistics corresponding to multiple windows are concatenated to generate a gradient statistical feature vector.

[0037] In the above embodiment, a sliding window with a preset number of bits is constructed. The preset number of bits refers to the length of the window, defined by the number of data points or the length of time. For example, assuming the sampling frequency of the temperature signal is 1 Hz (i.e., one temperature value is collected per second), a sliding window of length 10 corresponds to 10 seconds of temperature data. The window step size, i.e., the distance of each sliding movement, can be equal to the window length (no overlap) or less than the window length (with overlap). In practical applications, the step size is optimized based on the dynamic characteristics of the signal. As the sliding window traverses the temperature signal, a gradient sequence of temperature values ​​is calculated within each window. The gradient sequence reflects the rate of change of the temperature signal in the time dimension and is calculated by the difference between adjacent data points. First-order and second-order statistics are extracted from the gradient sequence. First-order statistics may include mean, minimum, and maximum values, reflecting the overall level of the gradient sequence. Second-order statistics may include variance and standard deviation, reflecting the volatility of the gradient sequence. For concrete illustration, assume the temperature sequence within a window is [20, 21, 22, 23, 24] (unit: degrees Celsius), and its gradient sequence is [1, 1, 1, 1]. This gradient sequence has a mean of 1 and a variance of 0, indicating that the temperature signal increases linearly at a constant rate within the window. The window step size is adjusted by incorporating the time-domain information of the time-frequency energy features. The timestamp information of the vibration signal's time-frequency energy features is used to dynamically adjust the sliding window step size. For example, when the vibration signal's time-frequency energy features show a significant increase in energy within a certain time period, indicating a change in the device's state, the window step size can be reduced to increase the analysis precision and more accurately capture the temperature signal changes. After calculating the statistics within each window, the statistics corresponding to multiple windows are concatenated to generate a gradient statistical feature vector. Specifically, each window generates a set of statistics, such as the mean and variance. Assuming there are N windows, the gradient statistical feature vector is a vector of length 2N that summarizes the temperature signal's changing characteristics over different time periods.

[0038] In one embodiment, the step of discretizing the first and second change features based on Shannon entropy to generate a preferred feature subset further includes: Discretize each feature dimension of the first change feature and the second change feature to generate a probability distribution of the feature; Calculating the entropy value of each feature dimension using the Shannon entropy formula according to the probability distribution to obtain an entropy value set of the first change feature and the second change feature; The information gain of the first change feature and the second change feature is calculated based on the entropy value set, an information gain ranking is generated, and feature dimensions below a preset information gain threshold are removed from the first change feature and the second change feature to form the preferred feature subset.

[0039] In the above embodiment, each feature dimension of the first change feature and the second change feature is discretized to generate a probability distribution of the feature. Discretization is the process of converting continuous or high-dimensional feature values ​​into discrete intervals. The frequency of each sample falling into these intervals is counted, and the probability of each interval is calculated to form a probability distribution of the dimension, and the continuous value of the high-dimensional feature is converted into a discrete probability distribution. Based on the generated probability distribution, the entropy value of each feature dimension is calculated using the Shannon entropy formula to obtain the entropy value set of the first change feature and the second change feature. Shannon entropy is used to measure the uncertainty of a random variable, and the expression is: ,in is the probability of the random variable X being in the i-th state. By calculating the entropy value for each feature dimension, a set of entropy values ​​is obtained. The entropy value reflects the uncertainty of the feature dimension: dimensions with lower entropy values ​​generally indicate a more concentrated data distribution and higher information content; dimensions with higher entropy values ​​may indicate a more even distribution and lower information content. Based on the entropy value set, the information gain of the first and second variation features is calculated, and an information gain ranking is generated. Information gain is a metric that measures the contribution of a feature dimension to a classification or prediction task and is defined as the difference between the overall entropy and the conditional entropy. For example, for a dimension of a vibration signal, the entropy of the entire dataset is calculated (based on the distribution of fault labels). The data is then grouped according to this dimension, and the conditional entropy of each group is calculated. The weighted sum is used to calculate the information gain. This process is repeated for each feature dimension to obtain the information gain values ​​for all dimensions and sort them in descending order to form an information gain ranking. Feature dimensions below a preset information gain threshold are removed from the first and second variation features to form a preferred feature subset. Assuming the preset threshold is 0.1, the information gain ranking is iterated, retaining dimensions with gain values ​​above 0.1 and removing those below 0.1. Through this screening process, redundant or low-information features are removed, and the subset retains the feature dimensions that contribute most to the classification task. To balance computational complexity and information integrity, the threshold or subset size can be adjusted through iterative optimization. For example, if the initial subset is too large, the threshold can be increased and rescreened; if the subset is too small, the threshold can be lowered to retain more dimensions, generating a preferred feature subset containing high-information features.

