Reciprocating compressor fault diagnosis method, device, equipment and storage medium

Abstract

The application relates to the technical field of fault diagnosis, and discloses a reciprocating compressor fault diagnosis method, device, equipment and storage medium, leakage correlation parameters in a cylinder of a target reciprocating compressor are acquired, then a non-steady-state fluctuation baseline formed by non-steady-state fluctuation is fitted in a local weighted regression mode, non-steady-state fluctuation noise is stripped through feature decoupling, pure fault features are extracted, the features are divided according to working condition stages and are adapted to a load sensing model, time sequence feature capturing is strengthened, finally, key features are focused through double-stage attention, a working condition compensation dynamic adjustment classification threshold value is combined, and a diagnosis result is output. Therefore, the application realizes the collaborative optimization of fluctuation noise separation, weak feature extraction and accurate classification of the start-stop transition working condition of the reciprocating compressor based on the combination of feature decoupling stripping of non-steady-state fluctuation noise and working condition stage division and load sensing BiGRU, and solves the problem that a traditional diagnosis method cannot adapt to the non-steady-state characteristics of the start-stop transition working condition.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and in particular to a method, apparatus, equipment and storage medium for diagnosing faults in reciprocating compressors. Background Technology

[0002] Reciprocating compressors, as core power equipment in industrial settings such as long-distance pipelines and chemical plants, undertake the critical tasks of gas compression and transportation. Their operational stability directly affects the safety and efficiency of the entire production system. In actual operation, compressors need to frequently respond to process load fluctuations and undergo start-stop transition conditions, i.e., the load rapidly increases from 0 to 30% of the rated load within 10 seconds. During this process, the gas compression inside the cylinder is in a highly unsteady state, and core thermodynamic parameters such as intake pressure, exhaust pressure, and cylinder wall temperature exhibit drastic nonlinear fluctuations. Cylinder leakage is the most common thermodynamic fault under this condition. Early faults with leakage amounts of only 3%-8% of the rated value are easily masked by unsteady-state fluctuation noise with amplitudes reaching 20% ​​of the rated value. Furthermore, the trends of multiple parameter changes are coupled, making fault characteristics difficult to identify.

[0003] Existing fault diagnosis methods have significant limitations: traditional fixed threshold methods rely on the range of steady-state operating parameters and cannot adapt to unsteady fluctuations; simple time series models (such as ordinary LSTM) do not consider the phase characteristics of start-stop conditions and are difficult to separate fluctuation noise from fault features; some machine learning-based methods are not optimized for unsteady data and have insufficient ability to identify weak coupling features, resulting in extremely low accuracy in diagnosing cylinder leakage under start-stop transition conditions, failing to effectively warn of early faults, and easily causing safety accidents such as equipment damage and production interruption. There is an urgent need for a precise diagnostic technology solution that adapts to the characteristics of start-stop transition conditions. Summary of the Invention

[0004] The present invention provides a method, apparatus, device and storage medium for diagnosing faults in reciprocating compressors, aiming to solve at least one of the above-mentioned technical problems.

[0005] To achieve the above objectives, the present invention provides a method for diagnosing faults in a reciprocating compressor, comprising the following steps:

[0006] Obtain the cylinder leakage correlation parameters of the target reciprocating compressor, construct the original parameter matrix, and preprocess the original parameter matrix to obtain the standardized parameter matrix;

[0007] Using the standardized parameter matrix as input, a local neighborhood window is allocated, and the non-steady-state fluctuation baseline formed by the non-steady-state fluctuation is fitted by a local weighted regression method. The fault feature matrix is ​​generated using the standardized parameter matrix and the non-steady-state fluctuation baseline.

[0008] Based on the fault feature matrix, an adaptive division of operating conditions is performed, and each operating condition is aligned, enhanced, and spliced. The spliced ​​features are then input into the load-aware BiGRU model to obtain the high-dimensional feature matrix output by the load-aware BiGRU model.

[0009] Using a stage attention calculation mechanism, the optimal feature vector is obtained by filtering and weighted fusion from the high-dimensional feature matrix. The optimal feature vector is then input into a multi-label classifier that integrates the operating condition compensation factor to obtain the fault diagnosis result of the reciprocating compressor.

[0010] Optionally, the step of obtaining the cylinder leakage correlation parameters of the target reciprocating compressor and constructing the original parameter matrix specifically includes:

[0011] Obtain cylinder leakage correlation parameters of the target reciprocating compressor; wherein, the cylinder leakage correlation parameters include the intake pressure sequence, exhaust pressure sequence and cylinder wall temperature sequence collected by sensors respectively deployed at the compressor intake port, compressor exhaust port and cylinder wall;

[0012] Based on a preset sampling frequency, preset cycle duration, and preset target number of reciprocating compressor start-stop cycles, data from several sampling points in the intake pressure sequence, exhaust pressure sequence, and cylinder wall temperature sequence are extracted to construct an original parameter matrix. The specific expression is as follows:

[0013] ;

[0014] In the formula, Represents the original parameter matrix, For the first The original value of inhalation pressure at all times. For the first The original value of the exhaust pressure at any time. For the first The original value of the cylinder wall temperature at any given time. Indicates the number of sampling points. .

[0015] Optionally, the steps of allocating a local neighborhood window using the standardized parameter matrix as input, fitting the non-steady-state fluctuation baseline formed by the non-steady-state fluctuation using a local weighted regression method, and generating a fault feature matrix using the standardized parameter matrix and the non-steady-state fluctuation baseline specifically include:

[0016] Based on the load change phases of the target reciprocating compressor during its start-stop cycle, for each moment... Dynamically allocate local neighborhood windows and use a Gaussian kernel function to calculate the time interval within the window. Based on the weights, construct a local weight matrix;

[0017] Using the standardized parameter matrix as input, combined with each time step... Using the local neighborhood window and local weight matrix, a local weighted regression algorithm is employed for each time step. No. The parameters are fitted to the unsteady-state fluctuation baseline to obtain the fluctuation baseline fitting value for each parameter;

[0018] The specific expression for the baseline fitting value of each parameter is as follows:

[0019]

[0020] In the formula, For the first Time of the first The baseline of non-steady-state fluctuations of each parameter For the first The local weight matrix at time t, For the first The parameter in the first... Local data matrix within the time-neighbor window;

[0021] Each moment The The standardized parameter values ​​are subtracted from the fluctuation baseline fitting values ​​to remove unsteady fluctuation noise, decouple the cylinder internal leakage fault characteristics, and integrate them to form a fault feature matrix.

[0022] The specific expression for decoupling the cylinder internal leakage fault characteristics is as follows:

[0023]

[0024] In the formula, For the first Time of the first The fault characteristic values ​​after decoupling of each parameter, For the first Time of the first The standardized parameter values ​​of each parameter.

