Multi-parameter fusion fault feature extraction and early warning method and system for reciprocating compressor

Abstract

The application discloses a kind of reciprocating compressor multi-parameter fusion fault feature extraction and early warning method and system, the combination of the present application is spliced to multi-scale feature extraction mechanism by multiple measuring points vibration signal, to simultaneously capture the trend characteristics of signal different scales, significantly improve the multi-resolution representation ability to non-stationary vibration signal;For each measuring point signal, extract time-frequency physical prior feature independently, and carry out depth nonlinear mapping by feature mapping network, so that the domain knowledge with clear physical meaning is explicitly injected into fault detection model in the form of feature vector;In addition, the present application maps vibration signal to discrete state space under different scales, extracts feature by state vector reconstruction and permutation combination, generates composite feature vector representing local fluctuation degree and time sequence evolution law of signal;Finally, the above three kinds of features are fused to form fault feature vector and input into fault detection model, so that the hierarchical early warning of reciprocating compressor can be realized according to the detection result.

Description

Technical Field

[0001] This invention belongs to the field of compressor fault early warning technology based on deep learning, specifically involving a method and system for multi-parameter fusion fault feature extraction and early warning of reciprocating compressors. Background Technology

[0002] Reciprocating compressors are core power equipment in industries such as petroleum, chemical, and natural gas transportation. Their operational reliability directly affects the safety and efficiency of the entire production system. Due to their complex structure, numerous vulnerable parts, and long-term operation under high temperature, high pressure, and complex alternating loads, key components (such as valves, piston rods, bearings, and crankshafts) inevitably experience wear, fatigue, and performance degradation. This leads to various faults, including valve leakage, piston rod breakage, and excessive bearing clearance. If these faults are not detected and warned of in a timely manner, they can reduce equipment operating efficiency and increase energy consumption, or even cause unplanned shutdowns or catastrophic safety accidents, resulting in huge economic losses and safety risks. Therefore, developing advanced fault diagnosis and graded early warning technologies for real-time and accurate condition monitoring and health assessment of reciprocating compressors has significant theoretical value and engineering application implications.

[0003] Currently, vibration signals are typically collected for fault diagnosis of reciprocating compressors. Existing fault diagnosis methods for reciprocating compressors commonly include traditional signal processing techniques and deep learning-based intelligent diagnostic techniques. However, these existing technologies have the following shortcomings: (1) Limited ability to extract dynamic features at multiple scales; the vibration signal of reciprocating compressor has strong non-stationary, nonlinear and multi-component coupling characteristics. Different fault modes are often reflected at different time scales. That is, high frequency components reflect local impact and transient response, while low frequency components reflect long period trend and overall vibration level. However, most existing methods adopt a single-scale feature extraction strategy, which makes it difficult to capture fault feature information at different time scales in the signal at the same time.

[0004] (2) Insufficient integration of physical prior knowledge and data-driven features; Although existing pure data-driven methods have strong nonlinear mapping capabilities, their "black box" characteristics lead to a lack of interpretability in the model, resulting in insufficient generalization ability when facing complex industrial scenarios such as variable working conditions and strong noise. Traditional physical feature extraction (such as time-domain statistical indicators and frequency-domain features) has clear physical meaning and good stability, but it often only stays at the shallow statistical level and fails to fully explore the deep nonlinear coupling relationship between features.

[0005] (3) Lack of systematic characterization of the local fluctuation degree and temporal evolution law of vibration signal; The occurrence and development of reciprocating compressor failure is a gradual dynamic process. Its vibration signal not only contains amplitude information, but also contains rich temporal evolution law and local fluctuation characteristics. However, existing methods focus on single-type feature extraction (such as extracting only statistical features), and fail to effectively extract feature information that can reflect the local fluctuation degree and temporal evolution law of vibration signal, making it difficult to fully characterize the health degradation trend of equipment.

[0006] Therefore, based on the aforementioned shortcomings, how to fully explore the multi-scale dynamic information, physical deterministic information and temporal evolution law contained in the vibration signals of multiple measuring points of reciprocating compressors, so as to achieve effective fusion of multi-source heterogeneous features and establish a fault early warning model oriented towards engineering practice, is a key technical problem that urgently needs to be solved in the field of intelligent operation and maintenance of reciprocating compressors. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for multi-parameter fusion fault feature extraction and early warning of reciprocating compressors, in order to solve the problems of limited multi-scale dynamic feature extraction capability, insufficient fusion of physical prior knowledge and data-driven features, and lack of systematic characterization of the local fluctuation degree and temporal evolution law of vibration signals in the existing technology, which leads to low fault detection accuracy.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, a method for multi-parameter fusion fault feature extraction and early warning of reciprocating compressors is provided, including: Obtain vibration signals corresponding to different measuring points on the reciprocating compressor; Multiple vibration signals are spliced ​​together to obtain a spliced ​​signal, and multi-scale time-series feature extraction is performed on the spliced ​​signal to obtain the multi-scale time-series depth feature vector of the reciprocating compressor. Time-frequency features are extracted from each vibration signal to obtain the time-frequency features of each vibration signal. Then, a feature mapping network is used to perform nonlinear mapping on each time-frequency feature to obtain the physical mapping features of each vibration signal. Based on the physical mapping features, the physical prior feature vector of the reciprocating compressor is generated. For any vibration signal among multiple vibration signals, at different scales, the vibration signal is mapped to a discrete state space to obtain discrete state sequences at different scales, and state vectors are reconstructed for each discrete state sequence to obtain reconstructed vectors at different scales. Extract the state arrangement and combination features of each reconstruction vector in the reconstruction space, and generate a composite feature vector corresponding to any vibration signal based on the state arrangement and combination features of each reconstruction vector in the reconstruction space. After all vibration signals have been traversed, a composite correlation vector is formed by using each composite feature vector. Each composite feature vector is used to characterize the local fluctuation degree and temporal evolution law of the corresponding vibration signal at multiple scales. Fault feature vectors are generated by using physical prior feature vectors, multi-scale time-series deep feature vectors, and composite correlation vectors. These fault feature vectors are then input into the fault detection model to perform fault classification and early warning for reciprocating compressors based on the fault detection results.

[0009] Based on the aforementioned disclosures, this invention constructs a multi-scale time-series deep feature extraction mechanism. This involves splicing vibration signals from multiple measurement points and extracting features at multiple scales to simultaneously capture the trend features of the signals at different scales, thereby obtaining a multi-scale time-series deep feature vector. In this way, the invention overcomes the shortcomings of traditional single-scale feature extraction methods in comprehensively representing complex fault modes, significantly improving the multi-resolution representation capability of non-stationary vibration signals. Furthermore, this invention independently extracts time-frequency features (i.e., physical prior information) from the signals at each measurement point and performs deep nonlinear mapping through a feature mapping network. This allows domain knowledge with clear physical meaning to be explicitly injected into the fault detection model in the form of feature vectors. Based on this, the mechanism retains the advantages of strong stability and good interpretability of traditional physical features while achieving complex coupling between features through neural networks. The automatic mining of composite relationships effectively compensates for the insufficient generalization ability caused by the "black box" nature of pure data-driven models. Simultaneously, this invention pioneers a composite feature extraction strategy based on discrete state space mapping. Specifically, at different scales, vibration signals are mapped to discrete state spaces. Through state vector reconstruction and permutation combination feature extraction, composite feature vectors that characterize the degree of local fluctuations and temporal evolution of the signal are generated. Based on this, the composite features of this invention fully explore the temporal transitions and permutation patterns between states, enabling more sensitive capture of subtle dynamic mutations during fault occurrence and development, and achieving a systematic characterization of the degree of local fluctuations and temporal evolution of vibration signals. Finally, using the three extracted features mentioned above to form fault features and inputting them into a fault detection model, graded early warning for reciprocating compressors can be achieved based on the fault detection results.

[0010] In one possible design, a multi-scale feature extraction model is used to extract multi-scale temporal features from the spliced ​​signal in order to obtain the multi-scale temporal depth feature vector of the reciprocating compressor. The multi-scale feature extraction model includes: a local feature extraction layer, a multi-scale temporal residual layer, a feature projection layer, and an adaptive pooling layer. The multi-scale temporal residual layer includes a feature concatenation unit, a residual unit, and multiple temporal feature extraction units. The local feature extraction layer is used to perform one-dimensional convolution and max pooling on the spliced ​​signal in sequence to obtain local feature vectors, and then input the local feature vectors into the residual unit and each temporal feature extraction unit respectively. For any one of the multiple temporal feature extraction units, the temporal feature extraction unit is used to perform a sliding window-style local convolution operation on the received local feature vector along the temporal dimension with a one-dimensional convolution kernel of a preset size and a preset stride to obtain the convolution response feature vector; Each temporal feature extraction unit is used to perform batch normalization on the convolutional response feature vector to obtain a normalized response feature vector, and to use the ReLU activation function to perform nonlinear activation mapping on the normalized response feature vector to obtain a temporal feature vector, which is then input to the feature concatenation unit. The size of the convolutional kernel of each temporal feature extraction unit is different to output temporal feature vectors of different scales. The residual unit is used to perform 1×1 convolution and batch normalization on the received local feature vectors to obtain residual connection features. The feature concatenation unit is used to concatenate the temporal feature vectors output by each temporal feature extraction unit along the channel dimension to obtain a temporal concatenation feature vector. It also adds the temporal concatenation feature vector to the residual connection feature vector element by element and performs nonlinear activation processing on the addition result to obtain a multi-scale temporal fusion feature vector after nonlinear activation processing. The feature projection layer is used to perform 1×1 convolution processing on the multi-scale temporal fusion feature vector to adjust the number of channels of the multi-scale temporal fusion feature vector to a preset dimension, so as to obtain the mapped feature vector. An adaptive pooling layer is used to perform adaptive pooling processing on the mapped feature vector to obtain the multi-scale temporal depth feature vector after adaptive pooling processing.

