Mechanical equipment operation state monitoring method, system and equipment based on deep learning and medium

Abstract

The invention relates to a deep learning-based mechanical equipment operation state monitoring method and system, equipment and a medium, and the method comprises the steps: carrying out the cutting and marking of time series data of a multi-source sensor of the equipment, and constructing a sample set; converting the sample into a three-dimensional time-frequency characteristic tensor; taking the tensor as input, calling a deep neural network with a built-in Monte Carlo random discard layer to carry out probabilistic training, and optimizing a negative log-likelihood loss function to obtain a probability prediction network model; the model is used for carrying out multiple times of forward propagation sampling on online monitoring data, and probability distribution prediction and confidence intervals of the remaining service life are obtained through aggregation output; and analyzing cognitive and accidental uncertainty in a prediction result, and generating a risk early warning and maintenance decision in combination with a threshold rule. According to the method, the defects that the prediction result is only a single deterministic value and the prediction reliability cannot be evaluated in the prior art are overcome, probabilistic prediction and uncertainty quantification of the remaining service life are realized, and a scientific risk basis is provided for maintenance decision making.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and predictive maintenance technology for mechanical equipment, specifically to a method, system, equipment and medium for monitoring the operating status of mechanical equipment based on deep learning. Background Technology

[0002] In the field of predictive and health management of industrial equipment, predicting the remaining useful life of machinery is a core task. This method aims to utilize condition monitoring data generated during equipment operation, such as vibration and temperature, to predict the remaining usable time from the current moment until a failure or performance degradation occurs, through data analysis models. Accurate prediction of remaining useful life provides crucial information for developing condition-based maintenance strategies, thereby enabling a shift from traditional scheduled or reactive maintenance to more efficient and economical predictive maintenance. This is of great significance for ensuring safe equipment operation, avoiding unplanned downtime, and reducing maintenance costs.

[0003] However, current deep learning-based methods for predicting remaining service life still have a significant limitation. Most existing models are deterministic prediction models, outputting a single, deterministic point estimate (e.g., "100 hours remaining"). This prediction approach fails to adequately consider the inherent randomness of equipment degradation and the model's own cognitive limitations when facing unknown operating conditions. A single prediction value cannot reflect the reliability or confidence level of the prediction result, making it difficult for maintenance personnel to assess the risk of potential prediction errors, and thus unable to make scientific and evidence-based choices between different decisions such as "immediate shutdown for maintenance" and "continued observation and operation." Summary of the Invention

[0004] Based on this, the purpose of the present invention is to provide a method, system, device and medium for monitoring the operating status of mechanical equipment based on deep learning, which can quantify and predict uncertainties and provide probabilistic prediction of remaining service life.

[0005] The objective of this invention is achieved through the following solution:

[0006] In a first aspect, the present invention provides a method for monitoring the operating status of mechanical equipment based on deep learning, comprising the following steps:

[0007] S1: Perform sliding window segmentation and labeling on the time-series data of the operating status synchronously collected by multi-source sensors deployed on mechanical equipment, divide it into multiple sample segments containing the corresponding real remaining service life values, and generate a set of sample pairs.

[0008] S2: Perform time-frequency transformation and feature fusion processing on the sample segments in the sample pair set, convert the channel signals in each sample segment into time-frequency spectrograms and stack them along the channel dimension to generate a three-dimensional time-frequency feature tensor.

[0009] S3: Using the three-dimensional time-frequency feature tensor as input and the real remaining lifetime value in the sample pair set as the supervision signal, a deep neural network with a built-in Monte Carlo random dropout layer is called for probabilistic training. By optimizing the negative log-likelihood loss function, the network learns the distribution of the predicted values ​​and generates a trained probabilistic prediction network model.

[0010] S4: The time series data to be monitored collected from the device to be monitored is processed by signal feature extraction and tensor construction, which is converted into a predictable feature tensor consistent with the three-dimensional time-frequency feature tensor structure. The probabilistic prediction network model is called to perform multiple forward propagation sampling of the predictable feature tensor, and the mean and variance parameters of the multiple outputs are aggregated to generate the probability distribution prediction of the remaining lifetime and the corresponding confidence interval.

[0011] S5: Perform uncertainty quantification analysis on the probability distribution prediction of the remaining useful life and the corresponding confidence interval, calculate the separated cognitive uncertainty and accidental uncertainty measures, and generate a decision signal containing risk level and maintenance recommendations by combining preset threshold rules.

[0012] In one embodiment, S1 of the deep learning-based mechanical equipment operation status monitoring method provided by the present invention specifically includes the following steps:

[0013] S11: The synchronous raw signal acquired by the vibration sensor and temperature sensor deployed on the mechanical equipment is subjected to anti-aliasing filtering and noise reduction preprocessing. The low-pass filter is used to filter out components with frequencies higher than the Nyquist frequency, and the empirical mode decomposition method is used to remove power frequency noise and random pulse interference from the signal to generate a clean multi-channel time sequence signal.

[0014] S12: Perform sliding segmentation on the clean multi-channel time-series signal. A fixed-length time window slides forward to cover multiple complete working cycles of the key rotating parts of the equipment, generating a sequence of original sample segments containing continuous dynamic information.

[0015] S13: Perform residual lifetime backtracking calculation and annotation processing based on the historical full-lifetime degradation trajectory of the device for the time point corresponding to each original sample segment sequence. Based on the time interval between the time point corresponding to each original sample segment sequence and the time interval between the device's performance failure or malfunction, assign a real residual lifetime value to each sample segment and generate a set of sample pairs.

[0016] In one embodiment, step S2 of the deep learning-based mechanical equipment operation status monitoring method provided by the present invention specifically includes the following steps:

[0017] S21: Perform time-frequency analysis on the original signals of each channel contained in each sample segment of the sample pair set, and convert the one-dimensional time series of each channel into a two-dimensional time spectrum map reflecting the distribution of signal energy in the time and frequency dimensions through continuous wavelet transform, thereby generating the time spectrum map of each channel.

[0018] S22: Perform time axis alignment and scale normalization on the time spectrum of each channel, interpolate and calibrate the time axis of each spectrum according to the unified start time point and sampling interval, and scale the amplitude values ​​of each spectrum matrix to the same numerical range to generate a standardized time spectrum group.

[0019] S23: The two-dimensional spectrum matrices of each channel in the standardized time-frequency spectrum group are spliced ​​and stacked along the new channel dimension to form a three-dimensional data block that simultaneously contains time, frequency and channel information, generating a three-dimensional time-frequency feature tensor.

[0020] In one embodiment, step S3 of the deep learning-based mechanical equipment operation status monitoring method provided by the present invention specifically includes the following steps:

[0021] S31: Construct a feature encoder with convolutional layers and attention modules as the core to extract local features and global dependencies from the three-dimensional time-frequency feature tensor and generate high-level semantic feature vectors.

[0022] S32: A Monte Carlo random dropout layer is embedded in the fully connected layer of the feature encoder to randomly mask some neuron connections during the forward propagation of the high-level semantic feature vectors, so as to simulate the posterior distribution of network parameters and generate feature vector samples with randomness.

[0023] S33: Based on feature vector samples with randomness, regression and uncertainty quantification are performed through parallel mean output layer and variance output layer respectively. The same feature vector sample is simultaneously mapped to conditional mean prediction and conditional log-variance prediction of remaining lifetime. The end-to-end optimization and update of all network parameters is performed by minimizing the negative log-likelihood loss function to generate a trained probability prediction network model.

[0024] In one embodiment, the expression for the negative log-likelihood loss function of the deep learning-based mechanical equipment operation status monitoring method provided by the present invention is as follows:

[0025]

[0026] in, This represents the negative log-likelihood loss function. This represents the number of samples in a training batch. This represents the true remaining useful life value for the i-th sample. Let be the conditional mean of the remaining useful life predicted by the network for the i-th sample. Let be the conditional variance of the remaining useful life predicted by the network for the i-th sample.