[0025] Optionally, the load change phases of the target reciprocating compressor during the start-stop cycle specifically include: the initial rise phase (0-3s), the rapid rise phase (3-7s), and the stabilization phase (7-10s); at each moment... The dynamically allocated local neighborhood windows specifically include: a local neighborhood window with 50 data points for the initial and rapid rise phases, and a local neighborhood window with 30 data points for the stable phase.

[0026] Optionally, the steps of performing adaptive segmentation of operating conditions based on the fault feature matrix, aligning, enhancing, and stitching each operating condition stage, and inputting the stitched features into the load-aware BiGRU model to obtain the high-dimensional feature matrix output by the load-aware BiGRU model specifically include:

[0027] Based on the load change stages of the target reciprocating compressor during its start-stop cycle, the decoupled fault feature matrix is ​​divided into stages, resulting in feature matrices for three stages. , , For the characteristic matrix , , Perform alignment, enhancement, and splicing processes;

[0028] Obtain the load normalization value of the target reciprocating compressor during the start-stop cycle, calculate the load-related gating weight at each time i using the Sigmoid activation function, and construct a load-aware BiGRU model with dynamic gating adjustment using the load-related gating weight.

[0029] The load-related gating weights include the load-related weights of the update gate and the load-related weights of the reset gate, and the specific expression is as follows:

[0030]

[0031]

[0032] In the formula, To update the load-related weights of the gate, To reset the load-related weights of the door, For the first Normalized load value at time 10:00 This is the load weight matrix. For bias terms, Use the Sigmoid activation function;

[0033] splicing features Input the load-aware BiGRU model and combine it with load-related gating weights to obtain the high-dimensional feature matrix output by the load-aware BiGRU model.

[0034] Optionally, the splicing features The steps for obtaining the high-dimensional feature matrix output by the load-sensing BiGRU model, which involves inputting the load-sensing BiGRU model and combining it with load-related gating weights, specifically include:

[0035] splicing features Input the load-aware BiGRU model, update the update gate and reset gate of the load-aware model by combining the load-related gating weights, and update the forward hidden state of the forward GRU and the backward hidden state of the backward GRU in the load-aware BiGRU model based on the gating output.

[0036] The expressions for the update gate and reset gate for updating load awareness are as follows:

[0037]

[0038]

[0039] The update expressions for the forward hidden state and the backward hidden state are as follows:

[0040]

[0041]

[0042] In the formula, For the first The decoupled eigenvector at time step 1. For the first Load-related gating weights at any given time. The state was hidden in the previous moment. For load sensing, update the gate and reset the gate output. This is the weight matrix. For bias terms, For element-wise multiplication, In the candidate hidden state, This is the hidden state updated at the current moment;

[0043] The forward hidden state and the backward hidden state are concatenated to form a bidirectional fused high-dimensional feature matrix, which is used as the output of the load-aware BiGRU model. The specific expression is as follows:

[0044] ;

[0045] In the formula, Forward GRU Always hide your status; For the backward GRU Always hide your status; The hidden state after bidirectional fusion, the hidden state after bidirectional fusion at all time points. Constructing a high-dimensional feature matrix .

[0046] Optionally, the step of using a stage attention computation mechanism to filter and weightedly fuse features from the high-dimensional feature matrix to obtain the optimal feature vector, and then inputting the optimal feature vector into a multi-label classifier that fuses operating condition compensation factors to obtain the reciprocating compressor fault diagnosis result, specifically includes:

[0047] High-dimensional feature matrix The process is divided into stages. Calculate the mean vector of the hidden states at each stage. The importance weights of each stage are calculated using an attention scoring function. ;

[0048] The feature matrices of each stage With the corresponding stage attention weights Perform a weighted summation to generate a feature matrix representing the importance of the fusion stage. For stage-weighted feature matrix Calculate the attention score for the hidden state at each time step, and obtain the attention weights at each time step using the Softmax function. ;

[0049] The stage-weighted feature matrix Hidden states at each time step and corresponding attention weights at each time step Perform weighted summation to generate the globally optimal feature vector. ;

[0050] Based on the load sequence of each stage, the load standard deviation is calculated and converted into the operating condition compensation factor. Based on the baseline threshold of each leakage level and the operating condition compensation factor, a dynamically adjusted classification threshold is generated.

[0051] The global optimal eigenvector Input a multi-label classifier, pass it through a fully connected layer and a sigmoid activation function, and output the predicted probability of each leakage level. Using the predicted probability By dynamically adjusting the classification threshold, fault diagnosis results for reciprocating compressors are generated.

[0052] Furthermore, to achieve the above objectives, the present invention also provides a reciprocating compressor fault diagnosis device, comprising:

[0053] A construction module is used to obtain the cylinder leakage correlation parameters of the target reciprocating compressor, construct the original parameter matrix, and preprocess the original parameter matrix to obtain the standardized parameter matrix;

[0054] The generation module is used to allocate local neighborhood windows with the standardized parameter matrix as input, fit the non-steady-state fluctuation baseline formed by non-steady-state fluctuations using local weighted regression, and generate a fault feature matrix using the standardized parameter matrix and the non-steady-state fluctuation baseline.

[0055] The segmentation module is used to perform adaptive segmentation of operating conditions based on the fault feature matrix, perform alignment, enhancement and splicing processing on each operating condition stage, input the spliced ​​features into the load-aware BiGRU model, and obtain the high-dimensional feature matrix output by the load-aware BiGRU model.

[0056] The diagnostic module is used to filter and weightedly fuse the high-dimensional feature matrix using a stage attention computing mechanism to obtain the optimal feature vector. The optimal feature vector is then input into a multi-label classifier that incorporates the operating condition compensation factor to obtain the fault diagnosis result of the reciprocating compressor.

[0057] In addition, to achieve the above objectives, the present invention also provides a reciprocating compressor fault diagnosis device, the reciprocating compressor fault diagnosis device comprising: a memory, a processor, and a reciprocating compressor fault diagnosis program stored in the memory and executable on the processor, wherein when the reciprocating compressor fault diagnosis program is executed by the processor, it implements the steps of the reciprocating compressor fault diagnosis method as described above.

[0058] In addition, to achieve the above objectives, the present invention also provides a storage medium storing a reciprocating compressor fault diagnosis program, wherein the reciprocating compressor fault diagnosis program, when executed by a processor, implements the steps of the reciprocating compressor fault diagnosis method described above.

