Reciprocating compressor fault early warning method, system, device and storage medium

Abstract

The application discloses a kind of reciprocating compressor fault early warning method, system, equipment and storage medium, the method includes: to the pre-processing of multi-source sensor data, obtain standardized data frame sequence set;Standardized data frame sequence set is based on time domain feature extraction, frequency domain feature extraction, angle domain order feature extraction and displacement feature extraction;To the multi-source feature fusion normalization of extracted time domain feature, frequency domain feature, angle domain order feature and displacement feature;Fusion normalized multi-source feature vector is input into fault diagnosis model, and outputs fault diagnosis report, carries out fault early warning to reciprocating compressor, and fault diagnosis model is based on mechanism rule engine and CNN-LSTM attention network combination construction.The application effectively integrates the vibration, displacement, key phase and other multi-source sensor data of reciprocating compressor, overcomes the technical bottleneck of single sensor information deficiency and poor adaptability of traditional diagnostic method, and improves the accuracy of fault diagnosis.

Description

Methods, systems, equipment and storage media for early warning of reciprocating compressor faults Technical Field

[0001] This invention relates to the field of compressor fault detection technology, and in particular to a method, system, device and storage medium for early warning of reciprocating compressor faults. Background Technology

[0002] Reciprocating compressors are widely used in key industrial fields such as petrochemicals and natural gas processing, serving as core power equipment to ensure continuous and safe operation of production processes. Due to their complex structure and harsh working environment, long-term operation under high pressure and variable load conditions can easily lead to wear, breakage, and loosening of critical components such as valves, piston rings, connecting rods, and crossheads. If these malfunctions are not detected and addressed in a timely manner, they may cause equipment shutdowns or even major safety accidents, resulting in serious economic losses and safety hazards.

[0003] Currently, fault diagnosis of reciprocating compressors in industrial settings mainly relies on single sensor signals such as vibration acceleration, combined with the experience of maintenance personnel. This traditional method has many limitations:

[0004] First, data from a single measuring point provides limited information dimensions, making it difficult to comprehensively and accurately reflect the overall operating status and potential fault characteristics of the equipment.

[0005] Second, in the case of multiple source sensors (such as vibration acceleration, piston rod displacement, key phase signal, etc.), there is a lack of effective data fusion mechanism, which makes it impossible to make full use of the complementary and redundant information between different types of sensors, resulting in low utilization of diagnostic information.

[0006] Third, fault diagnosis often uses fixed thresholds or simple statistical rules, which are sensitive to changes in operating conditions, have poor adaptability, and are prone to false alarms or missed alarms. The ability to intelligently identify complex fault modes relies heavily on expert experience, and the diagnostic results are highly subjective.

[0007] Therefore, how to accurately identify reciprocating compressor faults for early warning has become an urgent problem to be solved.

[0008] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0009] The main objective of this invention is to provide a method, system, device, and storage medium for early warning of reciprocating compressor faults, aiming to address the technical problem of how to accurately identify reciprocating compressor faults for early warning.

[0010] To achieve the above objectives, the present invention provides a method for early warning of reciprocating compressor faults, the method comprising:

[0011] Acquire multi-source sensor data from the reciprocating compressor and preprocess the multi-source sensor data to obtain a standardized data frame sequence set;

[0012] Based on the standardized data frame sequence set, time-domain feature extraction, frequency-domain feature extraction, angular-domain order feature extraction, and displacement feature extraction are performed.

[0013] The extracted time-domain features, frequency-domain features, angular-domain order features, and displacement features are fused and normalized using multi-source features.

[0014] The fused and normalized multi-source feature vectors are input into the fault diagnosis model, which outputs a fault diagnosis report. The fault diagnosis model is constructed based on a combination of a mechanism rule engine and a CNN-LSTM attention network.

[0015] The reciprocating compressor is given a fault warning based on the fault diagnosis report.

[0016] Optionally, the preprocessing of the multi-source sensor data to obtain a standardized data frame sequence includes:

[0017] Outlier detection is performed on the multi-source sensor data, and the outliers are repaired using cubic spline interpolation.

[0018] The crankcase vibration acceleration, crosshead slide vibration acceleration, cylinder vibration acceleration, piston rod directional displacement, and key phase signal were extracted from the repaired multi-source sensor data.

[0019] Adaptive digital filtering is applied to the vibration acceleration of the crankcase, the vibration acceleration of the crosshead slide, and the vibration acceleration of the cylinder, respectively.

[0020] The time reference is aligned for the filtered crankcase vibration acceleration, the filtered crosshead slide vibration acceleration, the filtered cylinder vibration acceleration, the piston rod directional displacement, and the key phase signal.

[0021] Multi-source sensor data aligned with the time reference is encapsulated into a standardized set of data frame sequences.

[0022] Optionally, temporal feature extraction is performed based on the standardized data frame sequence set, including:

[0023] Extract the standardized data frame sequence subsets corresponding to the crankcase vibration sensor, the crosshead slide vibration, and the cylinder vibration sensor from the standardized data frame sequence set, respectively.

[0024] A sequence of valid crankcase measurements is generated based on a subset of the standardized data frame sequence corresponding to the crankcase vibration sensor.