[0011] In one possible design, at different scales, any vibration signal is mapped to a discrete state space to obtain discrete state sequences at different scales, including: Obtain a set of scale factors, wherein the set of scale factors includes multiple different scale factors; Select the i-th scale factor from the set of scale factors as the division length, and divide any vibration signal into several signal blocks according to the division length, wherein the initial value of i is 1; Extract the local fluctuation feature value of each signal block, and use the local fluctuation feature value of each signal block to form a block sequence of any vibration signal under the i-th scale factor; Increment i by 1 and reselect the i-th scale factor from the set of scale factors until i is cycled from 1 to n, to obtain the block sequence of any vibration signal under different scale factors, where n is the total number of scale factors; Each block sequence is discretized and encoded to obtain several initial discrete state sequences. Each initial discrete state sequence is then linearly mapped to obtain discrete state sequences at different scales. The length of any block sequence is the same as the length of the discrete state sequence corresponding to that block sequence. Accordingly, reconstructing the state vectors of each discrete state sequence yields reconstructed vectors at different scales, including: Obtain the reconstructed embedding dimension and embedding delay, and use the embedding delay as a fixed interval; For any discrete state sequence, starting from the d-th element in the discrete state sequence, select c elements at fixed intervals to form the embedding vector corresponding to the d-th element using the selected c elements, where c represents the reconstructed embedding dimension and the initial value of d is 1. Increment d by 1, and starting from the d-th element in any discrete state sequence, select c elements forward at the fixed intervals until d equals 1. At that time, several embedding vectors are obtained, among which, This represents the length of any discrete state sequence. The fixed interval; Using several embedding vectors, a scale-based reconstruction vector is formed for any discrete state sequence.

[0012] In one possible design, each block sequence is discretized and encoded to obtain several initial discrete state sequences, including: For any block sequence, calculate the standard deviation and mean of the given block sequence; Based on the standard deviation and mean, and according to the following formula, the any block sequence is discretized and encoded. ; In the formula, This represents the j-th local fluctuation feature value in any of the block sequences. express The corresponding discretized coded value, As an intermediate variable for integration, Let represent the standard deviation and mean of any of the block sequences, respectively, where ,and Indicates the length of any of the block sequences; Accordingly, a linear mapping is performed on each initial discrete state sequence to obtain discrete state sequences at different scales after the linear mapping, which includes: For any initial discrete state sequence, a linear mapping is performed on the sequence according to the following formula; ; In the formula, express The linear mapping value, This is a rounding function. Represents the number of discretization levels. The offset represents the rounding off, wherein the corresponding discrete state sequence is formed by using the linear mapping value corresponding to each discretized coded value in any initial discrete state sequence.

[0013] In a possible design, any reconstruction vector contains several embedding vectors, each embedding vector corresponding to a state permutation and combination feature, and the same embedding vector corresponds to the same state permutation and combination feature. Accordingly, based on the state arrangement and combination characteristics of each reconstruction vector in the reconstruction space, a composite feature vector corresponding to any vibration signal is generated, including: For any reconstructed vector, the frequency of occurrence of each embedded vector in the reconstructed vector is calculated. Based on the frequency of occurrence of each embedding vector, the probability of the state pattern corresponding to each state permutation and combination feature is calculated. Based on the state pattern probability corresponding to each state permutation and combination feature, the composite feature entropy corresponding to any reconstruction vector is calculated, and after all reconstruction vectors have been traversed, the composite feature entropy of each reconstruction vector is obtained. Using all the composite feature entropies, a composite feature vector corresponding to any vibration signal is generated; The composite feature entropy corresponding to any reconstructed vector is calculated according to the following formula; ; In the formula, This represents the composite feature entropy corresponding to any of the reconstructed vectors. Indicates the first The probability of state patterns corresponding to the permutation and combination features of various states. This represents the total number of state permutation and combination features in any of the reconstructed vectors.

[0014] In one possible design, the fault detection model includes an input layer, a hidden layer, and an output layer. The fault detection model is constructed in the following manner. Obtain a training set, which contains multiple sample fault feature vectors of reciprocating compressors; A clustering algorithm based on local distribution is used to cluster the fault features of multiple samples in the training set to obtain the number of cluster centers. The parameter dimensions of the initial fault detection model are determined based on the number of nodes in the input layer, the number of nodes in the output layer, and the number of cluster centers. Based on the parameter dimensions, the initial fault detection model is initialized to obtain several sets of initial model parameters, and an initial population is formed using these sets of initial model parameters. Obtain the individual population at the t-th iteration and the prey set at the (t-1)-th iteration, wherein when t is 1, the individual population at the t-th iteration is the initial population, and the prey set at the (t-1)-th iteration is the initial prey set. Any initial prey is used to characterize a set of initial model parameters, and the initial prey set is the initial population. Using the individual population and prey set at the t-th iteration, a candidate population is formed. Based on the training set, the fitness of each candidate individual in the candidate population is calculated. Based on the fitness of each candidate individual, the prey set at the t-th iteration is determined. The number of prey in the prey set decreases as the number of iterations increases. Based on the prey set at the t-th iteration, determine the prey corresponding to each individual in the individual population at the t-th iteration; Determine if the iteration stopping condition is met; If not, calculate the hunting step size at the t-th iteration and the hunting intensity of each individual at the t-th iteration. Based on the hunting stride length, the hunting intensity of each individual at the t-th iteration and their corresponding prey, the position of each individual at the t-th iteration is updated to obtain the individual population at the (t+1)-th iteration. Increment t by 1, and reacquire the individual population at the t-th iteration and the prey set at the (t-1)-th iteration until the iteration stopping condition is met. Then, determine the optimal model parameters based on the prey corresponding to the individual with the highest fitness in the individual population at the time the iteration stopping condition is met. The initial fault detection model is updated using the optimal model parameters to obtain the fault detection model.

[0015] In one possible design, a clustering algorithm based on local distribution is used to cluster multiple sample fault features in the training set to obtain the number of cluster centers, including: Initialize the cluster counter, where the initial value of the cluster counter is 0, and the count of the cluster counter is used to represent the number of cluster centers; Obtain the sample set at the q-th clustering and the cluster confidence corresponding to the cluster center at the (q-1)-th clustering, wherein when q is 1, the sample set at the q-th clustering is the training set, and the cluster confidence corresponding to the cluster center at the (q-1)-th clustering is the initial value. For any sample fault feature in the sample set during the q-th clustering, obtain the neighborhood fault features of the fault feature ... Using the neighborhood fault features and the cluster confidence corresponding to the cluster center at the (q-1)th clustering, the local density index corresponding to the fault features of any sample is calculated. Based on the local density index and the local distribution index, the clustering confidence of any sample fault feature is calculated, and after all sample fault features in the sample set at the qth clustering time have been traversed, the clustering confidence of each sample fault feature in the sample set is obtained. From the fault features of each sample in the sample set, select the fault feature with the highest clustering confidence as the cluster center for the qth clustering, and increment the clustering counter by 1. Determine whether the clustering stopping condition is met; If not, then from the sample set of the qth clustering, delete the cluster centers of the qth clustering to obtain the sample set of the (q+1)th clustering. Increment q by 1 and reacquire the sample set for the qth clustering iteration until the clustering stopping condition is met. The number of cluster centers is determined based on the total count of the clustering counter when the clustering stopping condition is met.

[0016] In one possible design, based on the fitness of each candidate individual, the prey set for the t-th iteration is determined, including: Find the maximum number of iterations and the maximum number of prey. Calculate the number of prey in the t-th iteration based on the maximum number of iterations, the maximum number of prey, and the number of iterations t. The candidate individuals are sorted in descending order of fitness to obtain the sorted sequence; Filter the top from the sorted sequence There are _n_ candidate individuals, which form the prey combination in the _t_th iteration, where _n_... Let be the number of prey in the t-th iteration.

[0017] In one possible design, the hunting step size at the t-th iteration and the hunting intensity of each individual at the t-th iteration are calculated, including: The hunting intensity of any individual at the t-th iteration is calculated using the following formula. ; In the formula, This indicates the hunting intensity of any of the individuals mentioned. Let be the initial hunting intensity of any individual in the t-th iteration. This represents the hunting stamina coefficient of any individual. Decrease linearly from 1 to 0. Let represent the time influence coefficient and the location influence coefficient, respectively. Represents any of the individuals mentioned; Accordingly, based on the hunting step length, the hunting intensity of each individual at the t-th iteration, and their respective prey, the position of each individual at the t-th iteration is updated, including: For any individual at the t-th iteration, determine whether the hunting intensity of that individual is greater than the initial hunting intensity of that individual; If so, the hunting distance between any individual and its corresponding prey is calculated, and the position of any individual is updated based on the hunting distance and the hunting stride length to obtain the updated individual corresponding to any individual. Otherwise, a random jump update strategy is used to update the position of any individual to obtain the updated individual corresponding to any individual.

[0018] Secondly, a multi-parameter fusion fault feature extraction and early warning system for reciprocating compressors is provided, including: The acquisition unit is used to acquire vibration signals corresponding to different measuring points on the reciprocating compressor; The multi-scale feature extraction unit is used to splice multiple vibration signals to obtain a spliced ​​signal, and to extract multi-scale temporal features from the spliced ​​signal to obtain a multi-scale temporal depth feature vector of the reciprocating compressor. The physical prior feature extraction unit is used to extract time-frequency features from each vibration signal to obtain the time-frequency features of each vibration signal. Then, using a feature mapping network, nonlinear mapping is performed on each time-frequency feature to obtain the physical mapping features of each vibration signal. Based on the physical mapping features, the physical prior feature vector of the reciprocating compressor is generated. The reconstruction unit is used to map any vibration signal from multiple vibration signals to a discrete state space at different scales to obtain discrete state sequences at different scales, and to reconstruct the state vectors of each discrete state sequence to obtain reconstruction vectors at different scales. The composite feature extraction unit is used to extract the state arrangement and combination features of each reconstruction vector in the reconstruction space, so as to generate a composite feature vector corresponding to any vibration signal based on the state arrangement and combination features of each reconstruction vector in the reconstruction space, and after all vibration signals have been traversed, a composite correlation vector is formed by using each composite feature vector. In this case, any composite feature vector is used to characterize the local fluctuation degree and temporal evolution law of the corresponding vibration signal under multiple scales. The fault early warning unit is used to generate fault feature vectors by utilizing physical prior feature vectors, multi-scale time-series deep feature vectors, and composite correlation vectors, and inputs the fault feature vectors into the fault detection model to perform fault classification and early warning for reciprocating compressors based on the fault detection results.