[0040] In one embodiment, the steps of establishing a feature distribution baseline based on a Gaussian mixture model, calculating the relative entropy between the fused feature set and the feature distribution baseline, and generating a dynamic threshold sequence include: Acquire historical temperature signals and vibration signals of the moving equipment, perform Gaussian mixture model clustering processing on the fused feature set according to the historical temperature signals and vibration signals, and obtain a feature distribution baseline; Performing probability density estimation on the fused feature set according to the feature distribution baseline to obtain a feature probability distribution; Calculating the relative entropy of the probability density of the feature probability distribution of the fused feature set and the feature distribution baseline to generate an initial relative entropy sequence reflecting the difference in feature distribution; A time series dependency relationship in the initial relative entropy sequence is obtained, and a threshold sequence related to the abnormal state of the device is predicted based on the time series dependency relationship to generate a dynamic threshold sequence.

[0041] In the above embodiment, historical device temperature and vibration signals are acquired and, based on these signals, a Gaussian mixture model (GMM) clustering process is performed on the fused feature set to obtain a feature distribution baseline. A Gaussian mixture model is a probabilistic model that assumes the data is composed of multiple Gaussian distributions, each representing a potential operating state (e.g., normal operation, minor anomaly, etc.). Using the Expectation Maximum (EM) algorithm, the GMM model is fitted to obtain the mean, covariance, and weight of each Gaussian component, thereby constructing a feature distribution baseline. This baseline is essentially a probability density function that describes the feature distribution of the device under normal or historical operating conditions. Based on the feature distribution baseline, a probability density estimation is performed on the fused feature set to obtain a feature probability distribution. The currently collected device data (i.e., the new fused feature set) is then substituted into the GMM model to calculate its probability density with respect to the baseline distribution. The probability density reflects the likelihood that the current feature vector belongs to the baseline distribution. A high probability density indicates that the current feature vector is close to the historical baseline, indicating that the device is likely operating normally; otherwise, an anomaly may exist. The probability density estimation process essentially maps a high-dimensional feature space to a probability value. The relative entropy between the probability distribution of the fused feature set and the probability density of the baseline feature distribution is calculated to generate an initial relative entropy sequence reflecting the difference between the feature distributions. Relative entropy (Kullback-Leibler divergence) is a metric that measures the difference between two probability distributions. Since the probability density of the GMM model is typically continuous, numerical integration or Monte Carlo methods can be used to approximate the relative entropy to obtain an initial sequence that reflects the degree of difference between the current feature distribution and the historical baseline. Temporal dependencies in the initial relative entropy sequence are obtained and used to predict threshold sequences associated with device abnormalities, generating a dynamic threshold sequence. The relative entropy sequence is a time series whose values ​​vary over time and are affected by the device's operating status. Temporal dependencies can be extracted using time series analysis methods, such as autoregressive models (AR), to predict future relative entropy values. Based on these predicted values, a dynamic threshold sequence can be derived. The generation of dynamic thresholds is based on the anomaly detection assumption: when the relative entropy exceeds a certain threshold, the device is considered to be in an abnormal state. To make the thresholds adaptive, statistical methods (such as the mean plus a certain standard deviation) or machine learning methods (such as anomaly detection algorithms) can be combined to determine the thresholds. For example, assuming the mean of the relative entropy series is 0.2 and the standard deviation is 0.05, the threshold can be set to 0.2 + 3 × 0.05 = 0.35. If the relative entropy at a given moment exceeds 0.35, the system triggers an anomaly alarm. The above steps complete the process from data collection to dynamic threshold generation. Relative entropy is used to quantify the difference between the current state and the historical baseline, and dynamic thresholds are generated by combining time series analysis.