[0027] In one embodiment, step S4 of the deep learning-based mechanical equipment operation status monitoring method provided by the present invention specifically includes the following steps:

[0028] S41: Perform sliding window segmentation and time-frequency transformation processing on the time-series data to be detected collected online from the device to be monitored, and convert it into a predictable feature tensor with the same size and structure as the three-dimensional time-frequency feature tensor;

[0029] S42: Call the trained probability prediction network model, and while keeping its internal Monte Carlo random dropout layer active, perform multiple independent forward propagation calculations on the feature tensor to be predicted. Each propagation outputs a set of corresponding mean and variance parameters, generating multiple sets of prediction parameters.

[0030] S43: Perform statistical analysis on all mean parameters in multiple sets of prediction parameters to determine the central trend, and calculate the distribution range of the mean parameters to assess the volatility of the prediction. Aggregate the predictions to generate the probability distribution of remaining useful life and the corresponding confidence intervals. The confidence intervals are used to characterize the uncertainty of the prediction.

[0031] In one embodiment, step S5 of the deep learning-based mechanical equipment operation status monitoring method provided by the present invention specifically includes the following steps:

[0032] S51: Perform analytical processing on the probability distribution prediction of remaining useful life. From the parameter set of multiple forward propagation outputs that constitute the probability distribution prediction, extract the mean sequence formed by all mean prediction values ​​and the variance sequence formed by all variance prediction values ​​respectively, and generate the original parameter sequence for uncertainty decomposition.

[0033] S52: Perform statistical measurement calculations on the original parameter sequence, calculate the standard deviation of the mean sequence as a measure of the model's cognitive uncertainty, and calculate the square root of the mean of the variance sequence as a measure of the data's random uncertainty.

[0034] S53: Based on the confidence interval corresponding to the probability distribution prediction, the measurement of cognitive uncertainty and the measurement of data random uncertainty, combined with the preset lifespan threshold and uncertainty threshold, multi-condition joint judgment processing is carried out. According to the judgment result, different risk levels are matched and corresponding preventive maintenance suggestions are generated to form a decision signal.

[0035] Secondly, the present invention provides a deep learning-based mechanical equipment operation status monitoring system, which is configured with the following modules:

[0036] The sample segmentation and annotation module is used to perform sliding window segmentation and annotation on the time-series data of the operating status synchronously collected by multi-source sensors deployed on mechanical equipment, dividing it into multiple sample segments containing corresponding real remaining service life values, and generating a set of sample pairs.

[0037] The time-frequency feature fusion module is used to perform time-frequency transformation and feature fusion processing on sample segments in the sample pair set. It converts the channel signals in each sample segment into time-frequency spectrograms and stacks them along the channel dimension to generate a three-dimensional time-frequency feature tensor.

[0038] The probabilistic prediction network training module is used to train a deep neural network with a built-in Monte Carlo random dropout layer by taking a three-dimensional time-frequency feature tensor as input and the real remaining lifetime value in the sample pair set as the supervision signal. By optimizing the negative log-likelihood loss function, the network learns the distribution of the predicted values ​​and generates a trained probabilistic prediction network model.

[0039] The probability distribution prediction module is used to extract signal features and construct tensors from the time series data collected from the device to be monitored, convert it into a predictable feature tensor with the same three-dimensional time-frequency feature tensor structure, and call the probability prediction network model to perform multiple forward propagation samplings on the predictable feature tensor, aggregate the mean and variance parameters of multiple outputs, and generate the probability distribution prediction of the remaining lifetime and the corresponding confidence interval.

[0040] The uncertainty quantification decision module is used to perform uncertainty quantification analysis on the probability distribution prediction of the remaining useful life and the corresponding confidence interval, calculate the separated cognitive uncertainty and accidental uncertainty measures, and generate a decision signal containing risk level and maintenance recommendations in combination with preset threshold rules.

[0041] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the above-mentioned deep learning-based mechanical equipment operation status monitoring methods.

[0042] Fourthly, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the aforementioned deep learning-based mechanical equipment operation status monitoring methods.

[0043] In summary, the deep learning-based mechanical equipment operation status monitoring method provided in this application optimizes the negative log-likelihood loss for probabilistic training of the network. This enables the model to learn not only the conditional mean of the remaining service life but also its conditional variance, thus transforming the traditional deterministic point estimation output into a complete probability distribution prediction. By calling the trained network to perform multiple forward propagation samplings on online data, probabilistic prediction results containing confidence intervals can be aggregated and generated. These confidence intervals intuitively quantify the overall uncertainty range of the prediction. By analyzing the parameter set of multiple prediction outputs, the overall uncertainty can be separated into "cognitive uncertainty," which represents the cognitive limitations of the model, and "accidental uncertainty," which represents the inherent noise of the data. This provides a clear insight into the source of prediction uncertainty. By comprehensively utilizing probability distributions, confidence intervals, and the decomposed uncertainty measure, combined with preset rules for decision-making, a leap from "hard decision" relying on a single threshold to "flexible early warning" based on risk probability can be achieved. This provides scientific and reliable decision support for predictive maintenance of equipment, significantly improving the level of refined management and risk response capabilities of operation and maintenance decisions.

[0044] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0045] Figure 1 A flowchart illustrating a deep learning-based method for monitoring the operating status of mechanical equipment, provided in an embodiment of this application;

[0046] Figure 2 This is a schematic diagram of the process for generating a trained probability prediction network model provided in an embodiment of this application;

[0047] Figure 3 This is a schematic diagram of a deep learning-based mechanical equipment operation status monitoring system provided in another embodiment of this application. Detailed Implementation

[0048] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0049] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0050] In one embodiment, such as Figure 1 As shown, a deep learning-based method for monitoring the operating status of mechanical equipment is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0051] S1: Perform sliding window segmentation and labeling on the time-series data of the operating status synchronously collected by multi-source sensors deployed on mechanical equipment, divide it into multiple sample segments containing the corresponding real remaining service life values, and generate a set of sample pairs.

[0052] Specifically, the multi-source sensors include vibration sensors, temperature sensors, pressure sensors, and speed sensors. All sensors are connected to the data acquisition card via an industrial bus, and a timestamp synchronization mechanism is used to achieve synchronous acquisition of multi-channel signals. The acquired raw data is stored in a specified format, and the data content includes the sensor ID, acquisition timestamp, and corresponding physical quantity value. The sliding window cutting parameters are determined by the system based on the typical degradation cycle of the mechanical equipment. The system sets the window length and step size of the sliding window. The selection of the window length must ensure that the window contains the key temporal characteristics of the equipment's operating status, and the selection of the step size must meet the balance between data utilization and computational efficiency, avoiding information redundancy between adjacent sample segments.

[0053] The segmentation process is implemented by the system through a relevant data processing library. The system independently performs sliding window segmentation on the time-series data of each sensor channel, obtaining several one-dimensional time-series segments of uniform length. The true remaining service life is labeled using a fault endpoint backtracking method. The system first determines the fault occurrence time through the equipment fault record system. For any sample segment, the system specifies its start and end times, and the true remaining service life value corresponding to the sample segment is the difference between the fault occurrence time and the end time of the sample segment. Fault determination is based on the definitions in relevant national standards. The system confirms the fault occurrence through dual verification of sensor data mutation thresholds and manual inspection results. The system constructs a sample pair with each time-series segment obtained through sliding window segmentation and its corresponding true remaining service life value. Each sample pair contains multi-dimensional time-series data and a scalar remaining service life value. All sample pairs form a sample pair set. The system divides the sample set into a training set, a validation set, and a test set according to a set ratio. The divided datasets are used for model training, model validation, and performance evaluation, respectively.

[0054] S2: Perform time-frequency transformation and feature fusion processing on the sample segments in the sample pair set, convert the channel signals in each sample segment into time-frequency spectrograms and stack them along the channel dimension to generate a three-dimensional time-frequency feature tensor.

[0055] Specifically, the system selects an appropriate time-frequency transformation algorithm based on the characteristics of different types of sensor signals. Vibration signals use short-time Fourier transform (SFT), while slowly varying signals such as temperature and pressure use wavelet packet transform (WFT). During the SFT, the system selects a specific window function, sets the window length and overlap rate, and performs the Fourier transform to generate a time-frequency spectrum, which includes two dimensions: frequency and the number of time-domain frames. During the WFT, the system selects a specific wavelet basis, performs a decomposition operation to extract the energy features of each sub-band, and then reconstructs the time-frequency spectrum. The dimensions of the reconstructed time-frequency spectrum are consistent with those of the time-frequency spectrum generated by the SFT.