[0019] Thirdly, a device for extracting and warning of multi-parameter fusion fault features of reciprocating compressors is provided. Taking the device as an electronic device as an example, it includes a memory, a processor, and a transceiver that are connected in sequence. The memory is used to store computer programs, the transceiver is used to send and receive messages, and the processor is used to read the computer programs and execute the method for extracting and warning of multi-parameter fusion fault features of reciprocating compressors as described in the first aspect or any possible design in the first aspect.

[0020] Fourthly, a storage medium is provided, on which instructions are stored, which, when executed on a computer, perform the reciprocating compressor multi-parameter fusion fault feature extraction and early warning method as described in the first aspect or any possible design in the first aspect.

[0021] Fifthly, a computer program product containing instructions is provided, which, when executed on a computer, causes the computer to perform the reciprocating compressor multi-parameter fusion fault feature extraction and early warning method as described in the first aspect or any possible design of the first aspect.

[0022] Beneficial effects: (1) This invention combines multi-point vibration signal splicing with a multi-scale feature extraction mechanism to simultaneously capture the trend features of signals at different scales, significantly improving the multi-resolution representation capability of non-stationary vibration signals and overcoming the shortcomings of traditional single-scale methods in comprehensively representing complex fault modes. At the same time, time-frequency physical prior features are independently extracted for each measurement point signal and then subjected to deep nonlinear mapping through a feature mapping network. This allows domain knowledge with clear physical meaning to be explicitly injected into the fault detection model in the form of feature vectors. This not only retains the advantages of strong stability and good interpretability of physical features, but also effectively compensates for the insufficient generalization ability caused by the "black box" characteristics of pure data-driven models by automatically mining the complex coupling relationships between features through neural networks. Furthermore, this invention pioneers a composite feature extraction strategy based on discrete state space mapping. Specifically, vibration signals are mapped to discrete state spaces at different scales. Through state vector reconstruction and permutation combination feature extraction, a composite feature vector characterizing the local fluctuation degree and temporal evolution law of the signal is generated. This fully explores the temporal transitions and permutation patterns between states, enabling sensitive capture of subtle dynamic mutations during fault occurrence and development. Finally, the above three features are fused to form a fault feature vector, which is then input into a fault detection model. Based on the detection results, graded early warning for reciprocating compressors can be achieved. Therefore, a clear decision-making basis can be provided for engineering operation and maintenance.

[0023] (2) This invention establishes a three-branch parallel fusion architecture of multi-source heterogeneous features, realizing the synergistic complementarity of deep dynamic features, physical deterministic features and temporal evolution features. That is, this invention splices and fuses multi-scale temporal deep feature vectors, physical prior feature vectors and composite correlation vectors at the feature level to form a more comprehensive fault feature vector. The three feature extraction branches run independently and in parallel, respectively representing the health status of the equipment from three dimensions: "data-driven dynamic pattern mining", "knowledge-driven physical deterministic description" and "entropy-driven temporal evolution characterization". The three form effective information complementarity and cross-validation, which significantly improves the sensitivity and distinguishability of fault features to various fault modes.

[0024] (3) The redundant and complementary information of vibration signals from multiple measurement points is fully utilized, overcoming the inherent limitation of incomplete information in single-measurement-point diagnostic methods. This invention addresses the characteristics of reciprocating compressors, which have numerous vibration excitation sources and exhibit different response intensities at different measurement points. Through synchronous acquisition of multiple measurement points and parallel processing of multiple branches, it fully leverages the characterization information of different aspects of the same fault state from different measurement point signals. The combined use of multiple measurement point signals not only improves the coverage and accuracy of fault diagnosis but also ensures the reliability of diagnostic results even when the signal quality of one or several measurement points deteriorates, greatly enhancing the engineering practicality of the method. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the steps of the method for multi-parameter fusion fault feature extraction and early warning of reciprocating compressors provided in an embodiment of the present invention. Figure 2 This is a network structure diagram of the multi-scale feature extraction model provided in an embodiment of the present invention; Figure 3 This is a structural diagram of the reciprocating compressor multi-parameter fusion fault feature extraction and early warning system provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the accompanying drawings is only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0027] It should be understood that although the terms first, second, etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another. For example, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit, without departing from the scope of the exemplary embodiments of the invention.

[0028] It should be understood that the term "and / or" that may appear in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" that may appear in this document describes another relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " that may appear in this document generally indicates that the related objects before and after it are in an "or" relationship.

[0029] Example: See Figure 1As shown, the multi-parameter fusion fault feature extraction and early warning method for reciprocating compressors provided in this embodiment can be executed by a computer device with certain computing resources, including but not limited to servers, edge computers, or personal computers (PCs are multi-purpose computers of a size, price, and performance suitable for personal use; desktop computers, laptops, mini-laptops, tablets, and ultrabooks are all personal computers). It is understood that the aforementioned execution subject does not constitute a limitation on the embodiments of this application. Accordingly, the operation steps of this method can be, but are not limited to, the steps S1 to S6 below.

[0030] S1. Obtain vibration signals corresponding to different measuring points on the reciprocating compressor; in specific applications, for example, but not limited to, collecting vibration signals from different locations such as the casing, bearing housing, and valves of the reciprocating compressor, so as to make full use of the vibration signals at different measuring points to perform fault diagnosis of the reciprocating compressor.

[0031] After collecting vibration signals at different measurement points, signal preprocessing, such as denoising and standardization, can be performed. Then, the signals can be spliced ​​together. Finally, multi-scale feature extraction can be performed on the spliced ​​signals to capture the trend features of the signals at different scales, thereby significantly improving the multi-resolution characterization capability of non-stationary vibration signals. The multi-scale feature extraction process is shown in step S2 below.

[0032] S2. Multiple vibration signals are spliced ​​together to obtain a spliced ​​signal, and multi-scale temporal feature extraction is performed on the spliced ​​signal to obtain the multi-scale temporal depth feature vector of the reciprocating compressor. In a specific application, assuming that the multiple vibration signals are X1, X2 and X3, the spliced ​​signal is [X1, X2, X3]. Then, in this embodiment, a multi-scale feature extraction model is constructed to extract the multi-scale temporal features of the spliced ​​signal.

[0033] See Figure 2 As shown, this multi-scale temporal feature extraction model takes the spliced ​​signal as input, and its core objective is to automatically and hierarchically mine deep temporal features related to the fault state in the signal. For example, the model may include, but is not limited to, a local feature extraction layer, a multi-scale temporal residual layer, a feature projection layer, and an adaptive pooling layer, and the multi-scale temporal residual layer includes a feature splicing unit, a residual unit, and multiple temporal feature extraction units.

[0034] In practical implementation, the local feature extraction layer is used to sequentially perform one-dimensional convolution and max pooling on the spliced ​​signal. That is, one-dimensional convolution is used to perform preliminary local feature extraction on the spliced ​​signal, and then max pooling is used to downsample the convolutional feature map. This not only preserves the most significant features of each local region, but also further enhances the translation invariance of the model and reduces computational complexity. After obtaining the local feature vector through the above processing, the local feature vector can be input into the residual unit and each temporal feature extraction unit, thereby capturing fault mode information at different time scales in parallel in each temporal feature extraction unit.

[0035] Specifically, for any one of the multiple temporal feature extraction units, the temporal feature extraction unit is used to perform a sliding window-style local convolution operation on the received local feature vector along the temporal dimension (i.e., the time dimension) with a preset stride, to obtain a convolutional response feature vector; then, the temporal feature extraction unit is also used to perform a batch normalization operation on the convolutional response feature vector to obtain a normalized response feature vector, and to perform nonlinear activation mapping processing on the normalized response feature vector using the ReLU activation function to obtain a temporal feature vector that is input to the feature concatenation unit.

[0036] In this embodiment, the convolution kernels of each temporal feature extraction unit have different sizes, thereby outputting temporal feature vectors at different scales. Thus, the aforementioned temporal feature extraction units are equivalent to forming multiple parallel convolutional paths with convolution kernels of different sizes, enabling the multi-scale feature extraction model to perceive signal features from multiple time scales simultaneously. Specifically, small-sized convolution kernels focus on high-frequency transient details, while large-sized convolution kernels focus on low-frequency long-period trends. Finally, the features extracted by each scale path are spliced ​​together along the channel dimension to form a comprehensive description of the signal.

[0037] Furthermore, this embodiment also introduces a residual unit, which is used to sequentially perform 1×1 convolution and batch normalization on the received local feature vectors to obtain residual connection features (that is, to adjust the number of channels of the features through a 1×1 convolution to achieve dimension alignment); then, it can be sent to the feature concatenation unit to be added with the output of the multi-scale convolution to obtain a multi-scale temporal fusion feature vector.

[0038] Specifically, the feature concatenation unit concatenates the temporal feature vectors output by each temporal feature extraction unit along the channel dimension to obtain a temporal concatenation feature vector (forming a feature map that aggregates multi-scale information). It also adds the temporal concatenation feature vector element-wise to the residual connection feature vector and performs non-linear activation processing (using the ReLU activation function) on the addition result to obtain a multi-scale temporal fusion feature vector. Thus, the residual structure designed in this embodiment can effectively alleviate the gradient vanishing problem that may occur in deep networks, enabling the model to learn hierarchical fault features from low to high order and from simple to complex.