[0025] A sequence of valid measurement values ​​for the crosshead slide is generated based on a subset of standardized data frame sequences corresponding to the vibration of the crosshead slide.

[0026] A sequence of valid cylinder measurement values ​​is generated based on a subset of the standardized data frame sequence corresponding to the cylinder vibration sensor.

[0027] The mean, effective value, peak value, peak factor, and kurtosis are calculated based on the effective measurement sequence of the crankcase, the effective measurement sequence of the crosshead slide, and the effective measurement sequence of the cylinder, respectively.

[0028] Temporal features are extracted based on the mean, the effective value, the peak value, the peak factor, and the kurtosis.

[0029] Optionally, frequency domain feature extraction is performed based on the standardized data frame sequence set, including:

[0030] Based on the standardized data frame sequence set, obtain the effective measurement value sequence of the crankcase, the effective measurement value sequence of the crosshead slide, and the effective measurement value sequence of the cylinder;

[0031] Fourier transforms are performed on the effective measurement sequence of the crankcase, the effective measurement sequence of the crosshead slide, and the effective measurement sequence of the cylinder, respectively.

[0032] Determine the fundamental frequency, harmonic amplitude, total harmonic distortion, and spectral centroid based on the transformed spectrum;

[0033] Frequency domain features are extracted based on the fundamental frequency, the harmonic amplitude, the total harmonic distortion, and the spectral centroid.

[0034] Optionally, angular domain order feature extraction is performed based on the standardized data frame sequence set, including:

[0035] Extract a subset of standardized data frame sequences of the key phase signal from the standardized data frame sequence set;

[0036] The total number of bond phase pulses is determined based on a standardized subset of the bond phase signal data frame sequence;

[0037] Calculate the angular domain signal based on the total number of bond phase pulses;

[0038] Perform a Discrete Fourier Transform on the angular domain signal, and determine different order amplitudes based on the angular domain signal after the Discrete Fourier Transform.

[0039] Angular domain order features are extracted based on different order amplitudes.

[0040] Optionally, the multi-source feature fusion and normalization of the extracted time-domain features, frequency-domain features, angular-domain order features, and displacement features includes:

[0041] Multi-source feature fusion is performed using a weighted average method based on the extracted time-domain features, frequency-domain features, angular-domain order features, and displacement features.

[0042] The fused multi-source features are normalized using a normalization algorithm.

[0043] Optionally, the step of inputting the fused and normalized multi-source feature vectors into the fault diagnosis model and outputting a fault diagnosis report includes:

[0044] Input the fused and normalized multi-source feature vectors into the fault diagnosis model;

[0045] Based on the mechanism-based rule base, the fused and normalized features are matched by the mechanism rule engine to obtain the basic probability allocation of the mechanism.

[0046] Based on the CNN-LSTM attention network, the predicted probabilities of various faults are output through a fully connected layer according to the fused and normalized features.

[0047] The predicted probabilities of various faults and their corresponding probability distributions are weighted by confidence factors to obtain the basic probability allocation for deep learning.

[0048] The fault diagnosis result is obtained by using the Dempster combination rule based on the fundamental probability allocation of the mechanism and the fundamental probability allocation of deep learning.

[0049] Based on the fault diagnosis results, a fault diagnosis report is output through the fault diagnosis model.

[0050] Furthermore, to achieve the above objectives, the present invention also proposes a reciprocating compressor fault early warning system, the reciprocating compressor fault early warning system comprising:

[0051] The data preprocessing module is used to acquire multi-source sensor data of the reciprocating compressor and preprocess the multi-source sensor data to obtain a standardized data frame sequence set.

[0052] The feature extraction module is used to perform time-domain feature extraction, frequency-domain feature extraction, angular-domain order feature extraction, and displacement feature extraction based on the standardized data frame sequence set.

[0053] The fusion and normalization module is used to perform multi-source feature fusion and normalization on the extracted time-domain features, frequency-domain features, angular-domain order features, and displacement features;

[0054] The fault detection module is used to input the fused and normalized multi-source feature vectors into the fault diagnosis model and output a fault diagnosis report. The fault diagnosis model is constructed based on a combination of a mechanism rule engine and a CNN-LSTM attention network.

[0055] The fault early warning module is used to provide fault early warning for the reciprocating compressor based on the fault diagnosis report.

[0056] Furthermore, to achieve the above objectives, the present invention also proposes a reciprocating compressor fault early warning device, the device comprising: a memory, a processor, and a reciprocating compressor fault early warning program stored in the memory and executable on the processor, the reciprocating compressor fault early warning program being configured to implement the steps of the reciprocating compressor fault early warning method as described above.

[0057] Furthermore, to achieve the above objectives, the present invention also proposes a storage medium storing a reciprocating compressor fault early warning program, wherein when the reciprocating compressor fault early warning program is executed by a processor, it implements the steps of the reciprocating compressor fault early warning method described above.