[0059] The beneficial effects of this invention are as follows: It proposes a method, device, equipment, and storage medium for fault diagnosis of reciprocating compressors. By acquiring the cylinder leakage correlation parameters of the target reciprocating compressor, a standardized parameter matrix is ​​obtained after preprocessing. Then, a local weighted regression method is used to fit the unsteady-state fluctuation baseline formed by unsteady-state fluctuations. Using the standardized parameter matrix and the unsteady-state fluctuation baseline, unsteady-state fluctuation noise is removed through feature decoupling, and pure fault features are extracted. Next, features are divided according to operating condition stages and adapted to a load-sensing model to enhance the capture of temporal features. Finally, key features are focused through two-level attention, and the classification threshold is dynamically adjusted by combining operating condition compensation to output the diagnostic results. Therefore, this invention, based on feature decoupling to remove unsteady-state fluctuation noise and the combination of operating condition stage division and load-sensing BiGRU, achieves synergistic optimization of fluctuation noise separation, weak feature extraction, and accurate classification in the start-stop transition operating conditions of reciprocating compressors, solving the problem that traditional diagnostic methods cannot adapt to the unsteady characteristics of start-stop transition operating conditions. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the device structure of the hardware operating environment involved in the embodiments of the present invention;

[0061] Figure 2 This is a flowchart illustrating an embodiment of the reciprocating compressor fault diagnosis method of the present invention;

[0062] Figure 3 This is a structural block diagram of a reciprocating compressor fault diagnosis device according to an embodiment of the present invention. Detailed Implementation

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

[0064] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0065] like Figure 1 As shown, Figure 1 This is a schematic diagram of the device structure of the hardware operating environment involved in the embodiments of the present invention.

[0066] like Figure 1 As shown, the device may include: a processor 1001, such as a 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 Wi-Fi interface). The memory 1005 may be high-speed RAM or non-volatile memory, such as a disk drive. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0067] Those skilled in the art will understand that Figure 1 The structure of the device shown does not constitute a limitation on the device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0068] like Figure 1 As shown, the memory 1005, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and a reciprocating compressor fault diagnosis program.

[0069] exist Figure 1 In the terminal shown, network interface 1004 is mainly used to connect to the backend server and communicate data with it; user interface 1003 is mainly used to connect to the client (user terminal) and communicate data with it; while processor 1001 can be used to call the reciprocating compressor fault diagnosis program stored in memory 1005 and perform the following operations:

[0070] Obtain the cylinder leakage correlation parameters of the target reciprocating compressor, construct the original parameter matrix, and preprocess the original parameter matrix to obtain the standardized parameter matrix;

[0071] Using the standardized parameter matrix as input, a local neighborhood window is allocated, and the non-steady-state fluctuation baseline formed by the non-steady-state fluctuation is fitted by a local weighted regression method. The fault feature matrix is ​​generated using the standardized parameter matrix and the non-steady-state fluctuation baseline.

[0072] Based on the fault feature matrix, an adaptive division of operating conditions is performed, and each operating condition is aligned, enhanced, and spliced. The spliced ​​features are then input into the load-aware BiGRU model to obtain the high-dimensional feature matrix output by the load-aware BiGRU model.

[0073] Using a stage attention calculation mechanism, the optimal feature vector is obtained by filtering and weighted fusion from the high-dimensional feature matrix. The optimal feature vector is then input into a multi-label classifier that integrates the operating condition compensation factor to obtain the fault diagnosis result of the reciprocating compressor.

[0074] The specific embodiments of the present invention applied to the device are basically the same as the embodiments of the reciprocating compressor fault diagnosis method described below, and will not be repeated here.

[0075] This invention provides a method for diagnosing faults in a reciprocating compressor, referring to... Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the reciprocating compressor fault diagnosis method of the present invention.

[0076] In this embodiment, a method for diagnosing faults in a reciprocating compressor includes the following steps:

[0077] S100: Obtain the cylinder leakage correlation parameters of the target reciprocating compressor, construct the original parameter matrix, preprocess the original parameter matrix, and obtain the standardized parameter matrix;

[0078] S200: Assign a local neighborhood window using the standardized parameter matrix as input, fit the unsteady fluctuation baseline formed by the unsteady fluctuation using a local weighted regression method, and generate a fault feature matrix using the standardized parameter matrix and the unsteady fluctuation baseline.

[0079] S300: Based on the fault feature matrix, perform adaptive segmentation of operating conditions, perform alignment, enhancement and splicing processing on each operating condition stage, input the spliced ​​features into the load-aware BiGRU model, and obtain the high-dimensional feature matrix output by the load-aware BiGRU model.

[0080] S400: Using a stage attention calculation mechanism, the optimal feature vector is obtained by filtering and weighted fusion from the high-dimensional feature matrix. The optimal feature vector is then input into a multi-label classifier that integrates the working condition compensation factors to obtain the reciprocating compressor fault diagnosis result.

[0081] It should be noted that existing fault diagnosis methods have significant limitations: traditional fixed threshold methods rely on the range of steady-state operating parameters and cannot adapt to unsteady fluctuations; simple time series models (such as ordinary LSTM) do not consider the phase characteristics of start-stop conditions and are difficult to separate fluctuation noise from fault features; some machine learning-based methods are not optimized for unsteady data and have insufficient ability to identify weak coupling features, resulting in extremely low accuracy in diagnosing cylinder leakage under start-stop transition conditions, failing to effectively warn of early faults, and easily causing safety accidents such as equipment damage and production interruption.

[0082] To address the aforementioned issues, this embodiment acquires the cylinder leakage correlation parameters of the target reciprocating compressor. Using these parameters as the core, a standardized parameter matrix is ​​obtained after preprocessing. Then, a local weighted regression method is used to fit the non-steady-state fluctuation baseline. Utilizing the standardized parameter matrix and the non-steady-state fluctuation baseline, non-steady-state fluctuation noise is removed through feature decoupling, and pure fault features are extracted. Next, features are divided according to operating conditions and adapted to the load perception model to enhance the capture of time-series features. Finally, key features are focused through dual-level attention, and the classification threshold is dynamically adjusted in conjunction with operating condition compensation to output diagnostic results.

[0083] Therefore, this invention, based on feature decoupling to remove unsteady-state fluctuation noise and the combination of operating condition stage division and load sensing BiGRU, integrates the characteristics of operating condition stages into the sensing model, realizing the coordinated optimization of fluctuation noise separation, weak feature extraction and accurate classification in the start-stop transition operating condition of reciprocating compressors, and solving the problem that traditional diagnostic methods cannot adapt to the unsteady-state characteristics of start-stop transition operating conditions.

[0084] In a preferred embodiment, the step of obtaining the cylinder leakage correlation parameters of the target reciprocating compressor and constructing the original parameter matrix specifically includes:

[0085] S110: Obtain the cylinder leakage correlation parameters of the target reciprocating compressor; wherein, the cylinder leakage correlation parameters include the intake pressure sequence, exhaust pressure sequence and cylinder wall temperature sequence collected by sensors respectively installed at the compressor intake port, compressor exhaust port and cylinder wall;

[0086] S120: According to the preset sampling frequency, preset cycle duration, and preset target number of reciprocating compressor start-stop cycles, extract data from several sampling points in the intake pressure sequence, exhaust pressure sequence, and cylinder wall temperature sequence respectively, and construct the original parameter matrix. The specific expression is as follows:

[0087] ;

[0088] In the formula, Represents the original parameter matrix, For the first The original value of inhalation pressure at all times. For the first The original value of the exhaust pressure at any time. For the first The original value of the cylinder wall temperature at any given time. Indicates the number of sampling points. .