[0039] After feature concatenation is completed, a 1×1 convolutional layer is used to integrate and project the feature channels. Then, an adaptive pooling layer is used to generate a fixed-dimensional deep feature vector, namely, the feature projection layer, which is used to perform 1×1 convolution on the multi-scale temporal fusion feature vector to adjust the number of channels of the multi-scale temporal fusion feature vector to a preset dimension, so as to obtain the mapped feature vector. Then, the adaptive pooling layer is used to perform adaptive pooling on the mapped feature vector to obtain the multi-scale temporal deep feature vector after adaptive pooling.

[0040] It should be noted that adaptive pooling is a mature, built-in network layer component in deep learning, widely used in convolutional neural networks (CNNs). Unlike traditional pooling layers, it allows developers to directly specify the target size of the output feature map without manually calculating the kernel size and stride, thus enabling flexible handling of input data with different resolutions.

[0041] Therefore, by using the aforementioned multi-scale feature extraction model, the multi-scale temporal features of the spliced ​​signal can be extracted, thereby capturing the trend features of the signal at different scales. This overcomes the shortcomings of traditional single-scale methods in fully representing complex fault modes and significantly improves the multi-resolution representation capability of non-stationary vibration signals.

[0042] After completing the multi-scale feature extraction, the physical prior features can be extracted, as shown in step S3 below.

[0043] S3. Extract time-frequency features from each vibration signal to obtain the time-frequency features of each vibration signal. Then, use a feature mapping network to perform nonlinear mapping on each time-frequency feature to obtain the physical mapping features of each vibration signal. Based on these physical mapping features, generate the physical prior feature vector of the reciprocating compressor. In specific implementations, the time-frequency features of each vibration signal may include, but are not limited to: mean (measures the static offset level of the signal and evaluates the baseline stability of the equipment), standard deviation (quantifies the dispersion of the signal relative to the mean and characterizes the fluctuation amplitude of the signal), root mean square (reflects the effective energy level of the signal and is directly related to the power of the equipment), and skewness (describes the signal). The spectrum includes the following components: amplitude distribution asymmetry, kurtosis (capturing the impact component in the signal), peak-to-peak value (calculating the difference between the maximum and minimum values ​​of time series data within a period, describing the maximum dynamic range of the signal, and reflecting the extreme fluctuation characteristics of the signal), spectral centroid (calculating the energy-weighted average frequency of the signal spectrum, characterizing the concentrated frequency of the signal spectrum energy, and reflecting the distribution of the signal's main frequency band), main frequency (identifying the frequency component with the largest energy (amplitude) in the signal spectrum, corresponding to the frequency component with the largest amplitude in the spectrum, which is the basis for fault characteristic frequency identification), and spectral entropy (measuring the uniformity and uncertainty of the signal spectrum energy distribution, quantifying the regularity and complexity of the vibration signal spectrum).

[0044] Thus, after obtaining the aforementioned time-frequency domain features, nonlinear mapping of the features can be performed. That is, the nonlinear mapping model aims to explicitly integrate mature theories and expert prior knowledge in the field of vibration diagnosis into the model, so as to enhance the model's ability to understand the fault mechanism and its generalization performance. In this embodiment, nine classic physical prior feature indicators are extracted to characterize the health status of the compressor from multiple dimensions such as time domain, frequency domain and energy distribution. This feature set covers everything from the root mean square and mean values ​​that measure the overall energy and stability of the signal, to the kurtosis that is highly sensitive to early impact faults, as well as the spectral centroid and dominant frequency that reveal the spectral energy distribution and core frequency components. In this way, a feature set with complementary information and clear physical meaning can be formed, which is the key cornerstone for improving the diagnostic performance and interpretability of the model.

[0045] Among these, the aforementioned low-dimensional physical features often exhibit complex nonlinear coupling relationships, and their feature space differs significantly from the high-dimensional abstract feature space extracted by the multi-scale feature extraction model. Therefore, in order to fully explore the deep-level correlations between these features and achieve effective alignment and fusion with deep features in the same high-dimensional space, this embodiment designs a feature mapping network composed of a multilayer perceptron. This feature mapping network performs layer-by-layer nonlinear mapping and dimensionality enhancement on the input time-frequency features through hidden layers containing nonlinear activation functions. In this process, the model can learn and encode complex interaction patterns between features, ultimately transforming the original features into a physical feature vector, namely the physical mapping feature vector.

[0046] Furthermore, for example, a feature mapping network includes an input layer, two hidden layers, and an output layer. The input layer has 9 input nodes, the two hidden layers each contain 128 neurons, and the output layer outputs a 128-dimensional feature as the physical mapping feature. Finally, the physical mapping features of each vibration signal are concatenated to obtain the physical prior features.

[0047] Thus, in this embodiment, time-frequency physical prior features are independently extracted for each measurement point signal, and deep nonlinear mapping is performed through a feature mapping network. This allows domain knowledge with clear physical meaning to be explicitly injected into the fault detection model in the form of feature vectors. This retains the advantages of strong stability and good interpretability of physical features, and also effectively makes up for the problem of insufficient generalization ability caused by the "black box" characteristics of pure data-driven models by automatically mining the complex coupling relationship between features through neural networks.

[0048] After extracting the physical prior features, the local fluctuation degree and temporal evolution law of the signal can be systematically characterized, so as to fully explore the temporal transfer and arrangement pattern between the states of the signal, and thus more sensitively capture the weak dynamic mutations in the fault occurrence and development process; the systematic characterization process of the local fluctuation degree and temporal evolution law of the signal is shown in steps S4 and S5 below.

[0049] S4. For any vibration signal among multiple vibration signals, at different scales, map the vibration signal to a discrete state space to obtain discrete state sequences at different scales, and reconstruct the state vectors of each discrete state sequence to obtain reconstructed vectors at different scales; in specific applications, for example, but not limited to, the following steps S41 to S45 can be used to realize the discrete mapping of the vibration signal at different scales.

[0050] S41. Obtain a set of scale factors, wherein the set of scale factors includes multiple different scale factors; in a specific application, for example, the scale factors range from 1 to 10, which is equivalent to discretizing any vibration signal into blocks at 10 scales, and then performing discretization mapping on the obtained block sequence.

[0051] The block division process is shown in steps S42 to S44 below.

[0052] S42. Select the i-th scale factor from the set of scale factors as the division length, and divide any vibration signal into several signal blocks according to the division length, wherein the initial value of i is 1; in this embodiment, if the i-th scale factor is 2, then the division length is 2 to divide any vibration signal into several non-overlapping signal blocks; of course, when the scale factor is any other value, the signal division process is the same, and will not be described again here.

[0053] After obtaining multiple signal blocks, local fluctuation characteristic values ​​can be calculated, and based on the local fluctuation characteristic values ​​of each signal block, a block sequence under the i-th scale factor can be formed, as shown in step S43 below.

[0054] S43. Extract the local fluctuation feature value of each signal block, and use the local fluctuation feature value of each signal block to form a block sequence of the any vibration signal at the i-th scale factor. In specific applications, for example, but not limited to, the standard deviation of each signal block can be calculated as its corresponding local fluctuation feature value. In this way, the standard deviation of each signal block can be used to form the corresponding block sequence. Then, a new scale factor is selected, and the aforementioned steps S42 and S43 are repeated until all scale factors have been polled. Then, the block sequence of the any vibration signal at different scales can be obtained. The process is shown in step S44 below.

[0055] S44. Increment i by 1 and reselect the i-th scale factor from the scale factor set until i cycles from 1 to n, to obtain the block sequence of any vibration signal under different scale factors, where n is the total number of scale factors; in specific applications, when the scale factor is 1, its corresponding block sequence is the vibration signal itself.

[0056] Thus, based on the aforementioned steps, after dividing any vibration signal into blocks at different scales, the different block sequences can be mapped to the discrete state space, as shown in step S45 below.

[0057] S45. Discretize and encode each block sequence to obtain several initial discrete state sequences, and perform linear mapping on each initial discrete state sequence to obtain discrete state sequences at different scales after linear mapping. The length of any block sequence is the same as the length of the discrete state sequence corresponding to that block sequence.

[0058] In practical applications, the mapping process is illustrated by taking any segmented sequence as an example. For any segmented sequence, the standard deviation and mean of the segmented sequence can be calculated first, but are not limited to this example. Then, the segmented sequence is discretized and encoded based on the standard deviation and mean.

[0059] Optionally, for example, but not limited to, the following formula can be used to discretize and encode any block sequence.

[0060] ; In the formula, This represents the j-th local fluctuation feature value in any of the block sequences. express The corresponding discretized coded value, As an intermediate variable for integration, Let represent the standard deviation and mean of any of the block sequences, respectively, where ,and This indicates the length of any of the block sequences.

[0061] Thus, based on the aforementioned formula, after completing the discretization encoding of any block sequence, a linear mapping can be performed; for any initial discrete state sequence, for example, but not limited to, a linear mapping can be performed according to the following formula.

[0062] ; In the formula, express The linear mapping value, This is a rounding function. This represents the number of discretization levels (usually set to a positive integer between 6 and 8, which determines the granularity of the final state space). This represents the rounding offset (with a value of 0.5).

[0063] Thus, based on the aforementioned formula, after completing the linear mapping, the corresponding discrete state sequence can be formed by using the linear mapping value corresponding to each discretized code value in any initial discrete state sequence.

[0064] Of course, the mapping process of the other block sequences will not be described in detail in this embodiment.

[0065] Therefore, through the aforementioned steps S41 to S45, the vibration signal is mapped to the discrete state space at different scales to obtain discrete state sequences at different scales. Then, state vector reconstruction can be performed on each discrete state sequence. The role of state vector reconstruction is to reconstruct the phase space of the discrete state sequence to obtain multiple embedding vectors composed of linear mappings, and each embedding vector corresponds to a permutation and combination of state features.

[0066] The reconstruction process can be illustrated by taking any discrete state sequence as an example, and can be, but is not limited to, the steps S46 to S49 below.

