A Complex Equipment Fault Prediction Method and System Based on Multi-Dimensional Time-Series Feature Modulation

Abstract

The present invention relates to the technical field of fault prediction, and in particular, to a complex equipment fault prediction method and system based on multi-dimensional time series feature modulation. The method includes preprocessing the acquired sensor data; extracting initial features using the preprocessed sensor data; performing multi-dimensional time series feature encoding and multi-scale feature modulation based on the extracted initial features; refining the modulated features and smoothing the time series; and obtaining a fault classification result and a maintenance suggestion. The present invention eliminates the heterogeneity of multi-dimensional data through a data preprocessing module, and the DTFE module extracts global features from multiple domains, significantly improving the diagnostic accuracy, robustness, and real-time performance.

Description

Technical Field

[0001] The present invention relates to the technical field of fault prediction, and particularly to a complex equipment fault prediction method and system based on multi-dimensional time-series feature modulation. Background Art

[0002] As an important part of key industrial systems, the operating stability of complex petroleum equipment is directly related to the efficiency and safety of the entire production system. During the long-term service of the equipment, due to factors such as environmental changes, load impacts, and component aging, the equipment may experience a phenomenon where minor abnormalities gradually evolve into serious faults. To ensure that the equipment maintains good performance during operation, there is an urgent need for a health operation diagnosis method that can monitor the operating conditions of the equipment in real time, accurately identify potential risks, and predict possible fault trends. With the continuous development of industrial Internet of Things and artificial intelligence, more and more research attempts to use time-series data collected by multi-channel sensors to achieve equipment status assessment through a data-driven approach. However, existing diagnostic technologies still face many challenges. Especially in equipment systems with complex coupling relationships and multi-source heterogeneous signals, their recognition effects and generalization abilities far from meet the requirements of engineering applications.

[0003] Currently, the mainstream equipment health diagnosis methods mainly rely on traditional feature engineering or deep neural networks. By uniformly modeling the collected equipment signal data, the health status assessment results are then output. However, in practical applications, the operating state of the equipment often has obvious dynamic non-linearity and multi-scale evolution characteristics. Its signals contain both long-term trend evolution and short-term fluctuation abnormalities, resulting in the difficulty of accurately capturing key features by a single-scale and one-way modeling method. In addition, the interaction between different components may cause cross-channel abnormal responses, and there are significant deficiencies in existing methods for modeling the dependence relationship between the global and local states of the equipment. Moreover, some signal abnormalities initially manifest as minor fluctuations or short-term mutations. Due to the limited perception ability of traditional models, they are often difficult to detect in time, delaying the fault warning window and increasing maintenance risks. Therefore, there is an urgent need for an intelligent diagnosis method that integrates multi-dimensional structural information and time-series evolution laws to comprehensively improve the operating state modeling ability and real-time health judgment ability of complex equipment.

[0004] Therefore, the present invention proposes a complex equipment health operation diagnosis method and system based on multi-dimensional time-series feature modulation. Based on fully exploring the internal laws of equipment operation data, this method combines multi-dimensional time-series feature encoding and multi-dimensional time-series feature modulation to effectively capture the coupling characteristics between global trends and local abnormalities, and improves the stability and robustness of continuous state recognition through a time-series smoothing mechanism, thereby achieving accurate diagnosis and early warning of the operating health status of the equipment. Summary of the Invention

[0005] To solve the problems of the existing technology in multi-dimensional time series feature extraction, weak fault signal detection, and dynamic working condition adaptability, the present invention provides a complex equipment fault prediction method and system based on multi-dimensional time series feature modulation. By combining a deep one-dimensional convolutional neural network with multi-dimensional time series feature encoding (DTFE), multi-dimensional time series feature modulation (MTFT), time series smoothing, and fault classification techniques, the diagnostic accuracy, robustness, and real-time performance are comprehensively improved.

[0006] In the first aspect, a complex equipment fault prediction method based on multi-dimensional time series feature modulation provided by the present invention adopts the following technical solutions:

[0007] A complex equipment fault prediction method based on multi-dimensional time series feature modulation includes:

[0008] Obtain sensor data;

[0009] Preprocess the obtained sensor data;

[0010] Use the preprocessed sensor data for initial feature extraction;

[0011] Based on the extracted initial features, perform multi-dimensional time series feature encoding and multi-scale feature modulation;

[0012] Refine the modulated features and smooth the time series;

[0013] Obtain the fault classification result and maintenance suggestion.

[0014] Further, the preprocessing of the obtained sensor data includes performing data standardization on the obtained sensor data, then using wavelet denoising to decompose the signal and filter out high-frequency noise; and then performing data augmentation through random time translation, which is expressed as:

[0015]

[0016] Among them, is the denoised data; is the random time translation amount, which follows a uniform distribution ([[-10, 10]]).

[0017] Further, the initial feature extraction using the preprocessed sensor data includes performing multi-channel parallel convolution through a one-dimensional convolutional neural network to extract the short-term time series features of multi-dimensional data; and then normalizing the distribution of the convolutional output through layer normalization to process the multi-dimensional time series data to reduce the internal covariate shift, which is expressed as:

[0018]

[0019] Among them, is the convolutional output, and k = 1 is , where \(k = 2\) is ; respectively are the mean and variance of a small constant used to prevent division by zero; learnable parameters that respectively control scaling and translation.