[0089] In this embodiment, considering the direct impact of cylinder leakage on gas pressure and temperature (leakage leads to low intake pressure, insufficient exhaust pressure, and increased cylinder wall temperature), three core parameters are selected. These parameters are then synchronously collected using high-precision sensors. Data quality is ensured by setting a high sampling frequency (to adapt to high-frequency fluctuations) and multiple cycles of collection (to improve data representativeness). Finally, all collected data is matrix-integrated. For example, with a sampling frequency of 200Hz, a single cycle duration of 10s, and a target reciprocating compressor start-stop cycle count of 3, the number of sampling points M is 200×10×3=6000.

[0090] After constructing the original parameter matrix, it is necessary to preprocess the original parameter matrix to obtain a standardized parameter matrix. The preprocessing includes: using the 3σ criterion to quickly remove extreme outliers caused by sudden sensor failures, and then using Min-Max standardization to eliminate the dimensional differences between pressure (MPa) and temperature (K), while retaining the amplitude relationship of non-steady-state fluctuations, providing a standardized data foundation for subsequent feature decoupling.

[0091] In a preferred embodiment, the steps of allocating a local neighborhood window using the standardized parameter matrix as input, fitting the unsteady fluctuation baseline formed by the unsteady fluctuation using local weighted regression, and generating a fault feature matrix using the standardized parameter matrix and the unsteady fluctuation baseline specifically include:

[0092] S210: Based on the load change phase of the target reciprocating compressor during the start-stop cycle, for each moment... Dynamically allocate local neighborhood windows and use a Gaussian kernel function to calculate the time interval within the window. Based on the weights, construct a local weight matrix;

[0093] S220: Using the standardized parameter matrix as input, combined with each time step... Using the local neighborhood window and local weight matrix, a local weighted regression algorithm is employed for each time step. No. The parameters are fitted to the unsteady-state fluctuation baseline to obtain the fluctuation baseline fitting value for each parameter;

[0094] The specific expression for the baseline fitting value of each parameter is as follows:

[0095]

[0096] In the formula, For the first Time of the first The baseline of non-steady-state fluctuations of each parameter For the first The local weight matrix at time t, For the first The parameter in the first... Local data matrix within the time-neighbor window;

[0097] S230: Each moment The The standardized parameter values ​​are subtracted from the fluctuation baseline fitting values ​​to remove unsteady fluctuation noise, decouple the cylinder internal leakage fault characteristics, and integrate them to form a fault feature matrix.

[0098] The specific expression for decoupling the cylinder internal leakage fault characteristics is as follows:

[0099]

[0100] In the formula, For the first Time of the first The fault characteristic values ​​after decoupling of each parameter, For the first Time of the first The standardized parameter values ​​of each parameter.

[0101] In this embodiment, the window size is first dynamically adjusted based on the operating condition stage to ensure that the baseline fits the fluctuation characteristics of different stages. Then, the Gaussian kernel function is used to give higher weight to the recent data within the window to improve the accuracy of the baseline in tracking the fluctuation trend at the current moment. Finally, the original standardized signal is used to subtract the fluctuation baseline to separate the fault feature components related to leakage and integrate them to form a fault feature matrix.

[0102] For the baseline fit value of fluctuation for each parameter, this embodiment is based on locally weighted linear regression (LWR), using a weight matrix... The data within the window is weighted and fitted to the normal fluctuation trend (baseline) at the current moment. This ensures that the baseline only reflects non-steady-state load fluctuations while avoiding interference from fault characteristics. For decoupling the cylinder leakage fault characteristics, this embodiment is based on the signal superposition principle. Considering that the standardized signal consists of a non-steady-state fluctuation baseline and fault characteristics, the baseline (noise) is stripped away through subtraction, retaining the fault characteristic components.

[0103] Therefore, this invention proposes a feature decoupling method that combines dynamic local neighborhood windowing with Gaussian kernel weighted regression. This method adapts to the stage fluctuation characteristics of start-stop conditions and achieves accurate separation of unsteady-state fluctuation noise and fault features. It completely solves the core technical pain point that unsteady-state fluctuation noise masks weak fault features under start-stop conditions. By using dynamic windowing and weighted regression to improve baseline fitting accuracy, high-purity fault features are obtained after removing noise, providing reliable input for subsequent identification.

[0104] In a preferred embodiment, the load change phases of the target reciprocating compressor during the start-stop cycle specifically include: the initial rise phase (0-3s), the rapid rise phase (3-7s), and the stabilization phase (7-10s); at each moment... The dynamically allocated local neighborhood windows specifically include: a local neighborhood window with 50 data points for the initial and rapid rise phases, and a local neighborhood window with 30 data points for the stable phase.

[0105] In this embodiment, the load change patterns during start-stop operations are divided into stages based on the time pattern (0-3s rise, 3-7s rapid rise, 7-10s stabilization). A large window of 50 data points is allocated to address the strong fluctuations in the initial rise and stabilization phases, ensuring a good fit to the overall fluctuation trend. A smaller window of 30 data points is allocated to address the concentrated fault characteristics and gentle fluctuations during the rapid rise phase, preventing overfitting that could lead to the loss of fault characteristics. Based on the load change patterns during start-stop operations, the entire cycle is divided into three characteristic stages: initial rise, rapid rise, and stabilization, achieving refined adaptation of the operating conditions.

[0106] Specifically, the time pattern of load change during the start-stop operation of the reciprocating compressor is as follows: (1) Initial rise (0-3s): The load starts to rise from 0, the parameters fluctuate violently, the fault characteristics are scattered, and a large window is needed for fitting. (2) Rapid rise (3-7s): The load rises rapidly to 30% of the rated load, the parameters fluctuate gently, the fault characteristics are concentrated, and a small window is needed to retain details. (3) Stable period (7-10s): The load is maintained at 30% of the rated load, the parameters fluctuate less but there is steady-state drift, and a large window is needed to capture the trend.

[0107] Therefore, this invention dynamically allocates differentiated local neighborhood window sizes based on the fluctuation intensity and fault feature distribution at different stages, breaking through the limitations of traditional fixed windows and solving the problem of insufficient adaptability of traditional fixed windows in different operating conditions. Through stage division and dynamic window allocation, the baseline fitting captures the overall trend in the stage of violent fluctuations (early rise and stable period) and retains details in the stage of concentrated fault features (rapid rise period), thereby improving the accuracy of baseline fitting and the purity of fault feature extraction.