[0067] S46. Obtain the reconstruction embedding dimension and embedding delay, and use the embedding delay as a fixed interval; in specific implementation, for example, the reconstruction embedding dimension is 2 or 3, and the embedding delay can be, but is not limited to, 1 or 2.

[0068] After obtaining the reconstructed embedding dimension and embedding delay, the reconstruction of any discrete state sequence can be performed based on these, as shown in steps S47 to S49 below.

[0069] S47. For any discrete state sequence, starting from the d-th element in the discrete state sequence, select c elements at fixed intervals to form the embedding vector corresponding to the d-th element using the selected c elements, where c represents the reconstructed embedding dimension and the initial value of d is 1.

[0070] S48. Increment d by 1, and starting from the d-th element in any discrete state sequence, select c elements forward at the fixed intervals until d equals 1. At that time, several embedding vectors are obtained, among which, This represents the length of any discrete state sequence. The fixed interval is defined as such.

[0071] In this embodiment, assuming that any discrete state sequence is [3,1,4,2,5,3,1,6,2,4], the reconstruction embedding dimension is 3, and the embedding delay is 2, then the upper limit of the index is: 10-(3-1)×2=6. Therefore, d can only take values ​​from 1 to 6.

[0072] When d is 1, we take d1, d3, d5, which means the embedding vector corresponding to the d-th element is (3,4,5); similarly, when d is 2, we take d2, d4, d6, which means the embedding vector corresponding to the d-th element is (1,2,3); in this way, we increment d until d equals 6, which gives us 6 embedding vectors; based on this, the aforementioned reconstructed embedding dimension determines the number of elements contained in each embedding vector, while the embedding delay determines the time span between the elements in the embedding vector.

[0073] After constructing several embedding vectors, the aforementioned embedding vectors can be used to form the scale-based reconstruction vector corresponding to any discrete state sequence, as shown in step S49 below.

[0074] S49. Using several embedding vectors, a scale-based reconstruction vector is formed for any discrete state sequence; in this embodiment, each embedding vector represents a specific state permutation and combination feature, and the same embedding vector corresponds to the same state permutation and combination feature.

[0075] Based on the aforementioned steps S41 to S49, after reconstructing each discrete state sequence of any vibration signal, a composite feature vector can be generated to characterize the local fluctuation degree and temporal evolution law of any vibration signal at multiple scales by statistically analyzing the state arrangement and combination characteristics of the reconstruction vectors corresponding to each discrete state sequence. The process is shown in step S5 below.

[0076] S5. Extract the state arrangement and combination features of each reconstruction vector in the reconstruction space, and generate a composite feature vector corresponding to any vibration signal based on the state arrangement and combination features of each reconstruction vector in the reconstruction space. After all vibration signals have been traversed, use each composite feature vector to form a composite correlation vector. Each composite feature vector is used to characterize the local fluctuation degree and temporal evolution law of the corresponding vibration signal at multiple scales.

[0077] In specific applications, for example, but not limited to, the following steps S51 to S54 can be used to generate the composite feature vector corresponding to any vibration signal.

[0078] S51. For any reconstructed vector, count the frequency of each embedded vector in the reconstructed vector. Based on the previous example, assuming the embedded vector is (3,4,5), then (3,4,5) corresponds to a state permutation and combination feature. Therefore, the frequency of the embedded vector (3,4,5) in any reconstructed vector can be counted. Then, based on this, the state pattern probability of the state permutation and combination feature corresponding to the embedded vector (3,4,5) can be calculated. The process is shown in step S52 below.

[0079] S52. Calculate the state pattern probability corresponding to each state permutation and combination feature based on the occurrence frequency of each embedding vector; in this embodiment, the occurrence frequency of each embedding vector is divided by the total number of embedding vectors to obtain the state pattern probability of each state permutation and combination feature corresponding to each embedding vector.

[0080] Thus, after calculating the state pattern probability corresponding to each state permutation and combination feature, the composite feature entropy of any reconstructed vector can be calculated based on this, as shown in step S53 below.

[0081] S53. Based on the state pattern probabilities corresponding to the permutation and combination features of each state, calculate the composite feature entropy corresponding to any reconstruction vector, and obtain the composite feature entropy of each reconstruction vector after traversing all reconstruction vectors; in specific applications, for example, but not limited to, the following formula can be used to calculate the aforementioned composite feature entropy.

[0082] ; In the formula, This represents the composite feature entropy corresponding to any of the reconstructed vectors. Indicates the first The probability of state patterns corresponding to the permutation and combination features of various states. This represents the total number of state permutation and combination features in any of the reconstructed vectors.

[0083] Thus, based on the aforementioned formula, the composite feature entropy of any reconstructed vector can be calculated. Then, by processing the remaining reconstructed vectors in the aforementioned manner, the composite feature entropy of the reconstructed vectors of any vibration signal at different scales can be obtained. Then, based on each composite feature entropy, the composite feature vector of any vibration signal can be formed, as shown in step S54 below.

[0084] S54. Using all composite feature entropies, generate a composite feature vector corresponding to any vibration signal.

[0085] The composite feature entropy can characterize the local fluctuation degree and temporal evolution law of the vibration signal at multiple scales from the following three dimensions: (1) The composite feature entropy is calculated by dividing the signal into blocks at different scales. The standard deviation is used to construct the block sequences at different scales. The standard deviation directly reflects the dispersion and fluctuation amplitude of the signal within the time period. The correlation with the fault is as follows: taking a bearing as an example, when the bearing has excessive bearing clearance, wear, or impact faults, the vibration signal will generate violent impact pulses. These pulses will cause the standard deviation of the local segment to increase sharply. Therefore, the magnitude of the composite feature entropy value directly reflects the strength of the impact energy at this time scale. The stronger the impact, the greater the numerical fluctuation of the coarse-grained sequence, and the richer the basic information for subsequent mapping and entropy value calculation.

[0086] (2) It reflects the complexity and disorder of vibration waveforms at different scales; the composite feature entropy is calculated by the probability of state mode of state arrangement and combination features, and the uniformity of the state mode probability distribution determines the entropy value. If the vibration signal at this scale is close to a single regular sine wave or periodic impact (high mode repeatability and concentrated probability), the feature entropy value is low; if the signal contains a large amount of random noise, multi-frequency component aliasing or non-periodic fluctuation (multiple modes and dispersed probability), the feature entropy value is high; its correlation with faults is as follows: the vibration of the bearing of a normally operating compressor is relatively stable and regular, and the feature entropy value is usually low and stable; while the fault state (such as large gap) will produce modulation phenomenon, making the vibration waveform lose its regularity, and the arrangement and combination mode of the time series becomes chaotic and complex, resulting in a significant increase in the feature entropy value or different change trends.

[0087] (3) Characterize the multi-scale dynamic sequence characteristics at different scales (derived from the scale factor); when the scale factor is small, the complexity of the original high-frequency and detailed information is extracted, while when the scale factor is large, that is, the block window is lengthened, the complexity of the low-frequency and trend information is extracted; and the correlation with faults is: the reciprocating compressor is affected by the reciprocating motion of the piston, and its fault characteristics are often distributed on a specific fundamental frequency and its harmonics. Different types of faults (such as large-end tile faults vs. small-end tile faults) have different sensitivities to high-frequency impacts and low-frequency vibrations. Therefore, the curve of the characteristic entropy changing with the scale factor reflects the distribution law of fault characteristics in different frequency bands; if the curve shows abnormally high values ​​or abrupt changes at certain scales, it indicates that there is significant fault modulation energy in that frequency band.

[0088] In summary, the larger the characteristic entropy, the more significant the signal fluctuations (large standard deviation) at that scale, and the more abundant and non-repeating temporal arrangement patterns (strong randomness) it contains—common in severe gap friction or fluid excitation turbulence; conversely, the smaller the characteristic entropy, the more gradual the signal fluctuations at that scale, and the more repetitive they are (strong periodicity) it contains—common in normal operation or when only weak low-frequency interference exists.

[0089] Based on the aforementioned steps S51 to S54, the composite feature vectors of all vibration signals can be extracted. Then, the composite feature vectors can be spliced ​​together to form a composite correlation vector.

[0090] After extracting the three features, a fault feature vector for the reciprocating compressor can be formed based on this. The fault feature vector can then be used to identify faults in the reciprocating compressor. The process is shown in step S6 below.

[0091] S6. A fault feature vector is generated using physical prior feature vectors, multi-scale temporal deep feature vectors, and composite correlation vectors. This fault feature vector is then input into a fault detection model to perform fault classification and early warning for the reciprocating compressor based on the fault detection results. In this embodiment, before concatenating the aforementioned three feature vectors, batch normalization can be performed for scale alignment. After obtaining the fault feature vector, it can be compressed to a unified dimension using a fully connected layer (Linear) and an activation function (ReLU) to extract the most representative principal components and reduce the risk of overfitting. Finally, the obtained features are input into a pre-trained fault detection model (which may, but is not limited to, use an RBF neural network) to obtain the fault detection results for the reciprocating compressor, i.e., different fault types are detected (such as valve faults, piston faults, etc.). Based on this, a graded early warning can be performed according to the pre-set warning level for each detected fault type.

[0092] Therefore, through the multi-parameter fusion fault feature extraction and early warning method for reciprocating compressors described in detail in steps S1 to S6 above, this invention combines multi-point vibration signal splicing with a multi-scale feature extraction mechanism to simultaneously capture the trend features of signals at different scales, significantly improving the multi-resolution representation capability of non-stationary vibration signals and overcoming the shortcomings of traditional single-scale methods in comprehensively representing complex fault modes. Simultaneously, time-frequency physical prior features are independently extracted for each measuring point signal and subjected to deep nonlinear mapping via a feature mapping network. This allows domain knowledge with clear physical meaning to be explicitly injected into the fault detection model in the form of feature vectors. This retains the advantages of strong stability and good interpretability of physical features while automatically mining complex coupling relationships between features through neural networks, effectively compensating for the insufficient generalization ability caused by the "black box" characteristics of purely data-driven models. Furthermore, this invention pioneers a composite feature extraction strategy based on discrete state space mapping. Specifically, vibration signals are mapped to discrete state spaces at different scales. Through state vector reconstruction and permutation combination feature extraction, a composite feature vector characterizing the local fluctuation degree and temporal evolution law of the signal is generated. This fully explores the temporal transitions and permutation patterns between states, enabling sensitive capture of subtle dynamic mutations during fault occurrence and development. Finally, the above three features are fused to form a fault feature vector, which is then input into a fault detection model. Based on the detection results, graded early warning for reciprocating compressors can be achieved. Therefore, a clear decision-making basis can be provided for engineering operation and maintenance.