[0056] Furthermore, the system independently performs a corresponding time-frequency transformation operation on each channel signal of the sample segment, obtaining a two-dimensional time-frequency spectrum corresponding to the number of channels. The pixel value of each time-frequency spectrum corresponds to the amplitude or energy of the signal at a specific time-frequency point. The system normalizes all two-dimensional time-frequency spectra by mapping the pixel values ​​to a specified interval using a normalization formula, thereby eliminating the dimensional differences between signals from different channels. After normalization, the system stacks all two-dimensional time-frequency spectra along the channel dimension to construct a three-dimensional time-frequency feature tensor. This three-dimensional time-frequency feature tensor simultaneously preserves the frequency characteristics, temporal characteristics, and channel correlation characteristics of the signal from multiple sources, thus meeting the input requirements of deep neural networks.

[0057] S3: Using the three-dimensional time-frequency feature tensor as input and the real remaining lifetime value in the sample pair set as the supervision signal, a deep neural network with a built-in Monte Carlo random dropout layer is called for probabilistic training. By optimizing the negative log-likelihood loss function, the network learns the distribution of predicted values ​​and generates a trained probabilistic prediction network model.

[0058] Specifically, the deep neural network designed for the system uses a convolutional neural network as its backbone network, combined with residual connections to alleviate the gradient vanishing problem. The deep neural network structure, from top to bottom, consists of an input layer, a group of convolutional layers, a Monte Carlo random dropout layer, a global average pooling layer, and an output layer. The input layer receives the three-dimensional time-frequency feature tensor. The convolutional layer group consists of multiple convolutional blocks, each containing a convolutional layer, a batch normalization layer, and an activation function. The convolutional layers use kernels of a specific size, the stride is determined according to a set rule, and edge padding uses a specified pattern. The convolutional layer group extracts local time-frequency features from the three-dimensional time-frequency feature tensor. The Monte Carlo random dropout layer is placed after each convolutional block and remains enabled during both training and testing phases, simulating model uncertainty by randomly dropping neurons. The global average pooling layer compresses the features extracted by the convolutional layer group into a one-dimensional vector. The output layer adopts a dual-output structure, outputting the mean and variance of the predicted remaining lifetime, respectively. Based on the Gaussian distribution assumption, the variance is constrained to a positive value using a specific function.

[0059] During model training, the system batches inputs the 3D time-frequency feature tensors of the training set into the deep neural network, using the actual remaining lifetime values ​​in the sample pair set as supervision signals. Preferably, the system can use a negative log-likelihood loss function to optimize model parameters, minimizing the loss function through an optimizer and iteratively executing the training process. To prevent overfitting, the system employs an L2 regularization strategy and monitors the loss value of the validation set. When the validation set loss value meets a set condition, an early stopping mechanism is triggered, saving the current optimal model parameters, including network weights, architecture configuration, and training hyperparameters. After model training, the system verifies the model's training effect using a test set, calculates the model's predictive performance metrics, and ensures that the model meets the accuracy requirements for industrial applications.

[0060] S4: The time series data to be monitored is collected from the device to be monitored and processed by signal feature extraction and tensor construction. It is converted into a predictable feature tensor with the same three-dimensional time-frequency feature tensor structure. The probabilistic prediction network model is called to perform multiple forward propagation samplings on the predictable feature tensor. The mean and variance parameters of the multiple outputs are aggregated to generate the probability distribution prediction of the remaining lifetime and the corresponding confidence interval.

[0061] Specifically, during the prediction phase, the system collects time-series data on the operating status of the monitored equipment. The data acquisition process is consistent with the data acquisition during the model training phase to ensure that the sampling frequency and number of channels of the collected data match the training data. The system preprocesses the collected raw data, performing detrending processing using a linear trend removal method, and then denoising processing using a wavelet thresholding method. This preprocessing eliminates the influence of environmental interference and sensor noise on the signal.

[0062] After preprocessing, the system performs time-frequency transformation on the signal according to the time-frequency transformation parameters from the model training phase, generating a two-dimensional time-frequency spectrum corresponding to the number of channels. After normalization of the two-dimensional time-frequency spectrum, a feature tensor to be predicted is constructed by stacking along the channel dimension. The structure of the feature tensor to be predicted is consistent with the three-dimensional time-frequency feature tensor from the training phase. The system inputs the feature tensor to be predicted into the trained probabilistic prediction network model and performs multiple independent forward propagation samplings. During each forward propagation, the Monte Carlo random dropout layer randomly drops neurons according to a set mechanism, obtaining multiple sets of prediction results. Each set of prediction results includes the prediction mean and prediction variance.

[0063] Furthermore, the system aggregates parameters from multiple sets of prediction results, calculates the average of all predicted means and the statistical value of all predicted variances, and simultaneously calculates the dispersion of the sampling results, incorporating it into the calculation of the total variance. Based on the aggregated mean and total variance, the system determines the probability distribution type of the remaining useful life. According to a preset confidence level, it calculates the corresponding confidence interval using the quantiles of the normal distribution. The calculation of the confidence interval is based on the quantile characteristics of the standard normal distribution, ensuring that the confidence interval can reflect the reliable range of the prediction results.

[0064] S5: Perform uncertainty quantification analysis on the probability distribution prediction of the remaining useful life and the corresponding confidence interval, calculate the separated cognitive uncertainty and accidental uncertainty measures, and generate a decision signal containing risk level and maintenance recommendations by combining preset threshold rules.

[0065] Specifically, the system employs a prediction variance decomposition method to separate and quantify cognitive uncertainty and random uncertainty. Random uncertainty stems from the inherent randomness of the equipment degradation process, and its quantification is achieved by calculating the statistical value of the prediction variance from multiple sampling. Cognitive uncertainty arises from the model's limitations in understanding unknown operating conditions, which can be reduced by increasing the data sample size; its quantification is achieved by calculating the dispersion of the prediction mean from multiple sampling. The system calculates the total measure of both types of uncertainty, and simultaneously calculates the proportion coefficients of cognitive uncertainty and random uncertainty. The contribution of these two types of uncertainty to the total uncertainty is analyzed using these proportion coefficients, and the results of this analysis provide a basis for subsequent decision-making priorities.

[0066] Furthermore, the system formulates preset threshold rules based on industrial equipment operation and maintenance costs, failure losses, and historical data statistics. These threshold rules include a threshold for the average remaining useful life, a threshold for total uncertainty, and a threshold for the proportion of perceived uncertainty. All thresholds can be dynamically adjusted through configuration files to adapt to the operation and maintenance needs of different types of machinery and equipment. The system combines the average remaining useful life, total uncertainty, and the proportion of perceived uncertainty to classify risk levels. Risk level classification is executed according to a set judgment logic, determining high, medium, and low risk levels by combining different conditions.

[0067] For different risk levels, the system generates corresponding maintenance recommendations. At high risk levels, the system generates recommendations based on triggering conditions, such as shutdown for maintenance, shortening the monitoring cycle, or activating backup monitoring schemes. At medium risk levels, the system generates recommendations to maintain the regular monitoring frequency and track the average remaining service life trend. At low risk levels, the system generates recommendations to maintain the existing operation and maintenance strategy. The system integrates relevant information into decision signals, which include the device ID, prediction timestamp, average remaining service life, confidence interval, two types of uncertainty values ​​and their proportions, risk level code, and maintenance recommendation text. The decision signals are output in a standardized data format.

[0068] In summary, the deep learning-based mechanical equipment operation status monitoring method provided in this application optimizes the negative log-likelihood loss for probabilistic training of the network. This enables the model to learn not only the conditional mean of the remaining service life but also its conditional variance, thus transforming the traditional deterministic point estimation output into a complete probability distribution prediction. By calling the trained network to perform multiple forward propagation samplings on online data, probabilistic prediction results containing confidence intervals can be aggregated and generated. These confidence intervals intuitively quantify the overall uncertainty range of the prediction. By analyzing the parameter set of multiple prediction outputs, the overall uncertainty can be separated into "cognitive uncertainty," which represents the cognitive limitations of the model, and "accidental uncertainty," which represents the inherent noise of the data. This provides a clear insight into the source of prediction uncertainty. By comprehensively utilizing probability distributions, confidence intervals, and the decomposed uncertainty measure, combined with preset rules for decision-making, a leap from "hard decision" relying on a single threshold to "flexible early warning" based on risk probability can be achieved. This provides scientific and reliable decision support for predictive maintenance of equipment, significantly improving the level of refined management and risk response capabilities of operation and maintenance decisions.