[0108] In a preferred embodiment, the steps of performing adaptive segmentation of operating conditions based on the fault feature matrix, aligning, enhancing, and stitching each operating condition stage, and inputting the stitched features into the load-aware BiGRU model to obtain the high-dimensional feature matrix output by the load-aware BiGRU model specifically include:

[0109] S310: Based on the load change stages of the target reciprocating compressor during the start-stop cycle, the decoupled fault feature matrix is ​​divided into stages to obtain feature matrices for three stages. , , For the characteristic matrix , , Alignment, enhancement, and splicing processes are performed.

[0110] Specifically, based on the load change curves of the start-stop transition conditions (0-3s for the initial load rise, 3-7s for the rapid rise, and 7-10s for the steady-state period), the resulting decoupled fault characteristic matrix is ​​analyzed. The process involves dividing the data into stages, resulting in three sub-segment feature matrices. (0-3s, 200Hz×3s=600 data points) (3-7s, 200Hz×4s=800 data points) (7-10s, 200Hz × 3s = 600 data points). Then, each segment feature is zero-padded or truncated to a uniform length of 800 data points. Local standard deviation normalization is used to enhance feature differences within each stage, resulting in an aligned stage feature matrix. .

[0111] (1) Normalization formula for stage characteristics:

[0112] ;

[0113] (2) Zero-fill / truncation formula (in) (For example)

[0114] ;

[0115] In the formula, For the first Phase 1 Time of the first Decoupling eigenvalues ​​of each parameter For the first Phase 1 The mean of each parameter, Standard deviation To avoid tiny values ​​with a denominator of 0, For the first The eigenvalues ​​after stage alignment These correspond to the initial stage of growth, the rapid growth stage, and the stable stage, respectively.

[0116] S320: Obtain the load normalization value of the target reciprocating compressor during the start-stop cycle, calculate the load-related gating weight at each time i using the Sigmoid activation function, and construct a load-aware BiGRU model with dynamic gating adjustment using the load-related gating weight.

[0117] The load-related gating weights include the load-related weights of the update gate and the load-related weights of the reset gate, and the specific expression is as follows:

[0118]

[0119]

[0120] In the formula, To update the load-related weights of the gate, To reset the load-related weights of the door, For the first Normalized load value at time 10:00 This is the load weight matrix. For bias terms, This is the Sigmoid activation function.

[0121] S330: Features of splicing Input the load-aware BiGRU model and combine it with load-related gating weights to obtain the high-dimensional feature matrix output by the load-aware BiGRU model.

[0122] In this embodiment, the fault feature matrix is ​​first divided according to the operating stage. The feature length is unified by alignment processing and then spliced ​​to form a continuous time-series input. Then, the gating weight is calculated based on the load normalization value, so that the update gate is more active in the rapid rise period of high load change rate (enhancing feature capture) and the reset gate is more active in the stage of low load change rate (suppressing redundant information). Finally, the spliced ​​features are input into the load-aware BiGRU, and the high-dimensional feature matrix of time-dependent fusion stage is output.

[0123] Specifically, the three aligned stage feature matrices obtained Assembled in chronological order An improved load-aware BiGRU model is input. The model introduces load-related gating weights, dynamically adjusting the sensitivity of the update and reset gates based on the current load value. Higher feature capture weights are assigned to the rapid rise phase (high load change rate), while weights are appropriately reduced for the initial rise and stable phases (low load change rate). The model outputs forward and backward hidden states respectively, which, after fusion, yield a high-dimensional feature matrix containing stage-time dependencies. .

[0124] Specifically, regarding the load-related weights for the update gate and the reset gate, this embodiment considers normalizing the load value using the Sigmoid function. The weights are mapped to the 0-1 interval to achieve a linear relationship between load change and gating sensitivity. The higher the load change rate, the larger the update gate weight and the smaller the reset gate weight. The load normalization value... (Obtained by dividing the actual load value by 30% of the rated load), it can directly reflect the intensity of load changes in the current stage and fit the dynamic characteristics of the load under start-up and shutdown conditions.

[0125] Therefore, this embodiment proposes an adaptive division of operating conditions and binds the load change rate with the GRU gating mechanism when constructing the load-aware BiGRU model, thereby achieving dynamic adjustment of the sensitivity of feature capture in different stages. This solves the problem of insufficient adaptation of traditional GRU models to the characteristics of operating conditions. Through stage processing, the model focuses on the feature differences of different stages, and through load-aware gating, the feature capture of key stages (rapid rise period) is strengthened, thereby improving the pertinence and effectiveness of time series feature extraction.

[0126] In a preferred embodiment, the splicing feature The steps for obtaining the high-dimensional feature matrix output by the load-sensing BiGRU model, which involves inputting the load-sensing BiGRU model and combining it with load-related gating weights, specifically include:

[0127] S331: Features of splicing Input the load-aware BiGRU model, update the update gate and reset gate of the load-aware model by combining the load-related gating weights, and update the forward hidden state of the forward GRU and the backward hidden state of the backward GRU in the load-aware BiGRU model based on the gating output.

[0128] The expressions for the update gate and reset gate for updating load awareness are as follows:

[0129]

[0130]

[0131] The update expressions for the forward hidden state and the backward hidden state are as follows:

[0132]

[0133]

[0134] In the formula, For the first The decoupled eigenvector at time step 1. For the first Load-related gating weights at any given time. The state was hidden in the previous moment. For load sensing, update the gate and reset the gate output. This is the weight matrix. For bias terms, For element-wise multiplication, In the candidate hidden state, This is the hidden state updated at the current moment;

[0135] S332: The forward hidden state and the backward hidden state are concatenated to form a bidirectional fused high-dimensional feature matrix, which is used as the output of the load-aware BiGRU model. The specific expression is as follows:

[0136]

[0137] In the formula, Forward GRU Always hide your status; For the backward GRU Always hide your status; The hidden state after bidirectional fusion, the hidden state after bidirectional fusion at all time points. Constructing a high-dimensional feature matrix .

[0138] In this embodiment, the spliced ​​features are input into the GRU unit, and the gating output is updated by combining load-related weights, making the update gate more active during the rapid rise period and giving priority to accepting new fault features; by resetting the gate, redundant information in non-critical stages is suppressed, the hidden state is optimized, the forward GRU captures the temporal evolution of features (from past to present), and the backward GRU captures the backward correlation of features (from present to past), and the spliced ​​features form a high-dimensional feature matrix containing bidirectional dependencies.

[0139] Specifically, for the updates of the load-aware update gate and reset gate, this embodiment modulates the traditional gating output through load-related weights, so that the gating sensitivity is dynamically adjusted with load changes, strengthening the feature capture of key stages, making the gating output adapt to load changes, enhancing the update gate output during the rapid rise period, and weakening the reset gate output, thereby improving the efficiency of fault feature capture. For the updates of the forward hidden state and the backward hidden state, this embodiment considers the candidate hidden state to reflect the feature information at the current moment by calculating the historical state modulated by the reset gate and the current feature, while the current hidden state balances the historical state and the candidate state through the update gate, so as to achieve smooth updates of temporal features. Finally, when performing bidirectional fusion of the forward hidden state and the backward hidden state, the forward GRU is used to capture the temporal evolution from the past to the present, and the backward GRU is used to capture the backward correlation from the present to the past. After splicing, bidirectional information complementarity is achieved, improving the completeness of feature expression.