[0093] In one possible design, the second aspect of this embodiment, based on the first aspect of the embodiment, provides a construction process for an RBF neural network (which includes an input layer, a hidden layer, and an output layer), wherein the process of the fault detection model may be, but is not limited to, the steps shown below.

[0094] Step 1: Obtain the training set, which contains multiple sample fault feature vectors of reciprocating compressors. It also contains the label data corresponding to each sample fault feature vector, i.e., the specific fault type or no fault.

[0095] Step 2: A clustering algorithm based on local distribution is used to cluster the fault features of multiple samples in the training set to obtain the number of cluster centers. Based on the number of nodes in the input layer, the number of nodes in the output layer, and the number of cluster centers, the parameter dimensions of the initial fault detection model are determined. In practical applications, the key parameters of the RBF neural network mainly include the number of nodes in the input layer, hidden layer, and output layer; the center and width of each basis function in the hidden layer; and the weights of the output layer. Determining the number of nodes in the input and output layers is relatively simple, as it is determined by the dimension of the input data and the number of fault types, respectively. However, in traditional techniques, the number of hidden layer nodes needs to be determined manually in advance, which has strong randomness and seriously affects the performance of the neural network. Therefore, this embodiment provides a clustering algorithm based on local distribution to determine the number of hidden layer nodes.

[0096] The process of determining the number of hidden layer nodes (i.e., the number of cluster centers) is as follows.

[0097] Step 21: Initialize the cluster counter, where the initial value of the cluster counter is 0, and the count of the cluster counter is used to represent the number of cluster centers.

[0098] Step 22: Obtain the sample set at the q-th clustering and the cluster confidence corresponding to the cluster centers at the (q-1)-th clustering, wherein when q is 1, the sample set at the q-th clustering is the training set, and the cluster confidence corresponding to the cluster centers at the (q-1)-th clustering is the initial value; in this embodiment, the initial value is 0.

[0099] Thus, after obtaining the sample set at the q-th clustering and the cluster confidence corresponding to the cluster center at the (q-1)-th clustering, the local distribution index and local density index of each sample in the sample set can be calculated based on these, as shown in steps 23 and 24 below.

[0100] Step 23: For any sample fault feature in the sample set during the q-th clustering, obtain the neighborhood fault features of the any sample fault feature, and calculate the local distribution index of the any sample fault feature based on the neighborhood fault features. In this embodiment, a neighborhood radius can be preset. Then, the remaining sample fault features whose distance from the any sample fault feature is less than or equal to the neighborhood radius are taken as the neighborhood fault features of the any sample fault feature. Then, the average distance between the any sample fault feature and each neighborhood fault feature is calculated. Finally, the reciprocal of the average value is taken as its corresponding local distribution index. In this way, the local distribution index measures the compactness of its local data distribution.

[0101] After obtaining the local distribution index, the confidence of the cluster centers from the previous clustering can be used to calculate the local density index, as shown in step 24 below.

[0102] Step 24: Calculate the local density index corresponding to the fault features of any sample by using the neighborhood fault features and the cluster confidence corresponding to the cluster center at the (q-1)th clustering. In specific applications, the following formula can be used, but is not limited to, to calculate the aforementioned local density index.

[0103] ; In the formula, This represents the local density index corresponding to the fault characteristics of any of the samples. This represents the fault characteristics of any of the samples. This represents the first fault feature among the various neighboring fault features corresponding to any given sample fault feature. There are _nearby_fault_features, where G represents the total number of neighborhood_fault_features. Let r represent the L2 norm and r represent the neighborhood radius. This represents the cluster confidence score corresponding to the cluster centers in the (q-1)th clustering iteration. The cluster centers are represented in the (q-1)th clustering iteration. It is an intermediate variable, and .

[0104] Based on the aforementioned formula, after calculating the local density index of any sample fault feature, the clustering confidence of any sample fault feature can be calculated by combining the aforementioned local distribution index, as shown in step 25 below.

[0105] Step 25: Based on the local density index and the local distribution index, calculate the cluster confidence of any sample fault feature. After traversing all sample fault features in the sample set at the qth clustering, obtain the cluster confidence of each sample fault feature in the sample set. In this embodiment, the product of the local density index and the local distribution index is used as the cluster confidence of any sample fault feature. Thus, after calculating the cluster confidence of each sample fault feature in the sample set at the qth clustering based on the aforementioned method, the sample corresponding to the largest cluster confidence can be selected as the cluster center for this clustering, as shown in Step 26 below.

[0106] Step 26: Select the fault feature with the highest clustering confidence from the fault features of each sample in the sample set as the cluster center for the qth clustering, and increment the clustering counter by 1. In this embodiment, after determining the cluster center for the qth clustering, the number of clusters can be counted, and the clustering counter can be incremented by 1. Then, it can be determined whether the clustering stopping condition is met, as shown in Step 27 below.

[0107] Step 27: Determine whether the clustering stopping condition is met; In this embodiment, calculate the ratio between the cluster confidence of the cluster center at the qth clustering and the cluster confidence of the cluster center at the 1st clustering; Then, determine whether the ratio is less than a preset threshold (e.g., 0.5). If so, it is determined that the clustering stopping condition is met; otherwise, clustering needs to continue, as shown in steps 28 and 29 below.

[0108] Step 28: If not, then delete the cluster centers from the sample set of the qth clustering to obtain the sample set of the (q+1)th clustering. In this embodiment, after deleting the cluster centers from the sample set of the qth clustering, the sample set for the next clustering can be obtained. Then, steps 22 to 28 can be repeated until the clustering stopping condition is met. Then, the number of cluster centers can be obtained based on the count of the clustering counter, as shown in step 29 below.

[0109] Step 29. Increment q by 1 and reacquire the sample set for the qth clustering until the clustering stopping condition is met. Determine the number of cluster centers based on the total count of the clustering counter when the clustering stopping condition is met.

[0110] Based on steps 22 to 29 above, after obtaining the number of cluster centers, the number of cluster centers can be used as the number of hidden layer nodes in the RBF neural network. Then, the parameter dimension can be calculated by combining the number of nodes in the input layer and the output layer. The parameter dimension is: number of cluster centers S1 × (number of input layer nodes S2 + 1 + number of output layer nodes S3). Thus, the parameter dimension contains three segments: the first segment is S1 × S2, the second segment is S1, and the last segment is S1 × S3.

[0111] After calculating the parameter dimensions, the model parameters can be initialized based on these dimensions, as shown in step 3 below.

[0112] Step 3: Based on the parameter dimensions, perform parameter initialization processing on the initial fault detection model to obtain several sets of initial model parameters, and use these sets of initial model parameters to form an initial population. In this embodiment, the initial model parameters may include, but are not limited to, the radial basis function center, the radial basis function width, and the output layer weights. Therefore, based on the aforementioned initial model parameters, an initial population can be formed.

[0113] Specifically, the first segment (i.e., S1×S2 dimensions) in an initial individual corresponds to the center of the basis function, the second segment (i.e., S1 dimensions) corresponds to the width of the basis function, and the third segment (i.e., S1×S3 dimensions) corresponds to the connection weights of the output layer.

[0114] After constructing the initial population based on the aforementioned steps, swarm intelligence algorithms can be used to optimize parameters, as shown in steps 4 to 11 below.

[0115] Step 4: Obtain the individual population at iteration t and the prey set at iteration (t-1). When t is 1, the individual population at iteration t is the initial population, and the prey set at iteration (t-1) is the initial prey set. Any initial prey is used to represent a set of initial model parameters, and the initial prey set is the initial population. In this embodiment, the individual population represents the current model parameters, while the prey set represents candidate model schemes that have performed well in the past, which are the "targets" (which are also schemes) of the individuals. Therefore, the prey set at iteration (t-1) is the initial population itself. Thus, after obtaining the individual population at iteration t and the prey set at iteration (t-1), the prey set at the current iteration can be determined based on this, as shown in Step 5 below.

[0116] Step 5: Using the individual population and prey set at the t-th iteration, form a candidate population, and calculate the fitness of each candidate individual in the candidate population based on the training set. Based on the fitness of each candidate individual, determine the prey set at the t-th iteration, where the number of prey in the prey set decreases as the number of iterations increases.

[0117] In practice, a pre-selected fault detection model is constructed for each candidate individual based on its corresponding model parameters. Then, each pre-selected fault detection model is trained using a training set. Based on the output of each pre-selected fault detection model, the fitness of each individual is calculated. Specifically, the mean squared error (MSE) can be calculated based on the output of the pre-selected fault detection model (i.e., the classification result), and the reciprocal of the MSE is taken as the fitness. Therefore, the higher the fitness, the better the candidate individual. After calculating the fitness of each candidate individual, the prey set for the current iteration can be determined based on this. The process is as follows: First, obtain the maximum number of iterations and the maximum number of prey; then, based on the maximum number of iterations, the maximum number of prey, and the iteration number t, calculate the number of prey at the t-th iteration; where the number of prey at the t-th iteration is: In the formula, These represent the maximum number of prey and the maximum number of iterations, respectively. This is a rounding function; thus, as can be seen from the aforementioned formula, the number of prey gradually decreases in each iteration, which can continuously narrow down the candidate range of model schemes.