[0069] In one embodiment, S1 of the deep learning-based mechanical equipment operation status monitoring method provided by the present invention specifically includes the following steps:

[0070] S11: The synchronous raw signal acquired by the vibration sensor and temperature sensor deployed on the mechanical equipment is subjected to anti-aliasing filtering and noise reduction preprocessing. The low-pass filter is used to filter out components with frequencies higher than the Nyquist frequency, and the empirical mode decomposition method is used to remove power frequency noise and random pulse interference from the signal, generating a clean multi-channel timing signal.

[0071] Specifically, the system receives synchronous raw signals acquired simultaneously by vibration and temperature sensors deployed on the mechanical equipment. These synchronous raw signals contain continuous dynamic data on vibration and temperature during equipment operation. The synchronous acquisition of the sensors ensures the consistency of data timestamps through a time synchronization mechanism. To eliminate interference components introduced during signal transmission and acquisition, the system performs anti-aliasing filtering. Based on the Nyquist sampling theorem, the system uses a low-pass filter to filter the synchronous raw signals, removing high-frequency interference components above the Nyquist frequency. This prevents high-frequency components from folding into the effective frequency range, thus avoiding signal distortion and ensuring the compatibility of the signal frequency components with subsequent analysis requirements.

[0072] Furthermore, the system can employ Empirical Mode Decomposition (EMD) to denoise the signal. This method requires no pre-defined signal model and can adaptively decompose non-stationary and nonlinear signals. The system decomposes the filtered signal into several intrinsic mode functions (IMFs) and a residual component using EMD, with each IMF satisfying a specific stationarity condition. The system then filters the decomposed IMFs, identifying and removing components containing power frequency noise and random pulse interference, retaining only the effective IMFs that reflect the equipment's operating state. The filtered effective IMFs and the residual component are then reconstructed to generate a clean multi-channel time-series signal. This clean signal eliminates the influence of external interference and noise, truly preserving the essential characteristics of the equipment's operating state.

[0073] S12: Perform sliding segmentation on the clean multi-channel time-series signal. A fixed-length time window slides forward to cover multiple complete working cycles of the key rotating parts of the equipment, generating a sequence of original sample segments containing continuous dynamic information.

[0074] Specifically, after acquiring the clean multi-channel timing signal, the system converts the continuous multi-channel timing data into a sequence of sample segments containing continuous dynamic information of the equipment. Specifically, the system determines a fixed-length time window, with the core principle being to cover multiple complete working cycles of the equipment's key rotating components. The system determines the working cycle of the key rotating components by acquiring their operating parameters, and sets the time window length accordingly. This ensures that each time window can fully encompass the operating status information of the component across multiple working cycles, thereby comprehensively capturing the dynamic changes in the equipment's operation.

[0075] Furthermore, the system controls the time window to slide forward according to set rules, maintaining a constant window length during the sliding process. The step size of the window sliding is set based on the principle of balancing data continuity requirements and computational efficiency, ensuring that a reasonable degree of information overlap is maintained between adjacent sample segments and avoiding the loss of key dynamic information. The system performs the above sliding segmentation operation on the clean multi-channel time-series signal as a whole, sequentially extracting signal data within each time window in chronological order to form a continuous sequence of original sample segments. Each original sample segment contains synchronous data from both vibration and temperature channels and completely preserves the continuous operating status information of the equipment within the corresponding time interval, providing a structured signal carrier for subsequent labeling of the remaining service life.

[0076] S13: Perform residual lifetime backtracking calculation and annotation processing based on the historical full-lifetime degradation trajectory of the device for the time point corresponding to each original sample segment sequence. Based on the time interval between the time point corresponding to each original sample segment sequence and the time interval between the device's performance failure or malfunction, assign a real residual lifetime value to each sample segment and generate a set of sample pairs.

[0077] Specifically, the system accesses the device's historical full-lifecycle degradation trajectory database. This database stores operational data and time records for similar devices throughout their complete lifecycle, from initial operation to performance failure or malfunction, including the device's state characteristics at each stage and corresponding time node information. For each original sample segment sequence, the system extracts its corresponding time point information, which is the end timestamp of the sample segment sequence. Based on this timestamp, the system locates the device within its historical full-lifecycle degradation trajectory, determining the specific position of that time point within the device's entire lifecycle.

[0078] Furthermore, the system performs a backtesting calculation of remaining lifetime, determining the time interval from the point in time until the equipment experiences performance failure or malfunction. This time interval is the actual remaining lifetime value corresponding to the original sample segment sequence. The system assigns a corresponding actual remaining lifetime value to each original sample segment sequence and completes the annotation process. During annotation, it ensures a one-to-one correspondence between the timestamp of each sample segment and the actual remaining lifetime value, avoiding data misalignment. After annotation, the system constructs a sample pair with each original sample segment and its corresponding actual remaining lifetime value. Each sample pair contains multi-channel time-series signal data and a scalar-form remaining lifetime annotation value. All sample pairs are organized according to time order and data category to form a sample pair set. This sample pair set covers the status data and corresponding remaining lifetime information of the equipment at different operating stages.

[0079] In one embodiment, step S2 of the deep learning-based mechanical equipment operation status monitoring method provided by the present invention specifically includes the following steps:

[0080] S21: Perform time-frequency analysis on the original signals of each channel contained in each sample segment of the sample pair set. Convert the one-dimensional time series of each channel into a two-dimensional time spectrum map reflecting the distribution of signal energy in the time and frequency dimensions through continuous wavelet transform, and generate the time spectrum map of each channel.

[0081] Specifically, after acquiring each sample segment from the sample pair set, the system extracts the original signals of all channels contained in the sample segment. These original signals are one-dimensional time series, covering different types of equipment operating status data such as vibration and temperature. To fully explore the hidden time-frequency domain correlation features in the signals, the system independently performs time-frequency analysis processing on the one-dimensional time series of each channel, using continuous wavelet transform as the time-frequency analysis algorithm. Continuous wavelet transform achieves adaptive analysis of non-stationary signals by shifting the mother wavelet on the time axis and scaling it on the frequency axis, without the need for preset fixed time or frequency windows, and can simultaneously take into account the time domain resolution and frequency domain resolution of the signal.

[0082] Furthermore, the system selects an appropriate mother wavelet based on the characteristics of each channel signal and performs a global transformation operation on the one-dimensional time series by setting the translation step size and scaling parameters. During the transformation, the system calculates the inner product of the mother wavelet and the signal at different translation positions and different scaling scales, generating a two-dimensional time-frequency spectrum. The horizontal dimension of this two-dimensional time-frequency spectrum corresponds to the time axis, and the vertical dimension corresponds to the frequency axis. The value of each pixel in the graph reflects the energy distribution intensity of the signal at the corresponding time-frequency position, completely preserving the dynamic changes of the original signal in the time dimension and the component distribution information in the frequency dimension.

[0083] S22: Perform time axis alignment and scale normalization on the time spectrum of each channel, interpolate and calibrate the time axis of each spectrum according to the unified start time point and sampling interval, and scale the amplitude values ​​of each spectrum matrix to the same numerical range to generate a standardized time spectrum group.

[0084] Specifically, due to differences in installation location and signal transmission path between multi-channel sensors, the time axis of the time spectrum of each channel may have slight deviations. Furthermore, the different physical dimensions of the signals from different channels lead to variations in the amplitude range of the time spectrum matrix. These factors can affect the effectiveness of subsequent feature fusion. Therefore, the system performs time axis alignment and scale normalization on the time spectrum of all channels, determining a unified starting time point and sampling interval. This starting time point is based on the first data acquisition timestamp of the sample segment, and the sampling interval remains consistent with the acquisition interval of the original signal. The system performs time axis interpolation calibration on the time spectrum of each channel, supplementing or adjusting the sampling points on the time axis of the time spectrum through interpolation algorithms to ensure that the time axis scale and the number of sampling points of all time spectrums are completely consistent, achieving precise matching of the time-frequency characteristics of each channel in the time dimension.