[0140] Therefore, this invention proposes a load-aware gated update mechanism that integrates load-related weights into GRU gating computation, enabling hidden state updates to adapt to load changes. Simultaneously, it employs a bidirectional GRU structure to capture both the forward evolution and backward correlation of fault features, improving the completeness of temporal feature representation. Furthermore, bidirectional information fusion is achieved through hidden state concatenation, enhancing the expressive power of high-dimensional features. This invention solves the problems of traditional GRU's unidirectional capture of temporal dependencies and insensitivity to load changes. By using load-aware gating, hidden state updates focus on key stage features, and the bidirectional structure fully captures the temporal correlation of fault features, improving the recognizability of high-dimensional features.

[0141] In a preferred embodiment, the step of using a stage attention calculation mechanism to filter and weightedly fuse the high-dimensional feature matrix to obtain an optimal feature vector, and then inputting the optimal feature vector into a multi-label classifier that fuses operating condition compensation factors to obtain the reciprocating compressor fault diagnosis result, specifically includes:

[0142] S410: High-dimensional feature matrix The process is divided into stages. Calculate the mean vector of the hidden states at each stage. The importance weights of each stage are calculated using an attention scoring function. .

[0143] Specifically, the mean vector Importance weight The specific expression is:

[0144]

[0145]

[0146]

[0147] In the formula, For the first The characteristic matrix of the stage, For the first Mean of hidden state in each stage This is the stage attention weight matrix. For bias terms, For stage attention vectors, For the first Attention weights for each stage (0-1, summing to 1).

[0148] In this embodiment, by compressing stage features using the mean vector and calculating stage importance weights using nonlinear transformation and the Softmax function, the stage with the most concentrated fault features, i.e., the rapid rise phase, can be automatically identified. )of Typically, the highest weight is used to strengthen the features of this stage, thereby achieving coarse screening and focusing, and avoiding interference from features of non-critical stages.

[0149] S420: The feature matrices of each stage With the corresponding stage attention weights Perform a weighted summation to generate a feature matrix representing the importance of the fusion stage. For stage-weighted feature matrix Calculate the attention score for the hidden state at each time step, and obtain the attention weights at each time step using the Softmax function. .

[0150] Specifically, the feature matrix and attention weights at different times The specific expression is:

[0151]

[0152]

[0153]

[0154] In the formula, For stage-weighted feature matrix, For the first Stage attention weights For the first Stage feature matrix, for No. The hidden state at all times Let be the attention weight matrix at time step 1. For bias terms, Let be the attention vector at time step. For the first Attention weights at each moment (0-1, summing to 1).

[0155] In this embodiment, linear fusion of features from different stages is achieved through weighted averages. Stages with larger weights contribute more strongly to the fused features. Then, a nonlinear transformation maps high-dimensional features to a low-dimensional space for scoring, and the Softmax function normalizes the weights. This approach amplifies the significant rapid rise of fault characteristics while retaining useful information from other stages. Furthermore, it allows for further pinpointing the peak fault characteristic moment within critical stages (e.g., when the load reaches 20%-25% of the rated load). Strengthen the characteristics of these moments to achieve precise screening and improve the distinguishability of features.

[0156] S430: Weight the feature matrix for each stage Hidden states at each time step and corresponding attention weights at each time step Perform weighted summation to generate the globally optimal feature vector. .

[0157] Specifically, the globally optimal eigenvector The specific expression is:

[0158]

[0159] In the formula, This is the globally optimal eigenvector. For the first Attention weight at any moment for No. The status is always hidden.

[0160] In this embodiment, by weighting and fusing features within a time-series phase, the most representative feature information of the whole is extracted. This can condense the fault features of key phases and core moments into low-dimensional vectors, which avoids redundancy of full-time-series features and retains the most identifiable fault information.

[0161] S440: Calculate the load standard deviation based on the load sequence for each stage and convert it into a condition compensation factor. Based on the baseline thresholds and operating condition compensation factors for each leakage level, dynamically adjusted classification thresholds are generated. .

[0162] Specifically, operating condition compensation factor Dynamically adjust classification threshold The specific expression is:

[0163]

[0164]

[0165] In the formula, For the first Phase load sequence, For the standard deviation of the load, For adjustment coefficients, The operating condition compensation factor (the more stable the load, the better). The smaller, The closer to 1), These correspond to mild, moderate, and severe leaks, respectively. Baseline threshold (mild) Moderate Severe ), This is the compensation factor for the current operating conditions. This is the dynamically adjusted classification threshold.

[0166] In this embodiment, an exponential function is used to map the degree of load fluctuation to a compensation factor, thereby quantifying the stability of the operating condition. The compensation factor is then used to linearly correct the baseline threshold, balancing diagnostic accuracy and fault tolerance under different operating conditions. This provides a basis for adjusting the classification threshold: (The following is a separate, unrelated sentence: Large load fluctuations in the initial stage of an increase...) Small loads require lowering the threshold to avoid missed detections; stable loads during the steady period ( (For large load fluctuations), the threshold needs to be increased to avoid false alarms and adapt to the dynamic stability differences of the working conditions; it can also appropriately relax the threshold during periods of large load fluctuations to ensure that minor faults are not missed, and strictly control the threshold during periods of stable load to reduce false alarms, thus solving the problem of poor adaptability of traditional fixed thresholds under different working conditions.

[0167] S450: The globally optimal eigenvector Input a multi-label classifier, pass it through a fully connected layer and a sigmoid activation function, and output the predicted probability of each leakage level. Using the predicted probability By dynamically adjusting the classification threshold, fault diagnosis results for reciprocating compressors are generated.

[0168] Specifically, predicting probabilities The specific expression is:

[0169] ;

[0170] In the formula, For the classifier Class weight matrix, For the corresponding bias term, It is the Sigmoid activation function. For the first Predicted probability of leakage.

[0171] In this embodiment, a fully connected layer maps high-dimensional features into probability values, and the Sigmoid function realizes the probability output of multi-label classification. This can transform abstract fault features into intuitive probability values, providing a quantitative basis for fault level determination. At the same time, the magnitude of the probability value can reflect the confidence level of the diagnostic results.

[0172] As is easily understood, this embodiment first calculates the importance weight of each stage through stage attention, and highlights the characteristics of the key stage (rapid rise period) after weighted fusion. Then, it focuses on the core moment features within the stage through time attention to generate the globally optimal feature vector. Based on load stability, it calculates the compensation factor, dynamically adjusts the classification threshold of each leakage level, inputs the optimal feature vector into the multi-label classifier, outputs the predicted probability of each level, and combines the dynamic threshold to determine the fault level.