[0042] In one embodiment, the step of constructing a detection model based on the fused feature set and the dynamic threshold sequence, calculating the joint probability of temperature anomaly and vibration anomaly using the detection model, and outputting an abnormal pattern label includes: Performing nonlinear feature mapping on the fused feature set to obtain a deep feature representation; Adaptively calibrating the dynamic threshold sequence according to the deep feature representation to obtain a calibrated threshold sequence; Using the calibration threshold sequence as a constraint, analyzing the interaction between temperature and vibration features in the deep feature representation to generate a joint probability distribution of temperature anomalies and vibration anomalies; Performing variational Bayesian clustering on the abnormal patterns according to the joint probability distribution to obtain an initial abnormal pattern set; The clustering results in the initial abnormal pattern set are mapped to a predefined label set, a probability-based confidence value is assigned to each of the label sets, and an abnormal pattern label including an abnormal pattern label and a confidence score is output.

[0043] In the above embodiment, feature transformation is performed on the fused feature set to extract more expressive deep features. Nonlinear feature mapping can be achieved using deep learning models (such as neural networks). These raw features can be mapped into a high-dimensional space to generate a deep feature representation. This mapping is capable of capturing the complex nonlinear relationship between temperature and vibration. Based on the distribution characteristics of the deep feature representation and the statistical properties of the dynamic threshold sequence, the threshold sequence is recalibrated using kernel density estimation to adapt to the dynamic changes in feature distribution under different operating conditions. The calibration process analyzes the cluster structure of the deep feature representation in high-dimensional space and dynamically adjusts the upper and lower bounds of the threshold sequence to accurately distinguish between normal and abnormal states, generating a more robust calibration threshold sequence. Based on the deep feature representation and the calibration threshold sequence, multimodal collaborative analysis is performed on both to obtain a joint feature distribution. By constructing a temperature-vibration collaborative analysis module, the interactive relationship between temperature and vibration features in the deep feature representation is analyzed. In combination with the constraints of the calibration threshold sequence, a Bayesian network inference method is used to generate a joint probability distribution of temperature anomalies and vibration anomalies. The processing process uses conditional probability modeling to capture the coupling effects between temperature anomalies (such as overheating) and vibration anomalies (such as unbalanced vibration), generating a joint feature distribution containing multimodal interaction information. A variational Bayesian approach is used to perform probabilistic clustering on the joint feature distribution. By iteratively optimizing the latent variable model of the feature distribution, feature points are assigned to different abnormal pattern clusters, such as normal operation, mild vibration anomalies, or severe temperature-vibration coupling anomalies. The clustering process estimates the abnormal mode to which each feature point belongs by maximizing the posterior probability. A calibrated threshold sequence is introduced as a priori constraint to generate an initial abnormal pattern set containing preliminary classification results. The clustering results in this initial abnormal pattern set are mapped to a predefined set of labels, each of which is assigned a probability-based confidence value. The output is an abnormal pattern label consisting of the abnormal pattern label and the confidence score. The unsupervised clustering results are then associated with the known abnormal labels. In implementation, a mapping function (e.g., based on a nearest neighbor or softmax classifier) ​​can be used to align the clustering results with the predefined label set. The predefined label set may include categories such as normal and overheating fault. Each clustering result is assigned a confidence score based on its similarity to the label set. For example, in an industrial equipment scenario, suppose the initial abnormal pattern set contains two clusters, one highly correlated with the overheating fault label (confidence 0.9) and the other with the mechanical failure label (confidence 0.8). The output contains a set of abnormal pattern labels and their confidence scores.

[0044] In one embodiment, the step of performing time serialization processing on the abnormal pattern labels, predicting the failure probability in a future time period based on the time serialized abnormal pattern labels, and generating a failure prediction result includes: Mapping the abnormal pattern label and the timestamp association to a high-dimensional continuous vector space to generate a time series vector; Inputting the time series vector into a long short-term memory network for training to obtain a fault mapping feature; Performing variational autoencoding processing on the fault mapping features to obtain fault probability distribution parameters; generating an acceleration factor based on the first change feature and the second change feature of the fused feature set, and performing correction processing on the fault probability distribution parameter according to the acceleration factor to obtain a corrected probability distribution; Obtaining a potential spatial variable in the corrected probability distribution, generating a fault occurrence time series based on the spatial variable, and obtaining an estimated value of the remaining useful life; The remaining useful life estimate and the failure probability distribution parameters are subjected to multi-objective fusion processing to obtain a failure prediction result.