[0085] Preferably, the system can employ a unified normalization algorithm to scale the amplitude values ​​of the time-frequency matrix of each channel, mapping all amplitude values ​​to the same numerical range. This eliminates the influence of differences in dimensions and amplitude ranges between different channels, ensuring that the time-frequency characteristics of each channel have a unified numerical scale standard. After time axis alignment and scale normalization, the system obtains a standardized time-frequency spectrum group. All time-frequency spectra in this group have a unified time axis and consistent amplitude range, meeting the conditions for direct feature fusion.

[0086] S23: The two-dimensional spectrum matrices of each channel in the standardized time-frequency spectrum group are spliced ​​and stacked along the new channel dimension to form a three-dimensional data block that simultaneously contains time, frequency and channel information, generating a three-dimensional time-frequency feature tensor.

[0087] Specifically, the system fuses multiple two-dimensional standardized time-frequency spectrograms into structured feature data containing multi-dimensional information, clearly defining the dimensional composition of the three-dimensional time-frequency feature tensor. The first dimension corresponds to time information, the second to frequency information, and the third to channel information. Following a preset channel order, the system sequentially splices and stacks the two-dimensional spectrum matrices of each channel in the standardized time-frequency spectrogram group along the third dimension (the new channel dimension). During splicing, the system maintains the time and frequency dimensions of each two-dimensional spectrum matrix unchanged, only expanding the channel dimension to ensure that the time-frequency features of each channel do not interfere with each other and are completely preserved. After stacking, a three-dimensional data block is formed, which is the three-dimensional time-frequency feature tensor. This tensor integrates the dynamic features of the time dimension, the component features of the frequency dimension, and the multi-source correlation features of the channel dimension, achieving deep fusion of different types and different channel operating status data.

[0088] In one embodiment, step S3 of the deep learning-based mechanical equipment operation status monitoring method provided by the present invention specifically includes the following steps:

[0089] S31: Construct a feature encoder with convolutional layers and attention modules as the core to extract local features and global dependencies from the three-dimensional time-frequency feature tensor and generate high-level semantic feature vectors.

[0090] Specifically, the system constructs a feature encoder centered on convolutional layers and an attention module. This encoder comprises a sequence of convolutional layers and an attention mechanism layer, used to extract local features and global dependencies from the 3D time-frequency feature tensor in a hierarchical manner. The convolutional layer sequence consists of several levels of convolutional operations. Each convolutional layer performs cross-correlation operations with the 3D time-frequency feature tensor through a convolutional kernel to extract local time-frequency domain correlated features. The operation process follows the formula:

[0091]

[0092] in, This represents a local feature map of the convolution output. Represents the convolution kernel parameter matrix. This represents the input three-dimensional time-frequency feature tensor. This represents the bias vector of the convolutional layer. After each convolutional operation, the system performs batch normalization and non-linear activation operations. Batch normalization standardizes the mean and variance of the feature map, while activation operations perform non-linear transformations of the features. The attention mechanism layer aggregates global information by calculating the association weights at different locations in the feature map. The weight calculation follows the formula:

[0093]

[0094] in, This represents the attention weight at the j-th feature position. This represents the score value at the j-th feature position. This represents the total number of locations in the feature map. The system multiplies the local feature map with the attention weights to obtain a feature map that incorporates global dependencies. Then, it compresses the feature map into a one-dimensional vector through global average pooling. This vector is the high-level semantic feature vector, which integrates local details and global correlation information.

[0095] S32: A Monte Carlo random dropout layer is embedded in the fully connected layer of the feature encoder to randomly mask some neuron connections during the forward propagation of the high-level semantic feature vectors, so as to simulate the posterior distribution of network parameters and generate feature vector samples with randomness.

[0096] Specifically, the system embeds Monte Carlo dropout layers between the fully connected layers of the feature encoder. The fully connected layers map high-level semantic feature vectors to high-dimensional dense feature representations. The Monte Carlo dropout layers and fully connected layers form a cascaded transmission path, maintaining activation during both training and testing phases. The core operations of this layer follow the Bernoulli distribution sampling rule, and the survival state of neurons is determined by the formula... Confirmed, among which The survival flag represents the i-th neuron (a value of 1 indicates that the neuron is alive, and a value of 0 indicates that the neuron is blocked). Indicates the survival probability of a neuron. This represents the Bernoulli distribution. During the forward propagation of high-level semantic feature vectors through fully connected layers, the system independently samples each neuron according to the above formula, randomly shielding some neurons from connections to subsequent layers. These shielded neurons do not participate in feature calculation or signal transmission. In each forward propagation, the shielding combination of neurons changes randomly. Through multiple independent forward propagations, multiple sets of vectors with differing feature representations are generated. These vectors are the random feature vector samples, whose randomness stems from the random sampling of neuron survival states. This randomness can simulate the posterior distribution of network parameters, providing feature inputs containing uncertainty information for subsequent probability prediction.

[0097] S33: Based on feature vector samples with randomness, regression and uncertainty quantification are performed through parallel mean output layer and variance output layer respectively. The same feature vector sample is simultaneously mapped to conditional mean prediction and conditional log-variance prediction of remaining lifetime. The end-to-end optimization and update of all network parameters is performed by minimizing the negative log-likelihood loss function to generate a trained probability prediction network model.

[0098] Specifically, the system constructs parallel mean output layers and variance output layers, forming a complete probabilistic prediction network architecture with the feature encoder and Monte Carlo random dropout layer. The two output layers share the parameters of the preceding network and synchronously process feature vector samples with randomness. The mean output layer maps the feature vector samples to conditional mean predictions of remaining lifetime through a linear transformation, following the formula:

[0099]

[0100] in, Let represent the predicted conditional mean of the i-th sample. This represents the weight matrix of the mean output layer. This represents the i-th random feature vector sample. This represents the bias vector of the mean output layer. The variance output layer outputs the conditional logarithmic variance through a linear transformation, followed by an exponential transformation to obtain the conditional variance, ensuring that the variance is positive. The transformation process follows the formula below:

[0101]

[0102] in, Let represent the conditional variance of the i-th sample. This represents the weight matrix of the variance output layer. This represents the bias vector of the variance output layer. This represents an exponential function. During the model training phase, the system uses the negative log-likelihood loss function to calculate the prediction bias. The expression for the loss function is:

[0103]

[0104] in, This represents the negative log-likelihood loss function. This represents the number of samples in a training batch. This represents the true remaining useful life value for the i-th sample. Let be the conditional mean of the remaining useful life predicted by the network for the i-th sample. Let be the conditional variance of the remaining lifetime predicted by the network for the i-th sample. The system minimizes the loss function through an optimizer, iteratively updates all network parameters, monitors the loss value on the validation set, stops training when the convergence condition is met, saves the optimal model parameters, and generates a trained probability prediction network model.

[0105] In one embodiment, step S4 of the deep learning-based mechanical equipment operation status monitoring method provided by the present invention specifically includes the following steps:

[0106] S41: Perform sliding window segmentation and time-frequency transformation processing on the time-series data to be detected collected online from the device to be monitored, and convert it into a predictable feature tensor with the same size and structure as the three-dimensional time-frequency feature tensor.

[0107] Specifically, the system collects time-series data to be detected online from the device under monitoring. The acquisition process employs the same sensor configuration, sampling mechanism, and synchronization strategy as the model training phase, ensuring that the sampling frequency and number of channels of the time-series data to be detected match the training data, thus providing a foundation for feature consistency. After data acquisition, the system performs sliding window segmentation on the time-series data to be detected according to the sliding window segmentation rules set during model training. The window length and step size of the sliding window are completely consistent with the parameters used in the training phase, ensuring that the length of the segmented signal fragments is consistent with the length of the training sample fragments.