[0058] This invention first acquires multi-source sensor data from a reciprocating compressor and preprocesses the data to obtain a standardized data frame sequence set. Then, based on this set, it extracts temporal, frequency, angular, and displacement features. Next, it fuses and normalizes the extracted features, and inputs the fused and normalized multi-source feature vectors into a fault diagnosis model, outputting a fault diagnosis report. The fault diagnosis model is constructed based on a combination of a mechanism rule engine and a CNN-LSTM attention network. Finally, it provides fault warnings for the reciprocating compressor based on the fault diagnosis report. This invention effectively integrates multi-source sensor data such as vibration, displacement, and bond phase data from the reciprocating compressor, overcoming the technical bottlenecks of insufficient information from a single sensor and poor adaptability of traditional diagnostic methods. It achieves intelligent fault identification, precise location, and quantitative assessment of fault severity, significantly improving the accuracy, reliability, and predictability of fault diagnosis. Attached Figure Description

[0059] Figure 1 is a schematic diagram of the structure of a reciprocating compressor fault early warning device in the hardware operating environment of the embodiment of the present invention;

[0060] Figure 2 is a flowchart illustrating the first embodiment of the reciprocating compressor fault early warning method of the present invention;

[0061] Figure 3 is a structural block diagram of the first embodiment of the reciprocating compressor fault early warning system of the present invention.

[0062] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0063] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0064] Referring to Figure 1, Figure 1 is a schematic diagram of the structure of a reciprocating compressor fault early warning device in the hardware operating environment involved in the embodiment of the present invention.

[0065] As shown in Figure 1, the reciprocating compressor fault early warning device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be high-speed random access memory (RAM) or stable non-volatile memory (NVM), such as a disk storage device. Optionally, the memory 1005 may also be a storage system independent of the aforementioned processor 1001.

[0066] Those skilled in the art will understand that the structure shown in Figure 1 does not constitute a limitation on the reciprocating compressor fault warning device, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0067] As shown in Figure 1, the memory 1005, which serves as a storage medium, may include an operating system, a network communication module, a user interface module, and a reciprocating compressor fault early warning program.

[0068] In the reciprocating compressor fault early warning device shown in Figure 1, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the reciprocating compressor fault early warning device of the present invention can be set in the reciprocating compressor fault early warning device. The reciprocating compressor fault early warning device calls the reciprocating compressor fault early warning program stored in the memory 1005 through the processor 1001 and executes the reciprocating compressor fault early warning method provided in the embodiment of the present invention.

[0069] This invention provides a method for early warning of reciprocating compressor faults. Referring to Figure 2, which is a flowchart of the first embodiment of the method for early warning of reciprocating compressor faults according to this invention.

[0070] In this embodiment, the reciprocating compressor fault early warning method includes the following steps:

[0071] S1, acquire multi-source sensor data of the reciprocating compressor, and preprocess the multi-source sensor data to obtain a standardized data frame sequence set.

[0072] It is easy to understand that the executing entity in this embodiment can be a reciprocating compressor fault early warning system with functions such as data processing, network communication and program operation, or other computer equipment with similar functions. This embodiment does not limit it.

[0073] It should be noted that crankcase vibration sensors, crosshead slide vibration sensors, cylinder vibration sensors, displacement sensors, and key phase sensors are installed at specific locations on the reciprocating compressor.

[0074] In the specific implementation, multi-source sensor data is synchronously acquired on the reciprocating compressor to obtain signals from five key measuring points of the reciprocating compressor: crankcase vibration acceleration. Used to characterize the meshing state of the main bearing and gear; crosshead slide vibration acceleration. This reflects the impact and wear behavior of the connecting rod-slide pair; cylinder vibration acceleration. It carries information such as valve operation, cylinder pressure fluctuations, and piston ring sealing status; piston rod displacement in the X / Y directions. [μm] is used to monitor the piston rod swing trajectory, stuffing box alignment, and bond phase signal. As a reference for rotation angle, it enables periodic synchronous sampling under non-stationary operating conditions.

[0075] All channels are synchronously acquired under the drive of a unified clock source, which meets the requirements of the Nyquist theorem. The sampling frequency covers the main fault characteristic frequency band, the sampling duration covers a sufficient number of rotation cycles to ensure statistical stability, and the time synchronization error between multiple channels is controlled at the microsecond level to ensure the phase alignment accuracy of cross-modal signals.

[0076] After sensor initialization, the system establishes a unified time synchronization reference to eliminate time deviations between multiple nodes. Subsequently, the system uses the rising edge of the key phase signal output pulse as the synchronization trigger source to initiate the multi-source sensor data acquisition process. When the rising edge of the key phase signal is detected, the central controller sends a synchronization start command to each sensor, and all sensors begin sampling at the same time, ensuring strict alignment of all physical signals on the time axis. The acquisition process is continued for a preset duration, dynamically adjusted according to the compressor speed, to cover no less than six complete working cycles, ensuring that the acquired signals have sufficient periodic representativeness.

[0077] Furthermore, the preprocessing of multi-source sensor data to obtain a standardized data frame sequence involves: outlier detection of the multi-source sensor data, followed by repair of outliers using cubic spline interpolation; extraction of crankcase vibration acceleration, crosshead slide vibration acceleration, cylinder vibration acceleration, piston rod displacement, and key phase signals from the repaired multi-source sensor data; adaptive digital filtering of the crankcase vibration acceleration, crosshead slide vibration acceleration, and cylinder vibration acceleration; time-base alignment of the filtered crankcase vibration acceleration, filtered crosshead slide vibration acceleration, filtered cylinder vibration acceleration, piston rod displacement, and key phase signals; and encapsulation of the time-base aligned multi-source sensor data into a standardized data frame sequence set.