[0045] In the above embodiment, abnormal pattern labels are associated with timestamps and mapped into a high-dimensional continuous vector space. Each abnormal pattern label is bound to a corresponding timestamp, forming an ordered sequence of data points. After generating time series vectors, these vectors are fed into a long short-term memory (LSTM) network for training to extract fault mapping features. LSTM is a recurrent neural network specialized for processing time series data, capable of capturing long-term dependencies. During training, the time series vectors serve as input. The LSTM analyzes the vector sequence time-step by time-step through its gating mechanisms (such as the input gate, forget gate, and output gate), learning the temporal evolution of abnormal patterns and thereby defining specific fault patterns. After training, the LSTM outputs fault mapping features, which are abstract representations of the temporal dynamics of the abnormal pattern and contain information that indicates the onset of a fault. To further enhance the expressiveness of the features, the fault mapping features are processed using a variational autoencoder (VAE). The VAE compresses the fault mapping features into probability distribution parameters (mean or variance) in a latent space through an encoder, and then reconstructs the data through a decoder. The first and second variation features of the fused feature set are used to generate an acceleration factor, which is used to correct the fault probability distribution parameters. Specifically, this can be accomplished by constructing an acceleration factor model. The model takes the first and second variation characteristics as input and outputs a scalar value. The correction process involves adjusting the probability distribution parameters of the VAE output, such as scaling the mean or variance using an acceleration factor, to make the probability distribution more consistent with actual operating conditions. After obtaining the corrected probability distribution, latent space variables are extracted and used to generate a failure occurrence time series to estimate the remaining useful life (RUL). Latent space variables are sampling points in the VAE latent space, representing a stochastic expression of the failure probability. By decoding or mapping these variables, a virtual failure occurrence time series is generated. The remaining useful life estimate is calculated based on this series. For example, the average time of failure occurrence in the series is used as the RUL estimate. This estimate combines the randomness of the probability distribution with the dynamic nature of the time series, resulting in high prediction accuracy. The remaining useful life estimate is then combined with the failure probability distribution parameters for a multi-objective fusion process to generate the final fault prediction result. Multi-objective fusion involves simultaneously optimizing multiple prediction objectives, such as maximizing the accuracy of the RUL estimate and minimizing the error of the failure probability prediction. The specific implementation can adopt a multi-objective optimization algorithm (such as weighted sum method or Pareto optimization) to integrate the RUL estimation value and probability distribution parameters into a comprehensive prediction result.

[0046] In one embodiment, the step of prioritizing the fault types in the fault prediction results and generating a maintenance report according to the priority ranking results includes: Quantify the impact of each fault type in the fault prediction results to obtain a fault impact weight set; Dynamically prioritize the fault types according to the fault impact weight set to obtain a priority sequence; Optimize the allocation of maintenance resources according to the priority sequence to obtain a resource allocation plan; The resource allocation plan is matched with a preset maintenance strategy to generate the maintenance report.

[0047] In the above embodiment, the impact of each fault type in the fault prediction results is quantified. This quantification can be achieved by establishing a weighted evaluation model. The model inputs include the failure probability curve, the RUL estimate, and historical fault impact data. The comprehensive impact weight of each fault type is calculated to generate a set of fault impact weights. A weighted ranking method is used, using the fault impact weights as the primary ranking basis, while also incorporating other factors (such as fault urgency or maintenance resource availability) for auxiliary adjustment. The ranking result is a priority sequence that reflects the urgency and importance of fault handling. After obtaining the priority sequence, maintenance resources are optimized based on the sequence to generate a resource allocation plan. Maintenance resources include manpower, spare parts, tools, and time. The optimization goal is to maximize resource utilization efficiency and minimize the impact of faults. Resource allocation can be implemented using linear programming or heuristic algorithms. For example, the optimization objective function of the linear programming model can be to minimize total maintenance costs, with constraints including maintenance personnel availability, spare parts inventory, and fault priority. The operating parameter adjustment module can also be integrated into resource allocation optimization. For example, by reducing the speed or load of the wind turbine, the RUL of mechanical wear can be extended, thereby buying more time for repairing electrical faults. This coordinated optimization of parameter adjustment and resource allocation forms a closed-loop detection and control system, improving maintenance efficiency. The resource allocation plan is matched with the preset maintenance strategy to generate a maintenance report. The preset maintenance strategy can include rules such as preventive maintenance, emergency repairs, or regular inspections. The matching process aims to ensure that the resource allocation plan meets the strategy requirements and generates clear maintenance instructions. The matching process can be implemented through a rule engine or decision tree, comparing the maintenance tasks, personnel arrangements, and time plans in the resource allocation plan with the strategy rules one by one, forming a closed-loop detection and control system.