[0020] Furthermore, the initial feature extraction using the preprocessed sensor data further includes using 1×1 convolution to achieve the integration of multi-sensor features through channel adjustment and weighted fusion to reduce redundancy and enhance the collaboration of multi-dimensional data; to prevent the complex equipment data from being affected by noise residues and causing model overfitting, adaptive max pooling is used to reduce the computational burden by compressing the time dimension. Among them, Dropout enhances the model robustness by randomly discarding neurons, expressed as:

[0021]

[0022]

[0023] where is the fused feature, with the dimension of ; is the adaptive max pooling, with the window size (\(w = 2\)), the stride of 2, and the output dimension of ; is to randomly discard 30% of the neurons to prevent overfitting.

[0024] Furthermore, the multi-dimensional time-series feature encoding and multi-scale feature modulation based on the extracted initial features include performing multi-dimensional feature pooling based on time-domain pooling, frequency-domain pooling, and statistical-domain pooling respectively, and forming a multi-dimensional feature table by feature concatenation of the obtained time-domain, frequency-domain, and statistical-domain features; finally, the redundant information is compressed and the dimension is reduced through convolution to enhance the time-series correlation, improve the feature quality while reducing the computational burden, expressed as:

[0025]

[0026]

[0027]

[0028] where is the concatenated feature, with the dimension of ; is convolution, reducing the dimension to 128 channels; is the intermediate feature, with the dimension ; has the convolution kernel size (\(k = 3\)), the number of output channels (\(C = 64\)), the stride of 1, and the padding of 1; is the convolution output, with dimension ; is layer normalization to prevent gradient explosion; is the activation function, introducing non - linearity, and Z is the final output, with dimension .

[0029] Furthermore, the multi - dimensional time - series feature encoding and multi - scale feature modulation based on the extracted initial features further include extracting diverse time - series patterns through multi - scale convolution to enhance the perception ability of diverse fault patterns, obtaining features at different time scales; forming a unified feature representation by integrating multi - scale features; and finally reducing the computational complexity through feature dimensionality reduction to support real - time diagnosis after enhancing key frequency components through the frequency - domain attention mechanism and highlighting key sensor features through the channel - attention mechanism, which is expressed as:

[0030]

[0031]

[0032]

[0033] Among them, is the output of the channel attention, with dimension ; is the DTFE output; is a 1×1 convolution to increase the dimension of Z to 128 channels; is the fused feature, with dimension ; is a 1×1 convolution to reduce the dimension to 64 channels; is layer normalization; is the activation function.

[0034] Furthermore, the feature refinement and time - series smoothing of the modulated features include enhancing high - order features through convolution refinement and directly adding the MTFT output to the refined features based on the residual connection to ensure the original features, where the residual link captures high - order time - series patterns through one - dimensional convolution to enhance the expressive ability of the features; and using batch normalization to reduce the internal covariate shift by standardizing the feature distribution, improving the training stability and generalization ability, which is expressed as:

[0035]

[0036]

[0037]

[0038] Among them, It is a one-dimensional convolution with a convolution kernel size of k = 3, an output channel number (C = 128), a stride of 1, and a padding of 1; They are the mean and variance of the batch data, calculated by channel; It is a small constant to prevent division by zero; They are learnable parameters that respectively control scaling and translation; It is the MTFT output, with dimension ; It is a 1×1 convolution that increases the number of channels from 64 to 128.

[0039] Furthermore, the feature refinement and temporal smoothing of the modulated features also include compressing the time dimension through adaptive max pooling while retaining key temporal features; and since the temporal data of complex equipment exhibits non-stationary characteristics under dynamic working conditions, the instability of the diagnostic results caused by feature mutations, the temporal smoothing is achieved through exponential moving average to smooth the dynamic changes of the features and improve the feature consistency and diagnostic reliability, expressed as:

[0040]

[0041]

[0042] Among them, It is adaptive max pooling with a window size of w = 2 and a stride of 2; It is the pooled feature at the current moment, with dimension ; It is the smoothed feature at the previous moment, with dimension ; It is the smoothing factor that controls the weights of the new and old features.

[0043] Furthermore, the obtaining of the fault classification result and maintenance suggestions includes generating the fault probability through a multi-layer fully connected network to ensure the classification accuracy and robustness; and converting it into specific suggestions through rule mapping to support the decision-making in industrial scenarios to complete the diagnostic process, expressed as:

[0044]

[0045]

[0046]

[0047]

[0048] Among them, They are the parameters of the first fully connected layer; They are the intermediate features with a dimension of 32; is the parameter of the second fully-connected layer. The 5 in the fifth power of the real number matrix R represents the number of fault categories (normal, wear, crack, overheat, electrical fault) respectively; is the fractional score, with a dimension of 5; The normalization function generates a probability distribution; P is the fault probability, with a dimension of 5.

[0049] In a second aspect, a complex equipment fault prediction system based on multi-dimensional time series feature modulation includes:

[0050] A data acquisition module, configured to acquire sensor data;

[0051] A preprocessing module, configured to preprocess the acquired sensor data;

[0052] A feature extraction module, configured to perform initial feature extraction using the preprocessed sensor data;

[0053] A modulation module, configured to perform multi-dimensional time series feature encoding and multi-scale feature modulation based on the extracted initial features;

[0054] A smoothing module, configured to refine the modulated features and smooth the time series;

[0055] A classification module, configured to obtain a fault classification result and a maintenance suggestion.

[0056] In a third aspect, the present invention provides a computer-readable storage medium, in which multiple instructions are stored, and the instructions are adapted to be loaded and executed by a processor of a terminal device for the method for predicting complex equipment faults based on multi-dimensional time series feature modulation.

[0057] In a fourth aspect, the present invention provides a terminal device, including a processor and a computer-readable storage medium. The processor is used to implement each instruction; the computer-readable storage medium is used to store multiple instructions, and the instructions are adapted to be loaded and executed by the processor for the method for predicting complex equipment faults based on multi-dimensional time series feature modulation.