[0118] Then, the candidate individuals are sorted in descending order of fitness to obtain a sorted sequence; finally, the top performers can be selected from the sorted sequence. There are _n_ candidate individuals, which form the prey combination in the _t_th iteration, where _n_... Let t be the prey in the t-th iteration; based on this, in each iteration, selection is made according to fitness, that is, selecting the better individuals to form the prey set for this iteration.

[0119] After obtaining the prey set at the t-th iteration, the prey allocation for individuals at the t-th iteration can be carried out, as shown in step 6 below.

[0120] Step 6: Based on the prey set at iteration t, determine the prey corresponding to each individual in the individual population at iteration t. In this embodiment, first calculate the fitness of each individual at iteration t, then sort the individual population at iteration t in descending order of fitness to obtain the individual sequence. Next, assign the h-th prey in the prey set at iteration t to the h-th individual in the individual sequence (h is initially set to 1). Then, increment h by 1 and redistribute the prey until h equals 1. At that time, the prey corresponding to each individual in the population at the t-th iteration is determined; where, if If the number of individuals is less than the total number of individuals in the sequence, then the sequence is ordered... The remaining individuals are then uniformly assigned to the last prey in the prey set, i.e., the worst prey.

[0121] After assigning prey to each individual in the t-th iteration, it can be determined whether the iteration stopping condition is met, as shown in step 7 below.

[0122] Step 7: Determine whether the iteration stopping condition is met; In this embodiment, the iteration stopping condition can be, but is not limited to, reaching the maximum number of iterations t, or the fitness of the first prey (i.e. the prey with the highest fitness) in the prey set of the t-th iteration being greater than or equal to the fitness threshold; If the aforementioned iteration stopping condition is not met, a position update is required, the process of which is shown in steps 8 and 9 below.

[0123] Step 8. If not, calculate the hunting step length at the t-th iteration and the hunting intensity of each individual at the t-th iteration. In specific applications, this embodiment introduces a hunting step length to make individuals focus more on exploration at different stages of the search process. For example, but not limited to, the hunting step length can be calculated using the following formula.

[0124] ; In the formula, This represents the hunting step size in the t-th iteration. This is a hyperparameter that can be set according to actual use, such as 0.1 or 0.5.

[0125] Meanwhile, this embodiment introduces the individual's hunting intensity to simulate the dynamic consumption of an individual's physical strength as time and location change during the hunting process; for example, but not limited to, the hunting intensity of any individual at the t-th iteration can be calculated according to the following formula.

[0126] ; In the formula, This indicates the hunting intensity of any of the individuals mentioned. Let be the initial hunting intensity of any individual at the t-th iteration (a random number between (0,1)). This represents the hunting stamina coefficient of any individual. The formula for calculating it, which decreases linearly from 1 to 0, is as follows: , These represent the time influence coefficient and the location influence coefficient (preset values, such as 0.15 and 0.5), respectively. This refers to any of the individuals mentioned.

[0127] Thus, based on the aforementioned formula, after calculating the hunting stride length and hunting intensity, the position of each individual can be updated accordingly, as shown in step 9 below.

[0128] Step 9: Based on the hunting step length, the hunting intensity of each individual at iteration t, and their corresponding prey, update the position of each individual at iteration t to obtain the individual population at iteration (t+1). In practical application, for any individual at iteration t, first determine whether the hunting intensity of that individual is greater than the initial hunting intensity of that individual. If so, calculate the hunting distance between that individual and its corresponding prey, and update the position of that individual based on the hunting distance and the hunting step length to obtain the updated individual corresponding to that individual.

[0129] Optionally, the position update formula is: ; In the formula, Represents any of the individuals The corresponding updated individual, This represents the hunting step size in the t-th iteration. This indicates the hunting distance. This is the hunting coefficient, with a value of 1. A random number between [-1, 1] This refers to the prey corresponding to any of the individuals.

[0130] Wherein, when the hunting intensity of any individual is less than or equal to the initial hunting intensity of any individual, a random jump update strategy is adopted to update the position of any individual to obtain the updated individual corresponding to any individual.

[0131] Specifically, the random jump update formula is as follows: ; In the formula, For a random number between [0,1]; thus, Mapping the jump random number to [-1, 1] allows an individual to jump randomly in both directions (left and right or up and down) with the prey as the origin, thereby exploring a larger area of ​​unknown territory.

[0132] After updating the position of each individual, the aforementioned steps can be repeated until the iteration stopping condition is met, at which point the optimal model parameters can be obtained. The process is shown in step 10 below.

[0133] Step 10: Increment t by 1, and reacquire the individual population at iteration t and the prey set at iteration (t-1) until the iteration stopping condition is met. The optimal model parameters are determined based on the prey corresponding to the individual with the highest fitness in the individual population at the iteration stopping condition. In this embodiment, the prey set retains individuals with better fitness each time. Therefore, during the continuous iteration process, the historical best individuals will accumulate continuously, and the prey is allocated according to fitness. Based on this, when the iteration stopping condition is met, the model parameters of the prey corresponding to the individual with the highest fitness (that is, the first-ranked prey in the prey set at the iteration stopping condition) can be used as the optimal model parameters.

[0134] After obtaining the optimal model parameters, a fault detection model can be constructed based on them, as shown in step 11 below.

[0135] Step 11: Update the initial fault detection model using the optimal model parameters to obtain the fault detection model; In this embodiment, after updating the initial fault detection model using the optimal model parameters, the model can be retrained using the training data, that is, the fault detection model can be trained by taking the sample fault feature vector as input and the corresponding fault detection result as output.

[0136] Thus, through the aforementioned design, this invention allows "short-stressed" shells to make random large jumps by introducing hunting stride length and hunting intensity. In this way, population diversity is greatly maintained, and all individuals are prevented from prematurely gathering at local extreme points. Therefore, local optima are avoided, and the optimization accuracy of model parameters is improved.

[0137] Once the fault detection model is obtained, fault detection of reciprocating compressors can be performed based on it, thereby achieving graded early warning based on the detection results.

[0138] like Figure 3 As shown, the third aspect of this embodiment provides a software system for implementing the multi-parameter fusion fault feature extraction and early warning method for reciprocating compressors described in the first and second aspects of the embodiments, comprising: The acquisition unit is used to acquire vibration signals corresponding to different measuring points on the reciprocating compressor.

[0139] The multi-scale feature extraction unit is used to splice multiple vibration signals to obtain a spliced ​​signal, and to extract multi-scale temporal features from the spliced ​​signal to obtain a multi-scale temporal depth feature vector of the reciprocating compressor.

[0140] The physical prior feature extraction unit is used to extract time-frequency features from each vibration signal to obtain the time-frequency features of each vibration signal. Then, using a feature mapping network, nonlinear mapping is performed on each time-frequency feature to obtain the physical mapping features of each vibration signal. Based on the physical mapping features, the physical prior feature vector of the reciprocating compressor is generated.

[0141] The reconstruction unit is used to map any vibration signal from multiple vibration signals to a discrete state space at different scales to obtain discrete state sequences at different scales, and to reconstruct the state vectors of each discrete state sequence to obtain reconstruction vectors at different scales.

[0142] The composite feature extraction unit is used to extract the state arrangement and combination features of each reconstruction vector in the reconstruction space, so as to generate a composite feature vector corresponding to any vibration signal based on the state arrangement and combination features of each reconstruction vector in the reconstruction space, and after all vibration signals have been traversed, a composite correlation vector is formed by using each composite feature vector. Each composite feature vector is used to characterize the local fluctuation degree and temporal evolution law of the corresponding vibration signal at multiple scales.

[0143] The fault early warning unit is used to generate fault feature vectors by utilizing physical prior feature vectors, multi-scale time-series deep feature vectors, and composite correlation vectors, and inputs the fault feature vectors into the fault detection model to perform fault classification and early warning for reciprocating compressors based on the fault detection results.

[0144] The working process, working details and technical effects of the module provided in this embodiment can be found in the first and second aspects of the embodiment, and will not be repeated here.

[0145] like Figure 4 As shown, the fourth aspect of this embodiment provides a device for extracting and warning of multi-parameter fusion fault features of reciprocating compressors. Taking the device as an electronic device as an example, it includes: a memory, a processor, and a transceiver that are connected in sequence. The memory is used to store a computer program, the transceiver is used to send and receive messages, and the processor is used to read the computer program and execute the method for extracting and warning of multi-parameter fusion fault features of reciprocating compressors as described in the first and second aspects of the embodiments.

[0146] For specific examples, the memory may include, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, first-in-first-out (FIFO) memory, and / or first-in-last-out (FILO) memory, etc.; specifically, the processor may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor may be implemented using at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), PLA (Programmable Logic Array). The processor may also include a main processor and a coprocessor. The main processor, also known as the CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state.

[0147] In some embodiments, the processor may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. For example, the processor may not be limited to microprocessors of the STM32F105 series, reduced instruction set computer (RISC) microprocessors, x86 architecture processors, or processors with integrated neural network processing units (NPUs). The transceiver may be, but is not limited to, a Wi-Fi transceiver, a Bluetooth transceiver, a General Packet Radio Service (GPRS) transceiver, a ZigBee (a low-power LAN protocol based on the IEEE 802.15.4 standard) transceiver, a 3G transceiver, a 4G transceiver, and / or a 5G transceiver. Furthermore, the device may also include, but is not limited to, a power module, a display screen, and other necessary components.

[0148] The working process, working details and technical effects of the electronic device provided in this embodiment can be found in the first aspect of the embodiment, and will not be repeated here.

[0149] The fifth aspect of this embodiment provides a storage medium that stores instructions containing the multi-parameter fusion fault feature extraction and early warning method for reciprocating compressors as described in the first and second aspects of the embodiments. That is, the storage medium stores instructions that, when executed on a computer, perform the multi-parameter fusion fault feature extraction and early warning method for reciprocating compressors as described in the first and second aspects of the embodiments.

[0150] The storage medium refers to a carrier for storing data, which may include, but is not limited to, floppy disks, optical disks, hard disks, flash memory, USB flash drives, and / or memory sticks. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.