[0173] When calculating stage attention weights, a nonlinear transformation is used to map the stage mean features to attention scores. The stage weights are then normalized using the Softmax function to achieve key stage selection. When calculating time-based attention weights, the same method as for stage attention is used, focusing on core time-based features within each stage. The load standard deviation is then mapped to a compensation factor in the 0-1 range using an exponential function. The more stable the load, the closer the compensation factor is to 1. Finally, a linear correction is applied to the baseline threshold based on the compensation factor. When load fluctuations are large (…),… Lowering the threshold (small) can avoid missed detections; when the load is stable ( Increasing the threshold (by a large margin) can avoid false positives. Finally, a fully connected layer maps the high-dimensional optimal features to fault probabilities, and the sigmoid function is used to normalize the probabilities, adapting to the needs of multi-label classification.

[0174] Therefore, this invention adopts a two-level attention mechanism at both the stage and time point. By introducing a working condition compensation factor, load stability is quantified as the basis for adjusting the classification threshold, achieving dynamic threshold adaptation. At the same time, a multi-label classifier combined with a Sigmoid activation function is used to achieve simultaneous diagnosis of the presence and severity of faults. This solves the problems of inaccurate focusing and fixed classification thresholds in traditional attention mechanisms. By filtering key features layer by layer through two-level attention, the feature discrimination is improved. By dynamically adjusting the threshold through working condition compensation, the diagnostic accuracy and fault tolerance of different stages are balanced, achieving accurate output of fault classification.

[0175] Reference Figure 3 , Figure 3 This is a structural block diagram of an embodiment of the reciprocating compressor fault diagnosis device of the present invention.

[0176] like Figure 3 As shown, the reciprocating compressor fault diagnosis device proposed in this embodiment of the invention includes:

[0177] Module 10 is used to obtain the cylinder leakage correlation parameters of the target reciprocating compressor, construct the original parameter matrix, and preprocess the original parameter matrix to obtain the standardized parameter matrix.

[0178] The generation module 20 is used to allocate local neighborhood windows with the standardized parameter matrix as input, fit the non-steady-state fluctuation baseline formed by non-steady-state fluctuations using local weighted regression, and generate a fault feature matrix using the standardized parameter matrix and the non-steady-state fluctuation baseline.

[0179] The segmentation module 30 is used to perform adaptive segmentation of operating conditions according to the fault feature matrix, perform alignment, enhancement and splicing processing on each operating condition stage, input the spliced ​​features into the load-aware BiGRU model, and obtain the high-dimensional feature matrix output by the load-aware BiGRU model.

[0180] The diagnostic module 40 is used to filter and weightedly fuse the high-dimensional feature matrix using a stage attention computing mechanism to obtain the optimal feature vector, and input the optimal feature vector into a multi-label classifier that fuses the working condition compensation factor to obtain the reciprocating compressor fault diagnosis result.

[0181] Other embodiments or specific implementations of the reciprocating compressor fault diagnosis device of the present invention can be referred to the above-described method embodiments, and will not be repeated here.

[0182] Furthermore, the present invention also proposes a reciprocating compressor fault diagnosis device, which includes: a memory, a processor, and a reciprocating compressor fault diagnosis program stored in the memory and executable on the processor. When the reciprocating compressor fault diagnosis program is executed by the processor, it implements the steps of the reciprocating compressor fault diagnosis method as described above.

[0183] The specific implementation of the reciprocating compressor fault diagnosis device of this application is basically the same as the embodiments of the above-mentioned reciprocating compressor fault diagnosis method, and will not be repeated here.

[0184] Furthermore, this invention also proposes a readable storage medium, which includes a computer-readable storage medium storing a reciprocating compressor fault diagnosis program thereon. The readable storage medium may be... Figure 1 The memory 1005 in the terminal can also be at least one of ROM (Read-Only Memory) / RAM (Random Access Memory), magnetic disk, optical disk, etc. The readable storage medium includes several instructions to cause a reciprocating compressor fault diagnosis device with a processor to execute the reciprocating compressor fault diagnosis method described in various embodiments of the present invention.

[0185] The specific implementation in the readable storage medium of this application is basically the same as the embodiments of the above-described reciprocating compressor fault diagnosis method, and will not be repeated here.

[0186] It is understood that in the description of this specification, references to terms such as "one embodiment," "another embodiment," "other embodiments," or "first embodiment to Nth embodiment," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0187] 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.

[0188] 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.

[0189] 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 ROM / RAM, magnetic disk, optical disk) as described above, 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.

[0190] 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 diagnosing faults in a reciprocating compressor, characterized in that, Includes the following steps: Obtain the cylinder leakage correlation parameters of the target reciprocating compressor, construct the original parameter matrix, and preprocess the original parameter matrix to obtain the standardized parameter matrix; specifically including: Obtain cylinder leakage correlation parameters of the target reciprocating compressor; wherein, the cylinder leakage correlation parameters include the intake pressure sequence, exhaust pressure sequence and cylinder wall temperature sequence collected by sensors respectively deployed at the compressor intake port, compressor exhaust port and cylinder wall; Based on a preset sampling frequency, preset cycle duration, and preset target number of reciprocating compressor start-stop cycles, data from several sampling points in the intake pressure sequence, exhaust pressure sequence, and cylinder wall temperature sequence are extracted to construct an original parameter matrix. The specific expression is as follows: ； In the formula, Represents the original parameter matrix, For the first The original value of inhalation pressure at all times. For the first The original value of the exhaust pressure at any time. For the first The original value of the cylinder wall temperature at any given time. Indicates the number of sampling points. ; Using the standardized parameter matrix as input, a local neighborhood window is allocated. A locally weighted regression method is used to fit the baseline of the unsteady-state fluctuations. The standardized parameter matrix and the unsteady-state fluctuation baseline are then used to generate a fault feature matrix; specifically including: Based on the load change phases of the target reciprocating compressor during its start-stop cycle, for each moment... Dynamically allocate local neighborhood windows and use a Gaussian kernel function to calculate the time interval within the window. Based on the weights, construct a local weight matrix; Using the standardized parameter matrix as input, combined with each time step... Using the local neighborhood window and local weight matrix, a local weighted regression algorithm is employed for each time step. No. The parameters are fitted to the unsteady-state fluctuation baseline to obtain the fluctuation baseline fitting value for each parameter; The specific expression for the baseline fitting value of each parameter is as follows: ； In the formula, For the first Time of the first The baseline of non-steady-state fluctuations of each parameter For the first The local weight matrix at time t, For the first The parameter in the first... Local data matrix within the time-neighbor window; Each moment The The standardized parameter values ​​are subtracted from the fluctuation baseline fitting values ​​to remove unsteady fluctuation noise, decouple the cylinder internal leakage fault characteristics, and integrate them to form a fault feature matrix. The specific expression for decoupling the cylinder internal leakage fault characteristics is as follows: ； In the formula, For the first Time of the first The fault characteristic values ​​after decoupling of each parameter, For the first Time of the first The standardized parameter values ​​of each parameter; Specifically, the load change phases of the target reciprocating compressor during the start-stop cycle include: the initial rise phase (0-3s), the rapid rise phase (3-7s), and the stable phase (7-10s); at each moment... The dynamically allocated local neighborhood windows specifically include: a local neighborhood window with 50 data points for the initial and rapid rise phases, and a local neighborhood window with 30 data points for the stable phase. Based on the fault feature matrix, adaptive segmentation of operating conditions is performed. Alignment, enhancement, and splicing are then applied to each operating condition stage. The spliced ​​features are input into the load-aware BiGRU model to obtain a high-dimensional feature matrix output by the load-aware BiGRU model. Specifically, this includes: Based on the load change stages of the target reciprocating compressor during its start-stop cycle, the decoupled fault feature matrix is ​​divided into stages, resulting in feature matrices for three stages. , , For the characteristic matrix , , Perform alignment, enhancement, and splicing processes; Obtain the load normalization value of the target reciprocating compressor during the start-stop cycle, calculate the load-related gating weight at each time i using the Sigmoid activation function, and construct a load-aware BiGRU model with dynamic gating adjustment using the load-related gating weight. The load-related gating weights include the load-related weights of the update gate and the load-related weights of the reset gate, and the specific expression is as follows: ； ； In the formula, To update the load-related weights of the gate, To reset the load-related weights of the door, For the first Normalized load value at time 10:00 This is the load weight matrix. For bias terms, Use the Sigmoid activation function; splicing features Input the load-sensing BiGRU model and combine it with load-related gating weights to obtain the high-dimensional feature matrix output by the load-sensing BiGRU model; Using a stage attention calculation mechanism, the optimal feature vector is obtained by filtering and weighted fusion from the high-dimensional feature matrix. The optimal feature vector is then input into a multi-label classifier that integrates the operating condition compensation factor to obtain the fault diagnosis result of the reciprocating compressor.