[0108] After segmentation, the system selects the corresponding time-frequency transformation algorithm based on the signal type of each channel. Vibration signals undergo short-time Fourier transform, while slowly varying signals such as temperature undergo wavelet packet transform. The window function, wavelet basis, and number of decomposition levels for the time-frequency transform are maintained consistent with those used in the training phase, converting the one-dimensional time-series signal of each channel into a two-dimensional time-spectrum graph. The system performs normalization processing on all two-dimensional time-spectrum graphs, following the formula:

[0109]

[0110] in, This represents the normalized pixel value. This represents the original pixel values ​​of the time-frequency spectrogram. This represents the mean of the pixel values ​​in the spectrogram during the training set. This represents the standard deviation of pixel values ​​in the time-frequency spectrogram of the training set. After normalization, the system stacks the two-dimensional time-frequency spectrograms of each channel along the channel dimension to construct a feature tensor to be predicted. The size and dimensionality of this tensor are completely consistent with the three-dimensional time-frequency feature tensor in the training stage, satisfying the input format requirements of the probabilistic prediction network model.

[0111] S42: Call the trained probabilistic prediction network model, and while keeping its internal Monte Carlo random dropout layer active, perform multiple independent forward propagation calculations on the feature tensor to be predicted. Each propagation outputs a set of corresponding mean and variance parameters, generating multiple sets of prediction parameters.

[0112] Specifically, the system calls the trained probabilistic prediction network model and explicitly configures the activation state of the Monte Carlo random dropout layer during model loading to ensure that this layer remains active during the prediction phase, consistent with the operating mechanism during the training phase. The system inputs the feature tensor to be predicted into the probabilistic prediction network model and initiates multiple independent forward propagation calculations. Each forward propagation is an independent computation process, unaffected by the results of other propagations. During forward propagation, the Monte Carlo random dropout layer randomly determines the survival state of neurons according to a preset probability distribution. The combinations of neuron survival in different propagations are different, simulating the differences in the posterior distribution of model parameters through this mechanism. After each forward propagation, the model outputs a set of corresponding prediction parameters, including the conditional mean and conditional variance of the remaining lifetime. The output results follow the formula:

[0113]

[0114] in, This represents the mean prediction value of the k-th forward propagation. This represents the prediction variance of the k-th forward propagation. This represents the feature tensor to be predicted. These represent the fixed parameters of the probabilistic prediction network model. This represents the survival status of neurons in the Monte Carlo dropout layer during the k-th forward propagation. The system repeatedly performs forward propagation calculations until the preset number of propagations is reached. The predicted mean and variance parameters of all propagation outputs are then compiled and summarized to generate multiple sets of predicted parameters.

[0115] S43: Perform statistical analysis on all mean parameters in multiple sets of prediction parameters to determine the central trend, and calculate the distribution range of the mean parameters to assess the volatility of the prediction. Aggregate the predictions to generate the probability distribution of remaining useful life and the corresponding confidence intervals. The confidence intervals are used to characterize the uncertainty of the prediction.

[0116] Specifically, the system performs statistical analysis on all mean parameters in multiple sets of prediction parameters to determine the central trend of the remaining useful life prediction. The central trend is calculated using the mean statistical formula, which is expressed as follows:

[0117]

[0118] in, This represents the central tendency mean after aggregation. This indicates the total number of forward propagations. This represents the predicted mean of the k-th forward propagation. Simultaneously, the system calculates the distribution range of the mean parameter to assess prediction volatility; volatility quantification follows the formula:

[0119]

[0120] in, This represents the variance of the mean parameter. The system aggregates the central trend mean, variance of variance, and variance parameters from the prediction parameter set to generate a probability distribution of remaining useful life. This distribution comprehensively reflects the core estimate and uncertainty range of the prediction result. Based on the generated probability distribution, the system calculates the corresponding confidence interval according to a preset confidence level. The formula for calculating the confidence interval is:

[0121]

[0122] in, This represents the quantile at the corresponding confidence level under a standard normal distribution. The aggregated variance represents the total variance, and the confidence interval visually represents the uncertainty boundary of the prediction result.

[0123] In one embodiment, step S5 of the deep learning-based mechanical equipment operation status monitoring method provided by the present invention specifically includes the following steps:

[0124] S51: Perform analytical processing on the probability distribution prediction of remaining useful life. From the parameter set of multiple forward propagation outputs that constitute the probability distribution prediction, extract the mean sequence formed by all mean prediction values ​​and the variance sequence formed by all variance prediction values ​​respectively, and generate the original parameter sequence for uncertainty decomposition.

[0125] Specifically, after obtaining the probability distribution prediction results of the remaining useful life, the system separates the mean and variance related data from the parameter set output by multiple forward propagations, providing raw materials for subsequent uncertainty decomposition. The system first locates a dataset storing multiple sets of prediction parameters. This dataset contains the predicted mean and predicted variance output by all independent forward propagation processes, with each set of parameters uniquely associated with its corresponding propagation number. The system extracts all predicted means from the parameter set one by one according to the chronological or sequential order of the forward propagation, arranging these predicted mean values ​​in the extraction order to form a one-dimensional mean sequence. Simultaneously, the system extracts all predicted variances from the parameter set using the same arrangement rule, organizing these predicted variance values ​​sequentially to form a one-dimensional variance sequence. The mean sequence and variance sequence together constitute the original parameter sequence used for uncertainty decomposition, where the mean sequence reflects the central trend fluctuation of multiple prediction results, and the variance sequence reflects the model's uncertainty estimate of its own output in each prediction.

[0126] S52: Perform statistical measurement calculations on the original parameter sequence, calculate the standard deviation of the mean sequence as a measure of the model's cognitive uncertainty, and calculate the square root of the mean of the variance sequence as a measure of the data's random uncertainty.

[0127] Specifically, the system performs statistical measurement calculations based on the original parameter sequence. For the measurement of cognitive uncertainty, the system calculates the standard deviation of the mean sequence. This standard deviation reflects the dispersion of multiple predicted means, and its calculation follows the formula:

[0128]

[0129] in, This represents a measure of cognitive uncertainty. This represents the function for calculating standard deviation. Let K represent the mean sequence, and K represent the total number of forward propagations. Cognitive uncertainty stems from the model's limited understanding of unknown operating conditions, and its magnitude is related to the coverage of the model's training data. For measuring random uncertainty, the system first calculates the mean of the variance sequence, then takes the square root of that mean, following the formula:

[0130]

[0131] in, This represents a measure of random uncertainty. This represents the function for calculating the average value. The variance sequence represents the random uncertainty stemming from the inherent randomness of the equipment degradation process, unaffected by model optimization or data supplementation. The system quantifies the two types of uncertainty using the two formulas described above.

[0132] S53: Based on the confidence interval corresponding to the probability distribution prediction, the measurement of cognitive uncertainty and the measurement of data random uncertainty, combined with the preset lifespan threshold and uncertainty threshold, multi-condition joint judgment processing is carried out. According to the judgment result, different risk levels are matched and corresponding preventive maintenance suggestions are generated to form a decision signal.

[0133] Specifically, the system integrates the confidence interval, cognitive uncertainty metric, and random uncertainty metric corresponding to the probability distribution prediction, and combines them with preset lifetime and uncertainty thresholds to perform multi-condition joint judgment processing. The preset thresholds are formulated by the system based on industrial equipment operation and maintenance costs, failure loss statistics, and industry operation and maintenance standards. The lifetime threshold is used to divide the risk range of the remaining service life, and the uncertainty threshold is used to determine the reliability of the prediction results. All thresholds are stored in a configuration file and support dynamic adjustment according to equipment type. During the joint judgment process, the system first compares the mean predicted value of the remaining service life with the lifetime threshold, and simultaneously checks whether the coverage of the confidence interval meets the preset requirements. Secondly, it compares the cognitive uncertainty metric and the random uncertainty metric with their corresponding uncertainty thresholds to assess the impact of the two types of uncertainty on the prediction results.