[0058] In summary, the present invention has the following beneficial technical effects:

[0059] Compared with the deficiencies of the prior art in single-domain analysis, weak signal detection, dynamic working condition adaptability, and coordination of global and local features, the present invention eliminates the heterogeneity of multi-dimensional data through a data preprocessing module, extracts global features from multiple domains by a DTFE module, enhances the perception of weak fault signals by an MTFT module, optimizes the feature stability under dynamic working conditions by a feature refinement and smoothing module, and accurately predicts faults and generates suggestions by a fault classification and suggestion module. In scenarios such as fracturing equipment and wind power gearboxes, the early fault misdiagnosis rate is reduced from 25% to 8%, the misdiagnosis rate is reduced from 15% to 6%, the robustness of feature extraction is increased by 20%, and the single diagnosis time is shortened from 1.2 seconds to 0.5 seconds, significantly improving the diagnostic accuracy, robustness, and real-time performance.

[0060] Through the multi-dimensional time-series feature modulation technology, the present invention realizes the comprehensive analysis and processing of multi-dimensional data of complex equipment. The DTFE module extracts global features from the time domain, frequency domain, and statistical domain, the MTFT module combines multi-scale convolution and attention mechanism to enhance the perception of weak signals, and the collaborative design of multiple modules effectively copes with the challenges of dynamic working conditions, significantly improving the adaptability and reliability of fault prediction and diagnosis of complex equipment and meeting the requirements of high-reliability scenarios.

[0061] Through accurate fault prediction and diagnosis, the present invention optimizes the application effect of predictive maintenance strategies. In industrial scenarios such as fracturing equipment and wind power gearboxes, the technical solution supports efficient fault management, reduces the equipment downtime risk, improves the reliability and operation and maintenance efficiency of long-term operation, and provides strong support for the development of industrial intelligence. Description of the Drawings

[0062] Figure 1 It is a schematic diagram of a complex equipment fault prediction method based on multi-dimensional time-series feature modulation according to Embodiment 1 of the present invention.

[0063] Figure 2 It is an effect diagram of the misdiagnosis rate, misdiagnosis rate, and diagnosis time according to Embodiment 1 of the present invention.

[0064] Figure 3 It is a result diagram of the accuracy rate and suggestion applicability according to Embodiment 1 of the present invention.

[0065] Figure 4 It is a comprehensive performance radar chart according to Embodiment 1 of the present invention.

[0066] Figure 5 It is a comparison simulation diagram of the prediction lead time and the actual fault occurrence time according to Embodiment 1 of the present invention. Detailed Description of the Invention

[0067] The present invention will be further described in detail below with reference to the accompanying drawings.

[0068] Embodiment 1

[0069] Reference Figure 1 A complex equipment fault prediction method based on multi-dimensional time series feature modulation according to this embodiment includes:

[0070] Obtain sensor data;

[0071] Preprocess the obtained sensor data;

[0072] Use the preprocessed sensor data for initial feature extraction;

[0073] Perform multi-dimensional time series feature encoding and multi-scale feature modulation based on the extracted initial features;

[0074] Refine the modulated features and smooth the time series;

[0075] Obtain the fault classification result and maintenance suggestions;

[0076] Specifically:

[0077] (1). Multi-channel data preprocessing and feature initialization,

[0078] During the operation of complex equipment, sensors continuously collect multi-type signals such as temperature, vibration, current, voltage, and displacement. Since there are significant differences in amplitude, frequency, and dimension among signals of different channels, directly inputting them into the diagnostic model may cause feature suppression or response imbalance. Therefore, in this step, the multi-source sensor data (such as vibration, temperature, current) of complex equipment is first converted into standardized multi-channel time series features to eliminate noise and enhance weak fault signals, providing high-quality input for subsequent feature encoding.

[0079] 1) Data standardization,

[0080] The sensor data (such as vibration, temperature, current) of complex equipment has different dimensions and value ranges. For example, vibration signals are represented by acceleration (m / s²), and temperature is represented by degrees Celsius (°C). Direct processing will lead to unbalanced feature weights and affect the model's ability to fuse multi-dimensional data. Data standardization maps each sensor data to a unified range, eliminating dimension differences, and ensuring that subsequent convolution operations can fairly process multi-dimensional time series features. This is particularly important for complex equipment such as fracturing equipment because its multi-sensor data needs to be analyzed collaboratively to capture the operating state of multi-component coupling. The formula is as follows:

[0081]

[0082] Where, is the original value of sensor m at time t, , T is the length of the time series; is the minimum and maximum values of the data of sensor m.

[0083] 2) Wavelet denoising

[0084] Although data standardization unifies the dimension, the sensor data of complex equipment is still affected by noise. For example, the vibration signal of the wind power gearbox in the ultra-high pressure cutting system contains random noise due to wind speed changes, masking the weak signal of early wear. Direct feature extraction from the standardized data may cause noise to be misinterpreted as fault features, reducing the diagnostic accuracy. By using wavelet denoising to decompose the signal and filter out high-frequency noise, retaining the low-frequency and mid-frequency signals related to faults, the detection ability of weak fault signals is enhanced. The formula is as follows:

[0085]

[0086] Where is the sensor data after standardization, with a range of [0, 1]; is the wavelet denoising function, using Daubechies wavelet (db4); is the denoising threshold, used to control the intensity of noise filtering.