[0078] It should be noted that the collected multi-source sensor data is raw data.

[0079] In the specific implementation, the raw data collected by each sensor is processed to obtain data sequences for each channel. Continuity checks are performed on these data sequences. If discontinuous sequence numbers or timestamp jumps exceeding a preset threshold are found, an acquisition interruption is determined, and a resampling mechanism is triggered. Subsequently, outlier detection is performed on the data from each channel of each sensor.

[0080]

[0081] Calculate the mean of the sequence. with standard deviation For each sampling point x, the system checks whether it satisfies the above formula. If it does, it is marked as an outlier. For the marked outliers, the system uses cubic spline interpolation to repair them based on the adjacent valid data points. If the number of consecutive outliers exceeds 10% of the window length, the data segment is deemed invalid and marked as unusable.

[0082] After anomaly detection and repair are completed, adaptive digital filtering is performed on the vibration signals (i.e., crankcase vibration acceleration, crosshead slide vibration acceleration, and cylinder vibration acceleration).

[0083] The system calls a preset sixth-order Butterworth low-pass filter with a cutoff frequency of Dynamically set according to the current rotational speed n (unit: rpm) to meet the requirements. Where α is the safety margin coefficient, with a value ranging from 1.2 to 1.5. Let be the highest harmonic order of the fault characteristics to be analyzed, and The maximum allowable value is 12.8 kHz. This filter uses a bilinear transform method to convert the analog prototype into a digital filter, and is implemented with a three-stage cascaded second-order section (Biquad) structure to improve numerical stability and reduce the effects of finite word length. The filtering operation of each second-order section is performed through the following difference equation:

[0084]

[0085] in, and Let represent the input and output sample values ​​at the m-th time point, respectively; , , Forward coefficients; , The feedback coefficients are obtained by looking up tables online or calculating in real time based on the cutoff frequency. The system sequentially performs the above filtering operation on the vibration signal of each channel point by point to complete noise suppression.

[0086] After filtering, all sensor data (i.e., filtered crankcase vibration acceleration, filtered crosshead slide vibration acceleration, filtered cylinder vibration acceleration, piston rod directional displacement, and key phase signal) are organized into a unified data format. Each data record contains a timestamp and a measured value, forming a time-value pair. , ), where timestamp Determined by the following formula:

[0087]

[0088] in, The time of this data acquisition is the start time, synchronized to UTC standard time via the IEEE 1588 protocol, with an accuracy better than ±1μs; i is the sampling sequence number. This represents the sampling frequency of the corresponding channel; The residual deviation after clock drift compensation is estimated and corrected in real time by a Kalman filter. Data from all channels is aligned based on this time base, forming a multi-source dataset with precise time correlation. Subsequently, the processed multi-source data is encapsulated into a standardized data frame sequence set. This set includes subsets of standardized data frame sequences corresponding to crankcase vibration sensors, crosshead slide vibration, cylinder vibration sensors, piston rod displacement, and key phase signals. The structure of this standardized data frame sequence is defined as follows:

[0089]

[0090] In this data frame, Timestamp is a high-precision timestamp collected by the system; SensorID identifies the sensor number (e.g., P001, T003, etc.) representing sensors at different measuring points on the same device; Value is the filtered measurement value of the sensor (i.e., the filtered crankcase vibration acceleration, the filtered crosshead slide vibration acceleration, the filtered cylinder vibration acceleration) or piston rod displacement, key phase signal; and QualityFlag is the data quality flag (normal, interpolation repair, invalid, etc.). This data frame serves as the basic unit for subsequent analysis, ensuring clear data semantics and a consistent format.

[0091] S2, based on the standardized data frame sequence set, perform time-domain feature extraction, frequency-domain feature extraction, angular-domain order feature extraction, and displacement feature extraction.

[0092] Furthermore, the processing method for time-domain feature extraction based on the standardized data frame sequence set is as follows: extract the standardized data frame sequence subsets corresponding to the crankcase vibration sensor, the crosshead slide vibration, and the cylinder vibration sensor from the standardized data frame sequence set; generate the crankcase effective measurement value sequence based on the standardized data frame sequence subset corresponding to the crankcase vibration sensor; generate the crosshead slide effective measurement value sequence based on the standardized data frame sequence subset corresponding to the crosshead slide vibration; generate the cylinder effective measurement value sequence based on the standardized data frame sequence subset corresponding to the cylinder vibration sensor; calculate the mean, effective value, peak value, peak factor, and kurtosis based on the crankcase effective measurement value sequence, crosshead slide effective measurement value sequence, and cylinder effective measurement value sequence, respectively; and perform time-domain feature extraction based on the mean, effective value, peak value, peak factor, and kurtosis.

[0093] It should be understood that the subset of standardized data frame sequences corresponding to the sensor includes multiple standardized data frame sequences, and then valid measurement values ​​are extracted from each standardized data frame sequence according to time order to construct a valid measurement value sequence.

[0094] In the specific implementation, for a sensor's effective measurement value sequence x(1), x(2), ..., x(N), a set of standardized statistical features are calculated to reflect the signal's energy level and impulse characteristics within that period. The extracted features include: mean, RMS value, peak value, peak factor, and kurtosis.