[0048] Reference Figure 2 A detection sensor, applied to any of the above detection methods, comprising: The acquisition module 100 is used to acquire temperature and vibration signals from a moving device, synchronously process the temperature and vibration signals, and output a multidimensional data matrix. The acquisition module includes a temperature sensor unit, a vibration sensor unit, and a synchronous processing unit. The temperature sensor unit is used to acquire the temperature signal and vibration signal of the moving device in real time, and the vibration sensor unit is used to acquire the vibration signal of the moving device in real time. The synchronous processing unit is used to align the time axes and unify the sampling rates of the temperature and vibration signals to generate a multidimensional data matrix containing temperature and vibration data.

[0049] The feature extraction module 200 is configured to receive the multidimensional data matrix output by the acquisition module, extract a first variation feature of the vibration signal and a second variation feature of the temperature signal from the multidimensional data matrix, fuse the first variation feature and the second variation feature, and output a fused feature set. The feature extraction module includes a feature extraction unit and a feature fusion unit. The feature extraction unit uses a signal processing algorithm to extract the time-frequency domain features of the vibration signal as the first variation feature and the statistical features of the temperature signal as the second variation feature. The feature fusion unit uses a weighted fusion algorithm to perform multidimensional mapping on the first variation feature and the second variation feature to generate a fused feature set.

[0050] The threshold generation module 300 is configured to establish a feature distribution baseline based on a Gaussian mixture model, receive the fused feature set output by the feature extraction module, calculate the relative entropy between the fused feature set and the feature distribution baseline, and generate a dynamic threshold sequence. The threshold generation module includes a baseline modeling unit and a threshold calculation unit. The baseline modeling unit uses a Gaussian mixture model to perform probability distribution modeling on the historical fused feature set to generate a feature distribution baseline. The threshold calculation unit uses a relative entropy calculation algorithm to analyze the difference between the fused feature set and the feature distribution baseline to generate a dynamic threshold sequence reflecting the degree of anomaly.

[0051] Anomaly detection module 400 is configured to receive the fused feature set output by the feature extraction module and the dynamic threshold sequence output by the threshold generation module, construct a detection model, calculate the joint probability of temperature anomalies and vibration anomalies, and output an anomaly pattern label. The anomaly detection module includes a model construction unit and a probability calculation unit. The model construction unit constructs a probability detection model based on the fused feature set and dynamic threshold sequence, and the probability calculation unit analyzes the correlation between temperature anomalies and vibration anomalies using a joint probability calculation method to generate an anomaly pattern label.

[0052] The fault prediction module 500 is configured to receive the anomaly pattern labels output by the anomaly detection module, perform time-series processing on the anomaly pattern labels, and predict the probability of failure in a future time period based on the time-series anomaly pattern labels to generate a fault prediction result. The fault prediction module includes a serialization processing unit and a prediction unit. The serialization processing unit performs time series modeling on the anomaly pattern labels using a time series analysis algorithm, and the prediction unit calculates the probability of failure in a future time period using a probabilistic prediction algorithm to generate a fault prediction result containing the fault type and probability.

[0053] The central processing module 600 is configured to receive the fault prediction results output by the fault prediction module, prioritize the fault types in the fault prediction results, generate a maintenance report based on the priority ranking results, and coordinate the operations of the acquisition module, feature extraction module, threshold generation module, anomaly detection module, and fault prediction module. The central processing module includes a priority ranking unit, a report generation unit, and a coordination control unit. The priority ranking unit uses a priority ranking algorithm to sort the fault types and generate a priority sequence. The report generation unit generates a structured maintenance report based on the priority sequence. The coordination control unit uses a task scheduling algorithm to manage the data flow and processing timing of each module.