[0087] 3) Data augmentation

[0088] Since complex equipment is often in a dynamic working condition (such as the load change of the oil-water separation equipment), the distribution of time series data changes over time. Although the denoised data is relatively clean, static data cannot fully simulate the diversity of dynamic working conditions, which may lead to insufficient generalization ability of the model in practical applications. Data augmentation simulates the working condition changes through random time translation, enhancing the robustness of the model to the dynamic environment. The formula is as follows:

[0089]

[0090] Where is the denoised data; is the random time translation amount, following a uniform distribution ([-10, 10]).

[0091] 4) Initial feature extraction

[0092] The operating state of complex equipment is jointly characterized by multi-dimensional sensor data. Although the denoised and augmented data already has high quality, features that can reflect short-term time series patterns still need to be extracted to capture the local characteristics of fault signals (such as the instantaneous impact of vibration signals). Through one-dimensional convolutional neural network for multi-channel parallel convolution, the short-term time series features of multi-dimensional data are extracted, providing a basis for subsequent global feature encoding.

[0093]

[0094]

[0095] Among them, , is the enhanced multi-sensor data, T = 1024, M = 3; is a 1D convolution, the convolution kernel size (k = 3), the number of output channels (C = 32), the stride is 1, and the padding is 1.

[0096] 5) Layer normalization,

[0097] Since the high dynamicity of complex equipment data may lead to unstable training and affect model convergence. The distribution of the convolution output will be standardized through layer normalization, processing multi-dimensional time series data, reducing internal covariate shift, and improving training stability and generalization ability. The formula is as follows:

[0098]

[0099] Among them, is the convolution output, k = 1 is , k = 2 is ; are respectively the mean and variance of; is a small constant used to prevent division by zero; are learnable parameters, controlling scaling and translation respectively.

[0100] 6) Feature fusion,

[0101] Although parallel convolutions extract features of different depths, the operating state of complex equipment needs to be analyzed through the collaboration of multi-sensor data. It is necessary to fuse to generate a unified feature representation, reduce redundancy, and enhance the collaboration of multi-dimensional data. Here, 1×1 convolution is used to achieve the integration of multi-sensor features through channel adjustment and weighted fusion. The formula is as follows:

[0102]

[0103] Among them, are the normalized features, with dimensions and ; is convolution, which adjusts the number of channels from to 64; is convolution, which adjusts the number of channels from to 64.

[0104] 7) Pooling and regularization,

[0105] To prevent the complex equipment data from being affected by noise residues, resulting in model overfitting, adaptive max pooling is used to reduce the computational burden by compressing the time dimension (the original time dimension T = 1024), and Dropout is used to enhance the model's robustness by randomly discarding neurons. The formula is as follows:

[0106]

[0107]

[0108] Among them, is the fused feature, with a dimension of ; is the adaptive max pooling, with a window size (w = 2), a stride of 2, and an output dimension of ; is to randomly discard 30% of the neurons to prevent overfitting.

[0109] (2). Multidimensional time series feature encoding (DTFE),

[0110] The fault features of complex equipment are distributed in multiple domains. For example, the vibration signal of a tensioner device shows harmonic anomalies in the frequency domain, the temperature signal shows trend changes in the time domain, and the statistical domain reflects the non-normal distribution of the data. Traditional single-domain analysis (such as only the time domain) ignores the diversity of features, resulting in insufficient modeling of the global operating state and difficulty in capturing weak fault signals. DTFE integrates time domain, frequency domain, and statistical domain features through multidimensional pooling and feature fusion to extract the global time series features of complex equipment, capture the multidimensional characteristics of the operating state, and provide a comprehensive feature representation for subsequent feature modulation.

[0111] 1) Time domain pooling,

[0112] The time series data of complex equipment contains short-term trends and fluctuations. For example, the vibration signal of a fracturing equipment reflects load changes in a short period of time. Directly processing long time series data may ignore these local patterns and reduce the diagnostic accuracy. Time domain pooling extracts local time series features by calculating the short-time mean and standard deviation, captures the operating state under dynamic working conditions, and provides a time domain basis for global feature encoding. The output (X) of step 1 contains multidimensional local features, and short-time patterns need to be extracted through time domain pooling to construct a global representation:

[0113]

[0114]

[0115] Among them, is the output of step 1, , ; is the pooling window size, with a stride of 2; is the channel index; The short - term mean of channel (c). is the short - term standard deviation of channel c.

[0116] Final output , the number of channels is .

[0117] 2) Frequency - domain pooling,

[0118] In addition, the fault signals of complex equipment often show abnormalities in specific frequency components in the frequency domain. For example, the wear of the wind - power gearbox of the oil - water separation equipment leads to an increase in high - frequency harmonics. Although time - domain pooling extracts short - term trends, it cannot capture frequency - domain characteristics, limiting the detection ability for periodic faults. Frequency - domain pooling extracts the main frequency components through the Fast Fourier Transform (FFT) to enhance the perception of periodic fault signals to capture periodic characteristics and improve the global feature representation. The formula is as follows:

[0119]

[0120] where, is the time - series data of channel c, with a dimension of ; FFT is the Fast Fourier Transform, used to extract frequency components; are the first 16 main frequency components; is the frequency - domain feature, aligned with the time - series dimension through linear interpolation.

[0121] 3) Statistical - domain pooling,

[0122] Since the operation data of complex equipment often shows non - normal distribution, although time - domain and frequency - domain characteristics provide trend and periodic information, they cannot capture the statistical characteristics of data distribution, limiting the comprehensive description of abnormal states. Statistical - domain pooling enhances the expression ability of features for non - normal distribution by calculating skewness and kurtosis to improve multi - dimensional features. The formula is as follows:

[0123]

[0124]

[0125] where, is the eigenvalue of channel c at time t; are the mean and standard deviation of channel (c); is the skewness of channel c, reflecting the asymmetry of data distribution; is the kurtosis of channel (c), reflecting the sharpness of data distribution.