[0095]

[0096] Where Mean is the mean, RMS is the effective value, Peak is the peak value, CrestFactor is the peak factor, and Kurtosisi is the kurtosis. The standard deviation is denoted as .

[0097] When Kurtosisi > 4.5, it indicates the presence of significant pulse components in the signal, often used for early bearing or piston ring wear detection. The above characteristics constitute a subset of the sensor's time-domain features within the current cycle, supplemented by a cycle start timestamp. Combined with the SensorID identifier, a labeled feature tuple is formed.

[0098] Furthermore, the processing method for frequency domain feature extraction based on the standardized data frame sequence set is as follows: obtain the effective measurement value sequence of the crankcase, the effective measurement value sequence of the crosshead slide, and the effective measurement value sequence of the cylinder based on the standardized data frame sequence set; perform Fourier transform on the effective measurement value sequence of the crankcase, the effective measurement value sequence of the crosshead slide, and the effective measurement value sequence of the cylinder respectively; determine the fundamental frequency, harmonic amplitude, total harmonic distortion, and spectral centroid based on the transformed spectrum; and perform frequency domain feature extraction based on the fundamental frequency, harmonic amplitude, total harmonic distortion, and spectral centroid.

[0099] The processing method for obtaining the effective measurement value sequences of the crankcase, crosshead slide, and cylinder based on the standardized data frame sequence set is as follows: Extract the standardized data frame sequence subsets corresponding to the crankcase vibration sensor, the crosshead slide vibration sensor, and the cylinder vibration sensor from the standardized data frame sequence set, respectively; generate the effective measurement value sequence of the crankcase based on the standardized data frame sequence subset corresponding to the crankcase vibration sensor; generate the effective measurement value sequence of the crosshead slide based on the standardized data frame sequence subset corresponding to the crosshead slide vibration; and generate the effective measurement value sequence of the cylinder based on the standardized data frame sequence subset corresponding to the cylinder vibration sensor.

[0100] In the specific implementation, a Fast Fourier Transform (FFT) is performed on the vibration signal sequence within the same period (i.e., the effective measurement sequence of the crankcase, the effective measurement sequence of the crosshead slide, and the effective measurement sequence of the cylinder) to obtain its spectral representation X(f). The following key frequency domain parameters are then calculated and used as spectral features:

[0101] Fundamental frequency identification: Calculate the theoretical fundamental frequency based on the rotational speed n (rpm) provided by the same-period speed sensor. ;

[0102] Harmonic amplitude extraction: Extracting the first harmonic frequency 2nd harmonic 3rd harmonic 4x frequency The amplitude at the point , , , ;

[0103] Total Harmonic Distortion (THD):

[0104]

[0105] The degree of nonlinear distortion is indicated;

[0106] Spectral Centroid:

[0107]

[0108]

[0109] In the formula, This represents the actual frequency corresponding to the k-th frequency bin. Sampling rate, Let N be the amplitude corresponding to the k-th frequency bin, and N be the number of points in the FFT.

[0110] Reflecting the trend of concentrated spectral energy, faults often shift to higher frequencies. Finally, the above frequency domain features, along with the corresponding SensorID and timestamp, are stored in the feature dataset.

[0111] Furthermore, the processing method for extracting angular domain order features based on the standardized data frame sequence set is as follows: extract a subset of standardized data frame sequences of the key phase signal from the standardized data frame sequence set; determine the total number of key phase pulses based on the subset of standardized data frame sequences of the key phase signal; calculate the angular domain signal based on the total number of key phase pulses; perform a discrete Fourier transform on the angular domain signal, and determine different order amplitudes based on the angular domain signal after the discrete Fourier transform; extract angular domain order features based on the different order amplitudes.

[0112] In the specific implementation, to eliminate the influence of rotational speed fluctuations, angular domain resampling technology is used to process the vibration signal. The specific process is as follows:

[0113] Using the rising edge of the key phase signal as a reference, the interval between adjacent triggering moments is divided into one mechanical cycle. Let the total number of bond phase pulses within this period be... Then the angle sampling interval is Δθ = 2π / Based on the key-phase timestamp sequence The original vibration signal at the equiangular position was determined by linear interpolation. =m The amplitude at Δθ is used to obtain the angular domain signal x( Perform a Discrete Fourier Transform on the signal:

[0114]

[0115] Where n is the order number, N is the number of sampling points per revolution, j is the imaginary unit, and m is the summation index (representing the m-th isotropic sampling point). Key order amplitudes are extracted as features:

[0116] First order: A1 = |X(1)|, corresponding to imbalance fault;

[0117] Second order: A2 = |X(2)|, reflecting poor centering;

[0118] 0.5 order: A0.5 = |X(0.5N)|, indicating piston ring breakage or valve leakage;

[0119] z-order: Az = |X(z)|, where z is the piston number, representing the imbalance of piston force.

[0120] All order features, along with their corresponding SensorIDs and period start timestamps, are encapsulated into structured data output.

[0121] S3 performs multi-source feature fusion and normalization on the extracted time-domain features, frequency-domain features, angular-domain order features, and displacement features.