[0134] Furthermore, based on the results of multi-condition comparison, the system matches corresponding risk levels according to preset rules. Risk levels are divided into high, medium, and low, and the matching logic is determined based on the combination relationship of each condition. For different risk levels, the system generates corresponding preventative maintenance recommendations: high-risk levels correspond to emergency maintenance measures, medium-risk levels to enhanced monitoring and planned maintenance recommendations, and low-risk levels to routine operation and maintenance strategies. The system integrates the mean predicted value of remaining service life, confidence interval, two types of uncertainty measures, risk level, and maintenance recommendations into structured data, encapsulates it in a standardized format to form a decision signal, and this decision signal supports seamless integration with industrial operation and maintenance management systems, providing operation and maintenance personnel with comprehensive decision-making basis and realizing the transformation from prediction results to operation and maintenance actions.

[0135] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0136] Based on the same inventive concept, this application also provides a deep learning-based mechanical equipment operation status monitoring system for implementing the aforementioned deep learning-based mechanical equipment operation status monitoring method. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more deep learning-based mechanical equipment operation status monitoring system embodiments provided below can be found in the limitations of the deep learning-based mechanical equipment operation status monitoring method described above, and will not be repeated here.

[0137] Preferably, such as Figure 3 As shown, the present invention provides a deep learning-based mechanical equipment operation status monitoring system 600, which is configured with the following modules:

[0138] The sample cutting and labeling module 610 is used to perform sliding window cutting and labeling on the time-series data of the operating status synchronously collected by multi-source sensors deployed on mechanical equipment, dividing it into multiple sample segments containing corresponding real remaining service life values, and generating a set of sample pairs.

[0139] The time-frequency feature fusion module 620 is used to perform time-frequency transformation and feature fusion processing on sample segments in the sample pair set. It converts the channel signals in each sample segment into time-frequency spectrograms and stacks them along the channel dimension to generate a three-dimensional time-frequency feature tensor.

[0140] The probability prediction network training module 630 is used to train a deep neural network with a built-in Monte Carlo random dropout layer by taking a three-dimensional time-frequency feature tensor as input and the real remaining lifetime value in the sample pair set as the supervision signal. By optimizing the negative log-likelihood loss function, the network learns the distribution of the predicted values ​​and generates a trained probability prediction network model.

[0141] The probability distribution prediction module 640 is used to extract signal features and construct tensors from the time series data to be monitored collected from the device to be monitored, convert it into a predictable feature tensor with the same three-dimensional time-frequency feature tensor structure, and call the probability prediction network model to perform multiple forward propagation sampling of the predictable feature tensor, aggregate the mean and variance parameters of multiple outputs, and generate the probability distribution prediction of the remaining lifetime and the corresponding confidence interval.

[0142] The uncertainty quantification decision module 650 is used to perform uncertainty quantification analysis on the probability distribution prediction of the remaining useful life and the corresponding confidence interval, calculate the separated cognitive uncertainty and accidental uncertainty measures, and generate a decision signal containing risk level and maintenance recommendations in combination with preset threshold rules.

[0143] Preferably, the sample cutting and annotation module 610 provided in this application is configured with the following units:

[0144] The signal preprocessing and denoising unit is used to perform anti-aliasing filtering and denoising preprocessing on the synchronous raw signals collected synchronously by vibration sensors and temperature sensors deployed on mechanical equipment. It uses a low-pass filter to filter out components with frequencies higher than the Nyquist frequency and uses empirical mode decomposition to remove power frequency noise and random pulse interference from the signal, generating a clean multi-channel time sequence signal.

[0145] The signal sliding segmentation unit is used to slide and segment clean multi-channel time-series signals. A fixed-length time window slides forward to cover multiple complete working cycles of key rotating parts of the equipment, generating a sequence of original sample segments containing continuous dynamic information.

[0146] The Remaining Life Labeling Unit is used to perform backtracking calculation and labeling of the remaining life based on the historical full-life degradation trajectory of the device for the time point corresponding to each original sample segment sequence. Based on the time interval between the time point corresponding to each original sample segment sequence and the time interval between the device's performance failure or malfunction, a real remaining life value is assigned to each sample segment, generating a set of sample pairs.

[0147] Preferably, the time-frequency feature fusion module 620 provided in this application is configured with the following units:

[0148] The channel time-frequency analysis unit is used to perform time-frequency analysis processing on the original signals of each channel contained in each sample segment of the sample pair set. It converts the one-dimensional time series of each channel into a two-dimensional time spectrum map reflecting the distribution of signal energy in the time and frequency dimensions through continuous wavelet transform, and generates the time spectrum map of each channel.

[0149] The spectrum normalization unit is used to perform time axis alignment and scale normalization on the time spectrum of each channel. It interpolates and calibrates the time axis of each spectrum according to a unified start time point and sampling interval, and scales the amplitude values ​​of each spectrum matrix to the same numerical range to generate a standardized time spectrum group.

[0150] The feature tensor construction unit is used to stitch and stack the two-dimensional spectrum matrices of each channel in the standardized time-frequency spectrum group along the new channel dimension to form a three-dimensional data block that simultaneously contains time, frequency and channel information, and generate a three-dimensional time-frequency feature tensor.

[0151] Preferably, the probability prediction network training module 630 provided in this application is configured with the following units:

[0152] The semantic feature extraction unit is used to construct a feature encoder with convolutional layers and attention modules as its core, and to extract local features and global dependencies from the three-dimensional time-frequency feature tensor to generate high-level semantic feature vectors.

[0153] The random feature sampling unit is used to embed a Monte Carlo random drop-out layer in the fully connected layer of the feature encoder. It randomly masks some neuron connections during the forward propagation of the high-level semantic feature vector to simulate the posterior distribution of network parameters and generate feature vector samples with randomness.

[0154] The probabilistic model training unit is used to perform regression and uncertainty quantification processing on feature vector samples with randomness through parallel mean output layer and variance output layer, respectively. The same feature vector sample is simultaneously mapped to conditional mean prediction and conditional log-variance prediction of remaining useful life. The network parameters are optimized and updated end-to-end by minimizing the negative log-likelihood loss function to generate a trained probabilistic prediction network model.

[0155] Preferably, the probability distribution prediction module 640 provided in this application is configured with the following units:

[0156] The predictable tensor construction unit is used to perform sliding window segmentation and time-frequency transformation processing on the time-series data to be detected collected online from the device to be monitored, and convert it into a predictable feature tensor with the same size and structure as the three-dimensional time-frequency feature tensor.

[0157] The multi-round sampling unit of the model is used to call the trained probability prediction network model. While keeping the activation of its internal Monte Carlo random dropout layer, it performs multiple independent forward propagation calculations on the feature tensor to be predicted. Each propagation outputs a set of corresponding mean and variance parameters, generating multiple sets of prediction parameters.

[0158] The prediction result aggregation unit is used to perform statistical analysis on all mean parameters in multiple sets of prediction parameters to determine the central trend, and to calculate the distribution range of the mean parameters to assess the volatility of the prediction. The aggregation generates a probability distribution prediction of the remaining useful life and the corresponding confidence interval, which is used to characterize the uncertainty of the prediction.

[0159] Preferably, the uncertainty quantification decision module 650 provided in this application is configured with the following units:

[0160] The parameter sequence extraction unit is used to analyze the probability distribution prediction of the remaining useful life. From the parameter set of multiple forward propagation outputs that constitute the probability distribution prediction, it extracts the mean sequence formed by all mean prediction values ​​and the variance sequence formed by all variance prediction values, respectively, to generate the original parameter sequence for uncertainty decomposition.

[0161] The uncertainty decomposition unit is used to perform statistical measurement calculations on the original parameter sequence, calculate the standard deviation of the mean sequence as a measure of the model's cognitive uncertainty, and calculate the square root of the mean of the variance sequence as a measure of the data's random uncertainty.

[0162] The risk assessment and decision-making unit is used to predict the corresponding confidence interval based on the probability distribution, measure the cognitive uncertainty and the random uncertainty of the data, and perform multi-condition joint judgment processing in combination with the preset life threshold and uncertainty threshold. Based on the judgment results, different risk levels are matched and corresponding preventive maintenance suggestions are generated to form a decision signal.

[0163] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described deep learning-based mechanical equipment operation status monitoring method.

[0164] In one embodiment, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described deep learning-based mechanical equipment operation status monitoring method.