[0126] Final output , the number of channels is .

[0127] 4) Feature splicing

[0128] The operating state of complex equipment needs to be described through multi-domain feature collaboration, which requires combining time-domain trends, frequency-domain harmonics, and statistical distributions. Separate time-domain, frequency-domain, or statistical-domain features cannot comprehensively characterize fault modes. They need to be spliced to form multi-dimensional feature representations, improving the comprehensiveness of diagnosis. The formula is as follows:

[0129]

[0130] Where is the time-domain feature; is the frequency-domain feature; is the statistical-domain feature; is the spliced feature, and the number of channels is 128 + 16 + 128 = 272.

[0131] 5) Feature dimensionality reduction and enhancement

[0132] The real-time diagnosis requirements of complex equipment need more efficient feature representations. After splicing, the multi-dimensional feature has a large number of channels (272), increasing the computational complexity, and there may be redundancy among the features. Directly using them for subsequent processing will reduce efficiency. By using convolution to compress redundant information for dimensionality reduction, the temporal correlation will be enhanced, improving the feature quality while significantly reducing the computational burden. The formula is as follows:

[0133]

[0134]

[0135]

[0136] Where is the spliced feature, with a dimension of ; is convolution, reducing the dimension to 128 channels; is the intermediate feature, with a dimension of ; The convolution kernel size of is (k = 3), the number of output channels (C = 64), the stride is 1, and the padding is 1; is the convolution output, with a dimension of ; is layer normalization to prevent gradient explosion; is the activation function, introducing non-linearity. Z is the final output, with a dimension of

[0137] (3). Multi-dimensional Time Series Feature Modulation (MTFT)

[0138] The fault characteristics of complex equipment (such as cracks in engine blades) are manifested on different time scales (such as short-term vibration and long-term temperature drift) and frequency ranges. Traditional single-scale analysis is difficult to capture weak signals, resulting in a high misdiagnosis rate. MTFT captures diverse features through multi-scale convolution, and the frequency-domain and channel attention mechanisms enhance the weights of key fault signals, solving the problem of weak signal detection, modulating multi-dimensional time-series features, enhancing the perception of weak fault signals, and optimizing feature representation.

[0139] 1) Multi-scale convolution,

[0140] The fault signals of complex equipment have multi-time-scale characteristics. For example, the short-term vibration shock of a wind power gearbox reflects instantaneous anomalies, and the long-term temperature drift reflects cumulative wear. A single convolution kernel size cannot capture these features simultaneously and may ignore weak fault signals. The output (X) of step 1 contains local features, and multi-scale convolution is required to further extract diverse time-series patterns, enhancing the perception ability of diverse fault patterns and providing rich inputs for the subsequent attention mechanism. The formula is as follows:

[0141]

[0142] Among them, is the output of step 1; is the convolution kernel size, corresponding to the i-th branch respectively; One-dimensional convolution, the number of output channels (C = 32), the stride is 1, and the padding is 1, 2, 3, 4 respectively; represents the output of the i-th branch.

[0143] 2) Feature concatenation,

[0144] Multi-scale convolution generates features on different time scales, but using any single scale alone cannot comprehensively characterize the fault patterns of complex equipment. For example, the blade cracks of a tensioner device need to be analyzed by combining short-term and long-term features. Feature concatenation forms a unified feature representation by integrating multi-scale features, providing comprehensive inputs for the subsequent attention mechanism. The formula is as follows:

[0145]

[0146] Among them, represents the output of each scale convolution, with dimension ; is the concatenated feature, and the number of channels is .

[0147] 3) Frequency-domain attention,

[0148] The fault signals of complex equipment often exhibit anomalies at specific frequencies in the frequency domain. For example, electrical faults in fracturing equipment lead to the enhancement of specific frequency components. Although multi-scale features contain rich time-series information, they do not highlight the key signals in the frequency domain, which may result in the neglect of weak faults. Frequency-domain attention generates frequency weights through FFT and convolution, enhances the contribution of key frequency components, and improves the detection ability of weak fault signals. The formula is as follows:

[0149]

[0150]

[0151]

[0152]

[0153] Among them, is the multi-scale feature, with dimension ; is the fast Fourier transform, generating the frequency-domain representation. is the frequency-domain feature, with dimension ; is a 1×1 convolution, reducing the dimension to 64 channels; is the intermediate feature, with dimension ; is a 1×1 convolution, restoring to 128 channels; is the Sigmoid function, generating the attention weight; is the frequency-domain attention map, with dimension ; is element-wise multiplication; is the modulated feature, with dimension .

[0154] 4) Channel attention,

[0155] Since the contribution degrees of multi-dimensional sensor data of complex equipment are different. For example, the vibration signal of a wind power gearbox contributes more to fault detection than the temperature signal. Although frequency-domain modulation enhances the frequency features, it does not optimize the weight allocation between channels, which may lead to secondary features interfering with diagnosis. Channel attention generates channel weights through global pooling and fully connected layers, highlights the key sensor features, and improves the multi-dimensional data fusion efficiency to highlight the key channels. The formula is as follows:

[0156]

[0157]

[0158]

[0159]

[0160] Among them, is the frequency-domain modulation feature, with dimension ; is the global average pooling result, with dimension 128; The parameters of the first fully connected layer, reducing the dimension to 64; is the intermediate feature, with dimension 64; is the parameters of the second fully connected layer, restoring the 128-dimensionality; Channel attention map, with dimension 128; is the channel modulation feature, with dimension .