[0122] Furthermore, the processing method for multi-source feature fusion and normalization of the extracted time-domain features, frequency-domain features, angular-domain order features and displacement features is as follows: multi-source features are fused using a weighted average method based on the extracted time-domain features, frequency-domain features, angular-domain order features and displacement features; and the fused multi-source features are normalized using a normalization algorithm.

[0123] It should also be noted that the pre-processed piston rod directional displacement is directly used as the displacement feature.

[0124] In its implementation, to generate comprehensive features with device-level characterization capabilities, the system fuses similar features from different sensors (SensorID). First, the time-domain, frequency-domain, angular-domain, and displacement features extracted from each sensor within the same device cycle are concatenated in a preset order to form the original feature vector. To address data quality differences, a quality weighting factor is introduced. For participation in fusion computing, its value is determined based on the QualityFlag in the data frame: ("Normal": =1.0; "Interpolation Repair": =0.7; "Invalid": =0.0;). For the characteristics of the same physical quantity (such as the RMS values ​​of each channel), a weighted average method is used for fusion:

[0125]

[0126] in, Let k be the feature value of the k-th sensor. For its empirical sensitivity weight (e.g., vibration sensor) =0.3, pressure sensor =0.4), ensuring a higher contribution to key measurement points. The fused features are further processed using min-max normalization to eliminate the influence of dimensions and improve cross-cycle comparability:

[0127]

[0128] in , This represents the statistical extreme value of this feature under historical normal operating conditions of the equipment. The fused and normalized feature vector. As input to the fault diagnosis model, it ensures the stability and consistency of subsequent analyses.

[0129] S4. Input the fused and normalized multi-source feature vector into the fault diagnosis model and output a fault diagnosis report. The fault diagnosis model is constructed based on a combination of a mechanism rule engine and a CNN-LSTM attention network.

[0130] It should be noted that the fusion and normalization of multi-source feature vectors Each feature vector corresponds to a mechanical cycle and includes time-domain, frequency-domain, angular-domain order, and displacement features from multiple sources such as vibration, pressure, and temperature sensors.

[0131] The fault diagnosis model adopts a "dual-channel parallel diagnosis + evidence fusion decision" architecture. One channel builds a rule engine based on the equipment operation mechanism to make physical consistency judgments, while the other channel uses a lightweight CNN-LSTM network to capture the hidden patterns in the data. Finally, the two types of heterogeneous evidence are fused through Bayesian modified DS evidence theory to generate high-confidence diagnostic results, taking into account accuracy, interpretability and engineering reliability.

[0132] Furthermore, the process of inputting the fused and normalized multi-source feature vectors into the fault diagnosis model and outputting the fault diagnosis report is as follows: The fused and normalized multi-source feature vectors are input into the fault diagnosis model; based on the mechanism judgment rule base, the fused and normalized features are matched using a mechanism rule engine to obtain the basic mechanism probability allocation; based on a Convolutional Neural Network (CNN) - Long Short-Term Memory (LSTM) attention network, the predicted probabilities of various faults are output through a fully connected layer according to the fused and normalized features; the predicted probabilities of various faults and their corresponding probability distributions are weighted by a confidence factor to obtain the basic deep learning probability allocation; the fault diagnosis result is obtained through the Dempster combination rule based on the basic mechanism probability allocation and the basic deep learning probability allocation; and the fault diagnosis report is output through the fault diagnosis model based on the fault diagnosis result.

[0133] In its implementation, based on the dynamic model of reciprocating compressors and long-term operation and maintenance experience, an executable mechanism judgment rule base is constructed. The system determines the characteristics of the current cycle. Perform a match-by-match and output the basic probability assignment (BPA) with confidence level (i.e., mechanistic basic probability assignment). The output of each rule is converted into a basic probability assignment function. :

[0134]

[0135] in, ∈[0,1] is the rule confidence factor (calibrated based on historical verification accuracy), and Ω is the union of all faults.

[0136] Construct a CNN-LSTM attention network to capture nonlinear temporal patterns in feature sequences. Then, process the feature vectors from T consecutive cycles {...} The organization is for the input X∈ First, local temporal patterns are extracted using three stacked one-dimensional convolutional layers: each layer uses a convolutional kernel of size k=3, with the number of kernels being 64, 128, and 256 respectively, all equipped with ReLU activation function, batch normalization (BatchNorm1d), and Dropout (0.3), mapping the input sequence to 16×256 high-dimensional temporal features. The data is then fed into a single-layer bidirectional LSTM (both forward and backward hidden layers have a dimension of 128) to capture long-range dependencies between cycles, resulting in a sequence of hidden states containing historical and future contextual information. Building upon this, a time attention mechanism is introduced, utilizing learnable parameters. , and Calculate the attention weights for each cycle:

[0137]

[0138] The context vector is obtained by weighted summation. Finally, the predicted probabilities of various faults are output through the fully connected layer. The predicted probabilities and corresponding probability distributions of these various fault types are weighted by a confidence factor β (based on the validation set AUC) and used as the basic probability allocation (BPA) in DS evidence theory (i.e., deep learning basic probability allocation). Participate in integration decision-making.

[0139] The entire model is implemented in the PyTorch framework and trained offline using the Adam optimizer (learning rate 0.001, batch size=32). The loss function is weighted cross-entropy. After training, it is deployed on edge devices in ONNX format to meet industrial real-time requirements.