2. The reciprocating compressor fault diagnosis method as described in claim 1, characterized in that, splicing features The steps for obtaining the high-dimensional feature matrix output by the load-sensing BiGRU model, which involves inputting the load-sensing BiGRU model and combining it with load-related gating weights, specifically include: splicing features Input the load-aware BiGRU model, update the update gate and reset gate of the load-aware model by combining the load-related gating weights, and update the forward hidden state of the forward GRU and the backward hidden state of the backward GRU in the load-aware BiGRU model based on the gating output. The expressions for the update gate and reset gate for updating load awareness are as follows: ； ； The update expressions for the forward hidden state and the backward hidden state are as follows: ； ； In the formula, It is the Sigmoid activation function. For the first The decoupled eigenvector at time step 1. This represents the hyperbolic tangent activation function. For the first Load-related gating weights at any given time. The state was hidden in the previous moment. For load sensing, update the gate and reset the gate output. This is the weight matrix. For bias terms, For element-wise multiplication, In the candidate hidden state, This is the hidden state updated at the current moment; The forward hidden state and the backward hidden state are concatenated to form a bidirectional fused high-dimensional feature matrix, which is used as the output of the load-aware BiGRU model. The specific expression is as follows: ； In the formula, Forward GRU Always hide your status; For the backward GRU Always hide your status; The hidden state after bidirectional fusion, the hidden state after bidirectional fusion at all time points. Constructing a high-dimensional feature matrix .

3. The reciprocating compressor fault diagnosis method as described in claim 2, characterized in that, The steps of using a stage attention computation mechanism to filter and weightedly fuse features from the high-dimensional feature matrix to obtain the optimal feature vector, and then inputting the optimal feature vector into a multi-label classifier that fuses operating condition compensation factors to obtain the reciprocating compressor fault diagnosis result, specifically include: High-dimensional feature matrix The process is divided into stages. Calculate the mean vector of the hidden states at each stage. The importance weights of each stage are calculated using an attention scoring function. ; The feature matrices of each stage With the corresponding stage attention weights Perform a weighted summation to generate a feature matrix representing the importance of the fusion stage. For stage-weighted feature matrix Calculate the attention score for the hidden state at each time step, and obtain the attention weights at each time step using the Softmax function. ; The stage-weighted feature matrix Hidden states at each time step and corresponding attention weights at each time step Perform weighted summation to generate the globally optimal feature vector. ; Based on the load sequence of each stage, the load standard deviation is calculated and converted into the operating condition compensation factor. Based on the baseline threshold of each leakage level and the operating condition compensation factor, a dynamically adjusted classification threshold is generated. The global optimal eigenvector Input a multi-label classifier, pass it through a fully connected layer and a sigmoid activation function, and output the predicted probability of each leakage level. Using the predicted probability By dynamically adjusting the classification threshold, fault diagnosis results for reciprocating compressors are generated.

4. A fault diagnosis device for a reciprocating compressor, characterized in that, A method for performing fault diagnosis of a reciprocating compressor as described in any one of claims 1-3, comprising: A construction module is used to obtain the cylinder leakage correlation parameters of the target reciprocating compressor, construct the original parameter matrix, and preprocess the original parameter matrix to obtain the standardized parameter matrix; The generation module is used to allocate local neighborhood windows with the standardized parameter matrix as input, fit the non-steady-state fluctuation baseline formed by non-steady-state fluctuations using local weighted regression, and generate a fault feature matrix using the standardized parameter matrix and the non-steady-state fluctuation baseline. The segmentation module is used to perform adaptive segmentation of operating conditions based on the fault feature matrix, perform alignment, enhancement and splicing processing on each operating condition stage, input the spliced ​​features into the load-aware BiGRU model, and obtain the high-dimensional feature matrix output by the load-aware BiGRU model. The diagnostic module is used to filter and weightedly fuse the high-dimensional feature matrix using a stage attention computing mechanism to obtain the optimal feature vector. The optimal feature vector is then input into a multi-label classifier that incorporates the operating condition compensation factor to obtain the fault diagnosis result of the reciprocating compressor.

5. A fault diagnosis device for a reciprocating compressor, characterized in that, The reciprocating compressor fault diagnosis device includes: a memory, a processor, and a reciprocating compressor fault diagnosis program stored in the memory and executable on the processor. When the reciprocating compressor fault diagnosis program is executed by the processor, it implements the steps of the reciprocating compressor fault diagnosis method as described in any one of claims 1 to 3.

6. A storage medium, characterized in that, The storage medium stores a reciprocating compressor fault diagnosis program, which, when executed by a processor, implements the steps of the reciprocating compressor fault diagnosis method as described in any one of claims 1 to 3.

Images