[0161] 5) Feature fusion and dimensionality reduction,

[0162] The diagnosis of complex equipment needs to balance the global operating state and local fault features. For example, the overall vibration mode of a fracturing device reflects the health level, and the local crack signal determines the diagnostic accuracy. Although the channel attention feature optimizes the local features, it needs to be fused with the global features of DTFE to take both into account, and at the same time reduce the dimension to reduce the computational complexity and support real-time diagnosis. The formula is as follows:

[0163]

[0164]

[0165]

[0166] Among them, is the channel attention output, with dimension ; is the DTFE output; is a 1×1 convolution, raising the dimension of Z to 128 channels; is the fusion feature, with dimension ; is a 1×1 convolution, reducing the dimension to 64 channels; is layer normalization; is the activation function.

[0167] (4). Feature refinement and temporal smoothing,

[0168] Complex equipment (such as fracturing equipment) generates non-stationary time-series data under dynamic working conditions. Traditional feature processing methods are vulnerable to noise and mutations, resulting in unstable diagnosis. This step ensures feature stability through residual connections, refines high-order features through convolution, smooths dynamic changes through temporal smoothing, and optimizes the stability and consistency of multi-dimensional time-series features to meet the real-time diagnosis requirements under the dynamic working conditions of complex equipment.

[0169] 1) Convolution refinement, batch normalization, and residual connection,

[0170] Although the features of the MTFT output have been optimized, the fault modes of complex equipment include high-order timing features, which need to be further extracted to improve the diagnostic accuracy. Convolutional refinement captures high-order timing patterns through one-dimensional convolution, enhancing the expressive ability of features; and uses batch normalization to reduce internal covariate shift by normalizing the feature distribution, improving training stability and generalization ability; Since the diagnosis of complex equipment needs to retain the original feature information, although convolutional refinement enhances the high-order features, some original information may be lost. Residual connection ensures the retention of the original features by directly adding the MTFT output to the refined features, improving the model stability and diagnostic accuracy. The formula is as follows:

[0171]

[0172]

[0173]

[0174] Among them, is one-dimensional convolution, with a kernel size of (k = 3), an output channel number of (C = 128), a stride of 1, and a padding of 1; are the mean and variance of the batch data, calculated by channel; is a small constant to prevent division by zero; are learnable parameters, controlling scaling and translation respectively; is the MTFT output, with dimension ; is a 1×1 convolution, which raises the channel number from 64 to 128.

[0175] 2) Adaptive pooling and temporal smoothing,

[0176] The real-time diagnosis of complex equipment requires efficient computation. The time dimension of the residual connection features is still relatively high, increasing the computational complexity. Adaptive max pooling is used to compress the time dimension, reducing the computational burden while retaining the key timing features; and since the temporal data of complex equipment exhibits non-stationary characteristics under dynamic working conditions, feature mutations may lead to unstable diagnostic results. Temporal smoothing uses exponential moving average to smooth the dynamic changes of features, improving feature consistency and diagnostic reliability. The formula is as follows:

[0177]

[0178]

[0179] Among them, is adaptive max pooling, with a window size of w = 2 and a stride of 2; is the pooled feature at the current moment, with dimension ; is the smoothed feature at the previous moment, with dimension ; is the smoothing factor, which controls the weights of the new and old features.

[0180] (5). Fault classification and maintenance suggestions,

[0181] Complex equipment (such as fracturing equipment) has various types of faults (such as blade cracks, bearing wear), and accurate classification is required to formulate maintenance strategies. Traditional classification methods are not sensitive to weak fault signals and cannot provide real-time suggestions under dynamic working conditions. In this step, through global pooling, multi-task pre-training, and rule mapping, the refined multi-dimensional time series features are mapped into fault probabilities, generating maintenance suggestions for complex equipment to support accurate diagnosis and decision-making.

[0182] 1) Global average pooling and multi-task pre-training,

[0183] If the refined features directly containing the time series dimension are used for classification, it will increase the computational complexity, and the diagnosis of complex equipment requires comprehensive consideration of the global operating state. Global average pooling compresses the time series dimension, extracts global features, reduces the computational burden, and generates a comprehensive representation; the sample size of fault data for complex equipment is limited, and directly training a classifier may lead to overfitting and reduce the generalization ability. Multi-task pre-training uses historical data to initialize the classifier, enhances the adaptability of the model to small-sample scenarios, and improves the detection ability for rare faults. The formula is as follows:

[0184]

[0185]

[0186] where, is the output of step 4; is the global pooling feature, with dimension 128; are the parameters of the pre-trained fully connected layer.

[0187] 2) Fault classification and generation of maintenance suggestions,

[0188] Complex equipment has various types of faults, such as wear, cracks, overheating, etc. in the fracturing equipment system. It is necessary to accurately distinguish through a classifier to support maintenance decisions. Although global pooling and pre-training optimize the feature representation, it is necessary to generate fault probabilities through a multi-layer fully connected network to ensure classification accuracy and robustness; moreover, the diagnosis of complex equipment not only needs to identify the type of fault, but also needs to provide actionable maintenance suggestions. Although the fault probability provides the classification result, it needs to be transformed into specific suggestions through rule mapping to support decision-making in industrial scenarios to complete the diagnosis process. The formula is as follows:

[0189]

[0190]

[0191]

[0192]

[0193] Among them, is the parameter of the first fully connected layer; is the intermediate feature with a dimension of 32; is the parameter of the second fully connected layer, and the 5 in the fifth power of the real number matrix R are the number of fault categories (normal, wear, crack, overheat, electrical fault) respectively; is the fractional score with a dimension of 5; is the normalization function, generating a probability distribution; P is the fault probability with a dimension of 5.