[0140] Output the mechanism rules = , and deep learning output As two independent sources of evidence, they were fused using Dempster's combination rule:

[0141]

[0142] Among them, the conflict coefficient is:

[0143]

[0144] The consistency between the mechanistic assessment and the data model is reflected in the final diagnostic result:

[0145]

[0146] Simultaneously output the maximum confidence level (max(Bel)) and the conflict coefficient K. If K > 0.3, a "diagnostic conflict alarm" is triggered, indicating that manual intervention is required for analysis. The diagnostic results are packaged into a structured report, including the fault type and confidence level.

[0147] S5, based on the fault diagnosis report, provide a fault warning for the reciprocating compressor.

[0148] In practice, four response levels are dynamically defined: normal, minor, warning, and alarm. Fault diagnosis reports are analyzed. A fault type classified as normal with a confidence level higher than 0.8 is considered "normal." For minor faults with a severity level between 0.3 and 0.5 and a confidence level greater than 0.7, the "minor" level is triggered, and enhanced monitoring is initiated. For moderate faults with a severity level between 0.5 and 0.8 and a confidence level greater than 0.6, a "warning" state is entered, generating a preventative maintenance work order and recommending a shutdown for inspection within 7 days. If the fault is severe or has a severity level ≥ 0.8, or if the confidence level is low but the severity level is high (> 0.7), an "alarm" is immediately triggered, prompting an emergency shutdown.

[0149] It should be noted that the dynamic division of response strategies into four levels—normal, minor, warning, and alarm—and the generation of actionable maintenance recommendations based on historical trend analysis, achieves closed-loop management from "fault identification" to "operation and maintenance decision-making," thereby improving the intelligence and preventative level of equipment management.

[0150] In this embodiment, multi-source sensor data of the reciprocating compressor is first acquired and preprocessed to obtain a standardized data frame sequence set. Then, based on the standardized data frame sequence set, time-domain feature extraction, frequency-domain feature extraction, angular-domain order feature extraction, and displacement feature extraction are performed. Subsequently, the extracted time-domain, frequency-domain, angular-domain order features, and displacement features are fused and normalized. The fused and normalized multi-source feature vector is input into the fault diagnosis model, which outputs a fault diagnosis report. The fault diagnosis model is constructed based on a combination of a mechanism rule engine and a CNN-LSTM attention network. Finally, based on the fault diagnosis report, a fault warning is issued for the reciprocating compressor. This embodiment effectively integrates multi-source sensor data such as vibration, displacement, and bond phase data of the reciprocating compressor, overcoming the technical bottlenecks of insufficient information from a single sensor and poor adaptability of traditional diagnostic methods. It achieves intelligent fault identification, precise location, and quantitative assessment of fault severity, significantly improving the accuracy, reliability, and predictability of fault diagnosis.

[0151] Referring to Figure 3, which is a structural block diagram of the first embodiment of the reciprocating compressor fault early warning system of the present invention.

[0152] As shown in Figure 3, the reciprocating compressor fault early warning system proposed in this embodiment of the invention includes:

[0153] The data preprocessing module 3001 is used to acquire multi-source sensor data of the reciprocating compressor and preprocess the multi-source sensor data to obtain a standardized data frame sequence set.

[0154] Feature extraction module 3002 is used to perform time-domain feature extraction, frequency-domain feature extraction, angular-domain order feature extraction, and displacement feature extraction based on the standardized data frame sequence set;

[0155] The fusion normalization module 3003 is used to perform multi-source feature fusion normalization on the extracted time-domain features, frequency-domain features, angular-domain order features and displacement features;

[0156] The fault detection module 3004 is used to input the fused and normalized multi-source feature vector into the fault diagnosis model and output a fault diagnosis report. The fault diagnosis model is constructed based on a combination of a mechanism rule engine and a CNN-LSTM attention network.

[0157] The fault warning module 3005 is used to provide fault warnings for the reciprocating compressor based on the fault diagnosis report.

[0158] Other embodiments or specific implementations of the reciprocating compressor fault early warning system of the present invention can be referred to the above-described method embodiments, and will not be repeated here.

[0159] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0160] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0161] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0162] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for early warning of faults in a reciprocating compressor, characterized in that, The method includes the following steps: acquiring multi-source sensor data of a reciprocating compressor, and preprocessing the multi-source sensor data to obtain a standardized data frame sequence set; extracting time-domain features, frequency-domain features, angular-domain order features, and displacement features based on the standardized data frame sequence set; fusing and normalizing the extracted time-domain features, frequency-domain features, angular-domain order features, and displacement features; inputting the fused and normalized multi-source feature vector into a fault diagnosis model to output a fault diagnosis report, wherein the fault diagnosis model is constructed based on a combination of a mechanism rule engine and a CNN-LSTM attention network; providing fault warning for the reciprocating compressor based on the fault diagnosis report; and inputting the fused and normalized multi-source feature vector into the fault diagnosis model. The diagnostic model outputs a fault diagnosis report, including: inputting a fused and normalized multi-source feature vector into the fault diagnosis model; matching the fused and normalized features using a mechanism rule engine based on a mechanism determination rule base to obtain a basic mechanism probability allocation; outputting the predicted probabilities of various faults through a fully connected layer based on a CNN-LSTM attention network according to the fused and normalized features; calculating the predicted probabilities of various faults and their corresponding probability distributions using a confidence factor to obtain a basic deep learning probability allocation; obtaining the fault diagnosis result based on the basic mechanism probability allocation and the basic deep learning probability allocation using the Dempster combination rule; and outputting a fault diagnosis report based on the fault diagnosis result through the fault diagnosis model.