[0165] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0166] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0167] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for monitoring the operating status of mechanical equipment based on deep learning, characterized in that, Includes the following steps: S1: Perform sliding window segmentation and labeling on the time-series data of the operating status synchronously collected by multi-source sensors deployed on mechanical equipment, divide it into multiple sample segments containing the corresponding real remaining service life values, and generate a set of sample pairs. S2: Perform time-frequency transformation and feature fusion processing on the sample segments in the sample pair set, convert the channel signals in each sample segment into time-frequency spectrograms and stack them along the channel dimension to generate a three-dimensional time-frequency feature tensor. S3: Using the three-dimensional time-frequency feature tensor as input and the real remaining lifetime value in the sample pair set as the supervision signal, the deep neural network with the built-in Monte Carlo random dropout layer is called for probabilistic training. By optimizing the negative log-likelihood loss function, the network learns the distribution of the predicted values ​​and generates a trained probabilistic prediction network model. S4: The time series data to be monitored collected from the device to be monitored is processed by signal feature extraction and tensor construction, and converted into a predictable feature tensor consistent with the three-dimensional time-frequency feature tensor structure. The probability prediction network model is called to perform multiple forward propagation sampling on the predictable feature tensor, and the mean and variance parameters of the multiple outputs are aggregated to generate the probability distribution prediction of the remaining lifetime and the corresponding confidence interval. S5: Perform uncertainty quantification analysis on the probability distribution prediction of the remaining service life and the corresponding confidence interval, calculate the separated cognitive uncertainty and accidental uncertainty measures, and generate a decision signal containing risk level and maintenance recommendations in combination with preset threshold rules.

2. The method according to claim 1, characterized in that, S1 includes: S11: The synchronous raw signal acquired by the vibration sensor and temperature sensor deployed on the mechanical equipment is subjected to anti-aliasing filtering and noise reduction preprocessing. The low-pass filter is used to filter out components with frequencies higher than the Nyquist frequency, and the empirical mode decomposition method is used to remove power frequency noise and random pulse interference from the signal to generate a clean multi-channel time sequence signal. S12: The clean multi-channel time-series signal is subjected to sliding segmentation. A fixed-length time window slides forward to cover multiple complete working cycles of the key rotating parts of the equipment, generating a sequence of original sample segments containing continuous dynamic information. S13: Perform residual lifetime backtracking calculation and labeling processing on the time point corresponding to each of the original sample fragment sequences based on the historical full-lifetime degradation trajectory of the device. Based on the time interval between the time point corresponding to each of the original sample fragment sequences and the time interval between the device's performance failure or malfunction, assign a real residual lifetime value to each sample fragment to generate a set of sample pairs.

3. The method according to claim 1, characterized in that, S2 includes: S21: Perform time-frequency analysis on the original signals of each channel contained in each sample segment of the sample pair set, and convert the one-dimensional time series of each channel into a two-dimensional time spectrum map reflecting the distribution of signal energy in the time and frequency dimensions through continuous wavelet transform, thereby generating the time spectrum map of each channel. S22: Perform time axis alignment and scale normalization on the time spectrum diagrams of each channel, interpolate and calibrate the time axis of each spectrum diagram according to the unified start time point and sampling interval, and scale the amplitude values ​​of each spectrum matrix to the same numerical range to generate a standardized time spectrum diagram group. S23: The two-dimensional spectrum matrices of each channel in the standardized time-frequency spectrum group are spliced ​​and stacked along the new channel dimension to form a three-dimensional data block that simultaneously contains time, frequency and channel information, and a three-dimensional time-frequency feature tensor is generated.

4. The method according to claim 1, characterized in that, S3 includes: S31: Construct a feature encoder with convolutional layers and attention modules as the core, extract local features and global dependencies from the three-dimensional time-frequency feature tensor, and generate high-level semantic feature vectors. S32: A Monte Carlo random dropout layer is embedded in the fully connected layer of the feature encoder to randomly shield some neuron connections during the forward propagation of the high-level semantic feature vector, so as to simulate the posterior distribution of network parameters and generate feature vector samples with randomness. S33: Based on the random feature vector samples, regression and uncertainty quantification are performed through parallel mean output layer and variance output layer respectively. The same feature vector sample is simultaneously mapped to conditional mean prediction and conditional log-variance prediction of remaining lifetime. The end-to-end optimization and update of all network parameters is performed by minimizing the negative log-likelihood loss function to generate the trained probability prediction network model.

5. The method according to claim 4, characterized in that, The expression for the negative log-likelihood loss function is: in, This represents the negative log-likelihood loss function. This represents the number of samples in a training batch. This represents the true remaining useful life value for the i-th sample. Let be the conditional mean of the remaining useful life predicted by the network for the i-th sample. Let be the conditional variance of the remaining useful life predicted by the network for the i-th sample.

6. The method according to claim 1, characterized in that, S4 includes: S41: Perform sliding window segmentation and time-frequency transformation processing on the time-series data to be detected collected online from the device to be monitored, and convert it into a predictable feature tensor with the same size and structure as the three-dimensional time-frequency feature tensor; S42: Call the trained probability prediction network model, and while keeping its internal Monte Carlo random dropout layer active, perform multiple independent forward propagation calculations on the feature tensor to be predicted. Each propagation outputs a set of corresponding mean and variance parameters, generating multiple sets of prediction parameters. S43: Perform statistical analysis on all mean parameters in the multiple sets of prediction parameters to determine the central trend, and calculate the distribution range of the mean parameters to assess the volatility of the prediction. Aggregate the predictions to generate the probability distribution of the remaining useful life and the corresponding confidence intervals. The confidence intervals are used to characterize the uncertainty of the prediction.

7. The method according to any one of claims 1-6, characterized in that, S5 includes: S51: The probability distribution prediction of the remaining useful life is analyzed and processed. From the parameter set of multiple forward propagation outputs that constitute the probability distribution prediction, the mean sequence formed by all mean prediction values ​​and the variance sequence formed by all variance prediction values ​​are extracted respectively to generate the original parameter sequence for uncertainty decomposition. S52: Perform statistical measurement calculations on the original parameter sequence, calculate the standard deviation of the mean sequence as a measure of model cognitive uncertainty, and calculate the square root of the mean of the variance sequence as a measure of data random uncertainty. S53: Based on the confidence interval corresponding to the probability distribution prediction, the measurement of cognitive uncertainty and the measurement of data random uncertainty, and combined with the preset lifespan threshold and uncertainty threshold, multi-condition joint judgment processing is performed. According to the judgment result, different risk levels are matched and corresponding preventive maintenance suggestions are generated to form a decision signal.

8. A deep learning-based mechanical equipment operation status monitoring system, characterized in that, The system includes: The sample segmentation and annotation module is used to perform sliding window segmentation and annotation on the time-series data of the operating status synchronously collected by multi-source sensors deployed on mechanical equipment, dividing it into multiple sample segments containing corresponding real remaining service life values, and generating a set of sample pairs. The time-frequency feature fusion module is used to perform time-frequency transformation and feature fusion processing on the sample segments in the sample pair set, converting the channel signals in each sample segment into time-frequency spectrograms and stacking them along the channel dimension to generate a three-dimensional time-frequency feature tensor. The probability prediction network training module is used to take the three-dimensional time-frequency feature tensor as input and the real remaining lifetime value in the sample pair set as the supervision signal, call the deep neural network with built-in Monte Carlo random drop-out layer for probability training, optimize the negative log-likelihood loss function to make the network learn the distribution of prediction values, and generate a trained probability prediction network model. The probability distribution prediction module is used to extract signal features and construct tensors from the time series data to be monitored collected from the device to be monitored, convert it into a predictable feature tensor that is consistent with the three-dimensional time-frequency feature tensor structure, and call the probability prediction network model to perform multiple forward propagation sampling on the predictable feature tensor, aggregate the mean and variance parameters of the multiple outputs, and generate the probability distribution prediction of the remaining lifetime and the corresponding confidence interval. The uncertainty quantification decision module is used to perform uncertainty quantification analysis on the probability distribution prediction of the remaining useful life and the corresponding confidence interval, calculate the separated cognitive uncertainty and accidental uncertainty measures, and generate a decision signal containing risk level and maintenance recommendations in combination with preset threshold rules.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images