[0054] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for detecting temperature and vibration data of a moving device, characterized in that: include: collecting temperature signals and vibration signals of the moving equipment, and synchronously processing the temperature signals and vibration signals to obtain a multi-dimensional data matrix; extracting a first change feature of the vibration signal and a second change feature of the temperature signal from the multidimensional data matrix, and fusing the first change feature and the second change feature to obtain a fused feature set; Establishing a feature distribution baseline based on a Gaussian mixture model, calculating the relative entropy between the fused feature set and the feature distribution baseline, and generating a dynamic threshold sequence; Building a detection model based on the fusion feature set and the dynamic threshold sequence, calculating the joint probability of temperature anomaly and vibration anomaly through the detection model, and outputting an abnormal pattern label; Performing time serialization processing on the abnormal pattern labels, predicting the failure probability in a future time period based on the time serialized abnormal pattern labels, and generating a failure prediction result; The fault types in the fault prediction results are prioritized, and a maintenance report is generated according to the priority ranking results.

2. The method for detecting temperature and vibration data of a moving device according to claim 1, characterized in that: The step of collecting the temperature signal and the vibration signal of the moving device, synchronously processing the temperature signal and the vibration signal to obtain a multidimensional data matrix includes: Based on multi-channel data acquisition, the temperature signal and vibration signal of the moving equipment are collected, and the acquisition timestamp of each signal is recorded at the same time to generate the original signal set containing the temperature and vibration signal amplitude; According to the timestamp information in the original signal set, the temperature and vibration signals are time-aligned based on a linear interpolation algorithm to form a time-synchronized aligned signal set; According to the amplitude range of the alignment signal set, the signal values ​​in the alignment signal set are mapped to a fixed interval for normalization to obtain a normalized signal set; The time series, amplitude and phase information of the standardized signal set are subjected to multi-dimensional reconstruction processing to obtain a multi-dimensional data matrix.

3. The method for detecting temperature and vibration data of a moving device according to claim 1, wherein: The step of extracting the first change feature of the vibration signal and the second change feature of the temperature signal from the multidimensional data matrix, and fusing the first change feature and the second change feature to obtain a fused feature set includes: Decomposing the vibration signal in the multidimensional data matrix into a plurality of frequency band sub-signals to generate a first variation feature including a time-frequency energy feature; Analyzing the temperature signal in the multidimensional data matrix based on the time domain information of the time-frequency energy feature to generate a second change feature including a gradient statistical feature vector; Discretize the first and second change features based on Shannon entropy to generate a preferred feature subset; Constructing a graphical model based on the preferred feature subset, setting each feature dimension in the preferred feature subset as a node in the graphical model, and obtaining a temperature-vibration coupling feature; An operation change state of the moving device is obtained, and weighted processing is performed on the temperature-vibration coupling feature based on the operation change state to generate a fusion feature set.

4. The method for detecting temperature and vibration data of a moving device according to claim 3, characterized in that: The step of analyzing the temperature signal in the multidimensional data matrix based on the time domain information of the time-frequency energy feature to generate a second change feature including a gradient statistical feature vector further includes: Constructing a sliding window with a preset number of bits, traversing the temperature signal in the multidimensional data matrix with the sliding window, and calculating a gradient sequence of temperature values ​​in each window; Extracting first-order statistics and second-order statistics of the gradient sequence, and adjusting the step size of the window in combination with the time domain information of the time-frequency energy feature to generate statistics; The statistics corresponding to multiple windows are concatenated to generate a gradient statistical feature vector.

5. The method for detecting temperature and vibration data of a moving device according to claim 3, wherein: The step of discretizing the first and second change features based on Shannon entropy to generate a preferred feature subset further includes: Discretize each feature dimension of the first change feature and the second change feature to generate a probability distribution of the feature; Calculating the entropy value of each feature dimension using the Shannon entropy formula according to the probability distribution to obtain an entropy value set of the first change feature and the second change feature; The information gain of the first change feature and the second change feature is calculated based on the entropy value set, an information gain ranking is generated, and feature dimensions below a preset information gain threshold are removed from the first change feature and the second change feature to form the preferred feature subset.

6. The method for detecting temperature and vibration data of a moving device according to claim 1, characterized in that: The steps of establishing a feature distribution baseline based on a Gaussian mixture model, calculating the relative entropy between the fused feature set and the feature distribution baseline, and generating a dynamic threshold sequence include: Acquire historical temperature signals and vibration signals of the moving equipment, perform Gaussian mixture model clustering processing on the fused feature set according to the historical temperature signals and vibration signals, and obtain a feature distribution baseline; Performing probability density estimation on the fused feature set according to the feature distribution baseline to obtain a feature probability distribution; Calculating the relative entropy of the probability density of the feature probability distribution of the fused feature set and the feature distribution baseline to generate an initial relative entropy sequence reflecting the difference in feature distribution; A time series dependency relationship in the initial relative entropy sequence is obtained, and a threshold sequence related to the abnormal state of the device is predicted based on the time series dependency relationship to generate a dynamic threshold sequence.