2. The method as described in claim 1, characterized in that, The preprocessing of the multi-source sensor data to obtain a standardized data frame sequence includes: detecting outliers in the multi-source sensor data and repairing the outliers using cubic spline interpolation; extracting crankcase vibration acceleration, crosshead slide vibration acceleration, cylinder vibration acceleration, piston rod directional displacement, and key phase signal from the repaired multi-source sensor data; performing adaptive digital filtering on the crankcase vibration acceleration, crosshead slide vibration acceleration, and cylinder vibration acceleration respectively; aligning the filtered crankcase vibration acceleration, filtered crosshead slide vibration acceleration, filtered cylinder vibration acceleration, piston rod directional displacement, and key phase signal to a time reference; and encapsulating the time-reference-aligned multi-source sensor data into a standardized data frame sequence set.

3. The method as described in claim 1, characterized in that, Temporal feature extraction based on the standardized data frame sequence set includes: extracting standardized data frame sequence subsets corresponding to crankcase vibration sensors, crosshead slide vibration, and cylinder vibration sensors from the standardized data frame sequence set; generating a crankcase effective measurement value sequence based on the standardized data frame sequence subset corresponding to the crankcase vibration sensors; generating a crosshead slide effective measurement value sequence based on the standardized data frame sequence subset corresponding to the crosshead slide vibration; generating a cylinder effective measurement value sequence based on the standardized data frame sequence subset corresponding to the cylinder vibration sensors; calculating the mean, effective value, peak value, peak factor, and kurtosis based on the crankcase effective measurement value sequence, the crosshead slide effective measurement value sequence, and the cylinder effective measurement value sequence, respectively; and performing temporal feature extraction based on the mean, the effective value, the peak value, the peak factor, and the kurtosis.

4. The method as described in claim 1, characterized in that, Frequency domain feature extraction based on the standardized data frame sequence set includes: obtaining the crankcase effective measurement value sequence, the crosshead slide effective measurement value sequence, and the cylinder effective measurement value sequence based on the standardized data frame sequence set; performing Fourier transform on the crankcase effective measurement value sequence, the crosshead slide effective measurement value sequence, and the cylinder effective measurement value sequence, respectively; determining the fundamental frequency, harmonic amplitude, total harmonic distortion, and spectral centroid based on the transformed spectrum; and performing frequency domain feature extraction based on the fundamental frequency, the harmonic amplitude, the total harmonic distortion, and the spectral centroid.

5. The method as described in claim 1, characterized in that, Angular domain order feature extraction based on the standardized data frame sequence set includes: extracting a subset of standardized data frame sequences of the bond phase signal from the standardized data frame sequence set; determining the total number of bond phase pulses based on the subset of standardized data frame sequences of the bond phase signal; calculating the angular domain signal based on the total number of bond phase pulses; performing a discrete Fourier transform on the angular domain signal and determining different order amplitudes based on the angular domain signal after the discrete Fourier transform; and extracting angular domain order features based on the different order amplitudes.

6. The method as described in claim 1, characterized in that, The process of fusing and normalizing the extracted time-domain features, frequency-domain features, angular-domain order features, and displacement features includes: fusing the extracted time-domain features, frequency-domain features, angular-domain order features, and displacement features using a weighted average method; and normalizing the fused multi-source features using a normalization algorithm.

7. A reciprocating compressor fault early warning system, applied to the reciprocating compressor fault early warning method as described in claim 1, characterized in that, The system includes: a data preprocessing module for acquiring multi-source sensor data from the reciprocating compressor and preprocessing the multi-source sensor data to obtain a standardized data frame sequence set; a feature extraction module for extracting time-domain features, frequency-domain features, angular-domain order features, and displacement features based on the standardized data frame sequence set; a fusion and normalization module for fusing and normalizing the extracted time-domain features, frequency-domain features, angular-domain order features, and displacement features; a fault detection module for inputting the fused and normalized multi-source feature vectors into a fault diagnosis model and outputting a fault diagnosis report, wherein the fault diagnosis model is constructed based on a combination of a mechanism rule engine and a CNN-LSTM attention network; and a fault early warning module for providing fault early warning for the reciprocating compressor based on the fault diagnosis report.

8. A fault early warning device for a reciprocating compressor, characterized in that, The device includes: a memory, a processor, and a reciprocating compressor fault warning program stored in the memory and executable on the processor, the reciprocating compressor fault warning program being configured to implement the steps of the reciprocating compressor fault warning method as described in any one of claims 1 to 6.

9. A storage medium, characterized in that, The storage medium stores a reciprocating compressor fault early warning program, which, when executed by a processor, implements the steps of the reciprocating compressor fault early warning method as described in any one of claims 1 to 6.

Images