[0194] Experimental verification:

[0195] The operation data of the fracturing equipment was selected for the experiment, including multi-dimensional sound, light, and electrical sensor data such as vibration and temperature, to construct a test data set containing early faults and dynamic working conditions. A test data set containing early faults and dynamic working conditions was constructed. The load change rate was set (0, 0.2, 0.4, 0.6, 0.8, 1.0), where 0 represents no load change and 1.0 represents the maximum load fluctuation. The method of the present invention was compared with 5 baseline methods: SVM (Support Vector Machine), LSTM (Long Short-Term Memory Network), CNN (Convolutional Neural Network), RF (Random Forest), and DNN (Deep Neural Network). Fault prediction and diagnosis were performed on the same data set, and the fault prediction accuracy, diagnostic stability under dynamic working conditions, and applicability of maintenance suggestions were evaluated. The experimental results are as Figure 2 , Figure 3 , Figure 4 and shown in Table 1. Among them, in order to display different data indicators in the same radar chart, relevant processing was performed on the five indicators: ① Normalized to the [0, 1] interval; ② Reverse processing was performed on the indicators with "the smaller the better" (such as missed diagnosis rate, misdiagnosis rate, diagnosis time); ③ Direct normalization was performed on the indicators with "the larger the better" (accuracy, suggestion applicability). The method of the present invention is superior to the baseline methods under all load change rates.

[0196] Table 1 Data comparison of different methods under five major indicators

[0197] Indicator SVM LSTM CNN RF DNN The method of the present invention Early fault missed diagnosis rate 26% 22% 28% 24% 20% 8% Misdiagnosis rate 16% 14% 18% 15% 12% 6% Single diagnosis time 1.3 seconds 1.5 seconds 1.1 seconds 1.4 seconds 1.2 seconds 0.5 seconds Fault prediction accuracy 78% 80% 76% 75% 82% 92% Applicability of maintenance suggestions 72% 68% 70% 74% 78% 90%

[0198] The experimental results show that SVM has strong classification ability, but limited processing of multidimensional data. LSTM is suitable for time series data, but weak in multidimensional data fusion. CNN is good at feature extraction, but lacks perception of weak signals. RF has good robustness, but lacks prediction of complex fault modes. DNN has strong performance, but is slightly inferior in processing data heterogeneity. The method of the present invention is superior to the baseline method in all indicators, reflecting the comprehensive performance improvement of the technical solution.

[0199] In order to intuitively compare the performance of the "method of the present invention" with other methods (SVM, LSTM, CNN, RF, DNN) in terms of advance time of fault prediction, the present invention provides a simulation diagram of "advance time of prediction vs. actual fault occurrence time", such as Figure 5 The difference between the predicted failure time of each method and the actual failure time (100 seconds) is clearly shown through the line graph, reference line and lead time annotation.

[0200] Through simulation graph analysis, the method of the present invention performs well in fault prediction, and its 20-second prediction lead time is significantly better than SVM, LSTM, CNN, RF and DNN (8-12 seconds lead time). This method not only improves the timeliness of prediction, but also provides more response time for fault prevention, and has high practical value and application potential.

[0201] Example 2

[0202] This embodiment provides a complex equipment fault prediction system based on multi-dimensional time series feature modulation, including:

[0203] The data acquisition module is configured as follows:

[0204] A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, for a complex equipment fault prediction method based on multi-dimensional timing feature modulation.

[0205] A terminal device includes a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, wherein the instructions are suitable for being loaded and executed by the processor to perform a complex equipment fault prediction method based on multi-dimensional timing feature modulation.

[0206] The above are all preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A complex equipment fault prediction method based on multi-dimensional time series feature modulation, characterized in that, Including: Obtain sensor data; Preprocess the obtained sensor data; Perform initial feature extraction using the preprocessed sensor data; Perform multi-dimensional time-series feature encoding and multi-scale feature modulation based on the extracted initial features; The multi-dimensional time-series feature encoding and multi-scale feature modulation based on the extracted initial features includes performing multi-dimensional feature pooling respectively based on time-domain pooling, frequency-domain pooling, and statistical-domain pooling, and forming a multi-dimensional feature table by concatenating the obtained time-domain, frequency-domain, and statistical-domain features; finally, reducing redundant information and dimensionality through convolution to enhance time-series correlation, improve feature quality while reducing the computational burden, expressed as: ， ， ， Among them, is a splicing feature with a dimension of ; is a convolution that reduces the dimension to 128 channels; is an intermediate feature with a dimension of ; The convolution kernel size of k = 3, the number of output channels C = 64, the stride is 1, and the padding is 1; is the convolution output with a dimension of ; is layer normalization to prevent gradient explosion; is an activation function that introduces non-linearity, Z is the final output with a dimension of ; The multi-dimensional time-series feature encoding and multi-scale feature modulation based on the extracted initial features also includes further extracting diverse time-series patterns through multi-scale convolution to enhance the perception ability of diverse fault patterns, obtaining features of different time scales; forming a unified feature representation by integrating multi-scale features; finally, enhancing key frequency components through a frequency-domain attention mechanism and highlighting key sensor features through a channel attention mechanism, and then reducing the computational complexity through feature dimensionality reduction to support real-time diagnosis; Refine the modulated features and perform time-series smoothing; Obtain the fault classification result and maintenance suggestions.