7. The method for detecting temperature and vibration data of a moving device according to claim 1, characterized in that: The step of constructing a detection model based on the fusion feature set and the dynamic threshold sequence, calculating the joint probability of temperature anomaly and vibration anomaly by using the detection model, and outputting an abnormal pattern label includes: Performing nonlinear feature mapping on the fused feature set to obtain a deep feature representation; Adaptively calibrating the dynamic threshold sequence according to the deep feature representation to obtain a calibrated threshold sequence; Using the calibration threshold sequence as a constraint, analyzing the interaction between temperature and vibration features in the deep feature representation to generate a joint probability distribution of temperature anomalies and vibration anomalies; Performing variational Bayesian clustering on the abnormal patterns according to the joint probability distribution to obtain an initial abnormal pattern set; The clustering results in the initial abnormal pattern set are mapped to a predefined label set, a probability-based confidence value is assigned to each of the label sets, and an abnormal pattern label including an abnormal pattern label and a confidence score is output.

8. The method for detecting temperature and vibration data of a moving device according to claim 1, characterized in that: The step of performing time serialization processing on the abnormal pattern labels, predicting the fault probability in a future time period based on the time serialized abnormal pattern labels, and generating a fault prediction result includes: Mapping the abnormal pattern label and the timestamp association to a high-dimensional continuous vector space to generate a time series vector; Inputting the time series vector into a long short-term memory network for training to obtain a fault mapping feature; Performing variational autoencoding processing on the fault mapping features to obtain fault probability distribution parameters; generating an acceleration factor based on the first change feature and the second change feature of the fused feature set, and performing correction processing on the fault probability distribution parameter according to the acceleration factor to obtain a corrected probability distribution; Obtaining a potential spatial variable in the corrected probability distribution, generating a fault occurrence time series based on the spatial variable, and obtaining an estimated value of the remaining useful life; The remaining useful life estimate and the failure probability distribution parameters are subjected to multi-objective fusion processing to obtain a failure prediction result.

9. The method for detecting temperature and vibration data of a moving device according to claim 1, characterized in that: The step of prioritizing the fault types in the fault prediction results and generating a maintenance report according to the priority sorting results includes: Quantify the impact of each fault type in the fault prediction results to obtain a fault impact weight set; Dynamically prioritize the fault types according to the fault impact weight set to obtain a priority sequence; Optimize the allocation of maintenance resources according to the priority sequence to obtain a resource allocation plan; The resource allocation plan is matched with a preset maintenance strategy to generate the maintenance report.

10. A detection sensor, applied to the detection method according to any one of claims 1 to 9, characterized in that: include: An acquisition module is used to acquire temperature signals and vibration signals of the moving equipment, and synchronously process the temperature signals and vibration signals to output a multi-dimensional data matrix; a feature extraction module, configured to receive the multidimensional data matrix output by the acquisition module, extract a first change feature of the vibration signal and a second change feature of the temperature signal from the multidimensional data matrix, fuse the first change feature and the second change feature, and output a fused feature set; A threshold generation module is used to establish a feature distribution baseline based on a Gaussian mixture model, receive the fused feature set output by the feature extraction module, calculate the relative entropy between the fused feature set and the feature distribution baseline, and generate a dynamic threshold sequence; an anomaly detection module, configured to receive the fused feature set output by the feature extraction module and the dynamic threshold sequence output by the threshold generation module, construct a detection model, calculate the joint probability of temperature anomaly and vibration anomaly, and output an anomaly pattern label; A fault prediction module is configured to receive the abnormal pattern labels output by the abnormality detection module, perform time serialization processing on the abnormal pattern labels, predict the probability of failure in a future time period based on the time serialized abnormal pattern labels, and generate a fault prediction result; The central processing module is used to receive the fault prediction results output by the fault prediction module, prioritize the fault types in the fault prediction results, generate a maintenance report based on the priority sorting results, and coordinate the work of the acquisition module, feature extraction module, threshold generation module, anomaly detection module and fault prediction module.

Images