2. The complex equipment fault prediction method based on multi-dimensional time series feature modulation according to claim 1, wherein, The preprocessing of the obtained sensor data includes performing data standardization on the obtained sensor data, decomposing the signal using wavelet denoising and filtering out high-frequency noise; then performing data augmentation through random time translation, expressed as: ， Among them, is the denoised data; is the random time shift amount, which follows a uniform distribution, i.e., [-10, 10].

3. A complex equipment fault prediction method based on multi-dimensional time series feature modulation according to claim 2, characterized in that, The initial feature extraction using the preprocessed sensor data includes performing multi-channel parallel convolution through a one-dimensional convolutional neural network to extract short-term time-series features of multi-dimensional data; Then normalize the distribution of the convolutional output through layer normalization to process multi-dimensional time-series data to reduce internal covariate shift, expressed as: ， Among them, is the convolution output, k = 1 is , k = 2 is ; are respectively the mean and variance of; a small constant used to prevent division by zero; are learnable parameters that respectively control scaling and translation.

4. The complex equipment fault prediction method based on multi-dimensional time series feature modulation according to claim 3, characterized in that, The initial feature extraction using the preprocessed sensor data also includes using 1×1 convolution to achieve the integration of multi-sensor features through channel adjustment and weighted fusion to reduce redundancy and enhance the cooperation of multi-dimensional data; to prevent the complex equipment data from being affected by noise residues and causing model overfitting, use adaptive max pooling to reduce the computational burden by compressing the time dimension, where Dropout enhances the model robustness by randomly discarding neurons, expressed as: ， ， Among them, is the fusion feature, with a dimension of ; is the adaptive max pooling, with a window size of w = 2, a stride of 2, and an output dimension of ; is to randomly discard 30% of the neurons to prevent overfitting.

5. A complex equipment fault prediction method based on multi-dimensional time series feature modulation according to claim 4, characterized in that, The multi-dimensional time-series feature encoding and multi-scale feature modulation based on the extracted initial features also includes further extracting diverse time-series patterns through multi-scale convolution to enhance the perception ability of diverse fault patterns, obtaining features of different time scales; forming a unified feature representation by integrating multi-scale features; finally, enhancing key frequency components through a frequency-domain attention mechanism and highlighting key sensor features through a channel attention mechanism, and then reducing the computational complexity through feature dimensionality reduction to support real-time diagnosis, expressed as: ， ， ， Among them, is the output of channel attention, with dimension ; is the DTFE output; is a 1×1 convolution that upsamples Z to 128 channels; is the fused feature, with dimension ; is a 1×1 convolution that downsamples to 64 channels; is layer normalization; is the activation function.

6. A complex equipment fault prediction method based on multi-dimensional time series feature modulation according to claim 5, characterized in that The feature refinement and temporal smoothing of the modulated features include enhancing high-order features through convolutional refinement and directly adding the MTFT output to the refined features based on residual connections to ensure the original features, where the residual links capture high-order temporal patterns through one-dimensional convolution to enhance the expressive power of the features; and using batch normalization to reduce internal covariate shift by normalizing the feature distribution, improving training stability and generalization ability; expressed as: ， ， ， Among them, is a one-dimensional convolution with a convolution kernel size of k = 3, an output channel number of C = 128, a stride of 1, and a padding of 1; are the mean and variance of the batch data, calculated by channel; is a small constant to prevent division by zero; are learnable parameters that respectively control scaling and translation; is the MTFT output, with dimension ; is a 1×1 convolution that increases the number of channels from 64 to 128.

7. A complex equipment fault prediction method based on multi-dimensional time series feature modulation according to claim 6, characterized in that, The feature refinement and temporal smoothing of the modulated features further include compressing the time dimension through adaptive max pooling while retaining key temporal features; and since the temporal data of complex equipment exhibits non-stationary characteristics under dynamic working conditions, the instability of diagnostic results caused by feature mutations, the temporal smoothing smooths the dynamic changes of the features through exponential moving average, improving feature consistency and diagnostic reliability, expressed as: ， ， Among them, is adaptive max pooling, with window size w = 2 and stride 2; is the pooling feature at the current moment, with dimension ; is the smoothed feature at the previous moment, with dimension ; is the smoothing factor, controlling the weights of the old and new features.

8. A complex equipment fault prediction method based on multi-dimensional time series feature modulation according to claim 7, characterized in that The obtaining of the fault classification result and maintenance suggestions includes generating fault probabilities through a multi-layer fully connected network to ensure classification accuracy and robustness; and transforming them into specific suggestions through rule mapping to support decision-making in industrial scenarios to complete the diagnostic process, expressed as: ， ， ， ， Among them, is the parameter of the first fully connected layer; is the intermediate feature, with a dimension of 32; is the parameter of the second fully connected layer. The fifth power of the real number matrix R represents the number of fault categories, namely normal, wear, crack, overheat, and electrical fault; is the fractional score, with a dimension of 5; is the normalization function, generating a probability distribution; P is the fault probability, with a dimension of 5.

9. A complex equipment fault prediction system based on multi-dimensional time series feature modulation, which executes a complex equipment fault prediction method based on multi-dimensional time series feature modulation as described in claim 1, characterized in that, Including: A data acquisition module configured to acquire sensor data; A preprocessing module configured to preprocess the acquired sensor data; A feature extraction module configured to perform initial feature extraction using the preprocessed sensor data; A modulation module configured to perform multi-dimensional temporal feature encoding and multi-scale feature modulation based on the extracted initial features; A smoothing module configured to perform feature refinement and temporal smoothing on the modulated features; A classification module configured to obtain a fault classification result and maintenance suggestions.

Images