[0151] The working process, working details and technical effects of the storage medium provided in this embodiment can be found in the first and second aspects of the embodiment, and will not be repeated here.

[0152] The sixth aspect of this embodiment provides a computer program product containing instructions that, when executed on a computer, cause the computer to perform the reciprocating compressor multi-parameter fusion fault feature extraction and early warning method as described in the first and second aspects of the embodiments, wherein the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.

[0153] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for multi-parameter fusion fault feature extraction and early warning of reciprocating compressors, characterized in that, include: Obtain vibration signals corresponding to different measuring points on the reciprocating compressor; Multiple vibration signals are spliced ​​together to obtain a spliced ​​signal, and multi-scale time-series feature extraction is performed on the spliced ​​signal to obtain the multi-scale time-series depth feature vector of the reciprocating compressor. Time-frequency features are extracted from each vibration signal to obtain the time-frequency features of each vibration signal. Then, a feature mapping network is used to perform nonlinear mapping on each time-frequency feature to obtain the physical mapping features of each vibration signal. Based on the physical mapping features, the physical prior feature vector of the reciprocating compressor is generated. For any vibration signal among multiple vibration signals, at different scales, the vibration signal is mapped to a discrete state space to obtain a discrete state sequence at different scales, and the state vector of each discrete state sequence is reconstructed to obtain a reconstructed vector at different scales. Extract the state arrangement and combination features of each reconstruction vector in the reconstruction space, and generate a composite feature vector corresponding to any vibration signal based on the state arrangement and combination features of each reconstruction vector in the reconstruction space. After all vibration signals have been traversed, a composite correlation vector is formed by using each composite feature vector. Each composite feature vector is used to characterize the local fluctuation degree and temporal evolution law of the corresponding vibration signal at multiple scales. Fault feature vectors are generated by using physical prior feature vectors, multi-scale time-series deep feature vectors, and composite correlation vectors. These fault feature vectors are then input into the fault detection model to perform fault classification and early warning for reciprocating compressors based on the fault detection results.

2. The method according to claim 1, characterized in that, A multi-scale feature extraction model is used to extract multi-scale temporal features from the spliced ​​signal to obtain the multi-scale temporal depth feature vector of the reciprocating compressor. The multi-scale feature extraction model includes: a local feature extraction layer, a multi-scale temporal residual layer, a feature projection layer, and an adaptive pooling layer. The multi-scale temporal residual layer includes a feature concatenation unit, a residual unit, and multiple temporal feature extraction units. The local feature extraction layer is used to perform one-dimensional convolution and max pooling on the spliced ​​signal in sequence to obtain local feature vectors, and then input the local feature vectors into the residual unit and each temporal feature extraction unit respectively. For any one of the multiple temporal feature extraction units, the temporal feature extraction unit is used to perform a sliding window-style local convolution operation on the received local feature vector along the temporal dimension using a one-dimensional convolution kernel of a preset size and a preset stride, to obtain a convolution response feature vector; Each temporal feature extraction unit is used to perform batch normalization on the convolutional response feature vector to obtain a normalized response feature vector, and to use the ReLU activation function to perform nonlinear activation mapping on the normalized response feature vector to obtain a temporal feature vector, which is then input to the feature concatenation unit. The size of the convolutional kernel of each temporal feature extraction unit is different to output temporal feature vectors of different scales. The residual unit is used to perform 1×1 convolution and batch normalization on the received local feature vectors to obtain residual connection features. The feature concatenation unit is used to concatenate the temporal feature vectors output by each temporal feature extraction unit along the channel dimension to obtain a temporal concatenation feature vector. It also adds the temporal concatenation feature vector to the residual connection feature vector element by element and performs nonlinear activation processing on the addition result to obtain a multi-scale temporal fusion feature vector after nonlinear activation processing. The feature projection layer is used to perform 1×1 convolution processing on the multi-scale temporal fusion feature vector to adjust the number of channels of the multi-scale temporal fusion feature vector to a preset dimension, so as to obtain the mapped feature vector. An adaptive pooling layer is used to perform adaptive pooling processing on the mapped feature vector to obtain the multi-scale temporal depth feature vector after adaptive pooling processing.

3. The method according to claim 1, characterized in that, Mapping any vibration signal to a discrete state space at different scales yields discrete state sequences at different scales, including: Obtain a set of scale factors, wherein the set of scale factors includes multiple different scale factors; Select the i-th scale factor from the set of scale factors as the division length, and divide any vibration signal into several signal blocks according to the division length, wherein the initial value of i is 1; Extract the local fluctuation feature value of each signal block, and use the local fluctuation feature value of each signal block to form a block sequence of any vibration signal under the i-th scale factor; Increment i by 1 and reselect the i-th scale factor from the scale factor set until i is circulated from 1 to n, to obtain the block sequence of any vibration signal under different scale factors, where n is the total number of scale factors; Each block sequence is discretized and encoded to obtain several initial discrete state sequences. Each initial discrete state sequence is then linearly mapped to obtain discrete state sequences at different scales. The length of any block sequence is the same as the length of the discrete state sequence corresponding to that block sequence. Accordingly, reconstructing the state vectors of each discrete state sequence yields reconstructed vectors at different scales, including: Obtain the reconstructed embedding dimension and embedding delay, and use the embedding delay as a fixed interval; For any discrete state sequence, starting from the d-th element in the discrete state sequence, select c elements at fixed intervals to form the embedding vector corresponding to the d-th element using the selected c elements, where c represents the reconstructed embedding dimension and the initial value of d is 1. Increment d by 1, and starting from the d-th element in any discrete state sequence, select c elements forward at the fixed intervals until d equals 1. At that time, several embedding vectors are obtained, among which, This represents the length of any discrete state sequence. The fixed interval; Using several embedding vectors, a scale-based reconstruction vector is formed for any discrete state sequence.

4. The method according to claim 3, characterized in that, Each block sequence is discretized and encoded to obtain several initial discrete state sequences, including: For any block sequence, calculate the standard deviation and mean of the given block sequence; Based on the standard deviation and mean, and according to the following formula, the any block sequence is discretized and encoded. ； In the formula, This represents the j-th local fluctuation feature value in any of the block sequences. express The corresponding discretized coded value, As an intermediate variable for integration, Let represent the standard deviation and mean of any of the block sequences, respectively, where ,and Indicates the length of any of the block sequences; Accordingly, a linear mapping is performed on each initial discrete state sequence to obtain discrete state sequences at different scales after the linear mapping, which includes: For any initial discrete state sequence, a linear mapping is performed on the sequence according to the following formula; ； In the formula, express The linear mapping value, This is a rounding function. Represents the number of discretization levels. The offset represents the rounding off, wherein the corresponding discrete state sequence is formed by using the linear mapping value corresponding to each discretized coded value in any initial discrete state sequence.

5. The method according to claim 1, characterized in that, Any reconstructed vector contains several embedding vectors, each embedding vector corresponds to a state permutation and combination feature, and the same embedding vector corresponds to the same state permutation and combination feature; Accordingly, based on the state arrangement and combination characteristics of each reconstruction vector in the reconstruction space, a composite feature vector corresponding to any vibration signal is generated, including: For any reconstructed vector, the frequency of occurrence of each embedded vector in the reconstructed vector is calculated. Based on the frequency of occurrence of each embedding vector, the probability of the state pattern corresponding to each state permutation and combination feature is calculated. Based on the state pattern probability corresponding to each state permutation and combination feature, the composite feature entropy corresponding to any reconstruction vector is calculated, and after all reconstruction vectors have been traversed, the composite feature entropy of each reconstruction vector is obtained. Using all the composite feature entropies, a composite feature vector corresponding to any vibration signal is generated; The composite feature entropy corresponding to any reconstructed vector is calculated according to the following formula; ； In the formula, This represents the composite feature entropy corresponding to any of the reconstructed vectors. Indicates the first The probability of state patterns corresponding to the permutation and combination features of various states. This represents the total number of state permutation and combination features in any of the reconstructed vectors.

6. The method according to claim 1, characterized in that, The fault detection model consists of an input layer, a hidden layer, and an output layer. The fault detection model is constructed in the following manner. Obtain a training set, which contains multiple sample fault feature vectors of reciprocating compressors; A clustering algorithm based on local distribution is used to cluster the fault features of multiple samples in the training set to obtain the number of cluster centers. The parameter dimensions of the initial fault detection model are determined based on the number of nodes in the input layer, the number of nodes in the output layer, and the number of cluster centers. Based on the parameter dimensions, the initial fault detection model is initialized to obtain several sets of initial model parameters, and an initial population is formed using these sets of initial model parameters. Obtain the individual population at the t-th iteration and the prey set at the (t-1)-th iteration, wherein when t is 1, the individual population at the t-th iteration is the initial population, and the prey set at the (t-1)-th iteration is the initial prey set. Any initial prey is used to characterize a set of initial model parameters, and the initial prey set is the initial population. Using the individual population and prey set at the t-th iteration, a candidate population is formed. Based on the training set, the fitness of each candidate individual in the candidate population is calculated. Based on the fitness of each candidate individual, the prey set at the t-th iteration is determined. The number of prey in the prey set decreases as the number of iterations increases. Based on the prey set at the t-th iteration, determine the prey corresponding to each individual in the individual population at the t-th iteration; Determine if the iteration stopping condition is met; If not, calculate the hunting step size at the t-th iteration and the hunting intensity of each individual at the t-th iteration. Based on the hunting stride length, the hunting intensity of each individual at the t-th iteration and their corresponding prey, the position of each individual at the t-th iteration is updated to obtain the individual population at the (t+1)-th iteration. Increment t by 1, and reacquire the individual population at the t-th iteration and the prey set at the (t-1)-th iteration until the iteration stopping condition is met. Then, determine the optimal model parameters based on the prey corresponding to the individual with the highest fitness in the individual population at the time the iteration stopping condition is met. The initial fault detection model is updated using the optimal model parameters to obtain the fault detection model.

7. The method according to claim 1, characterized in that, A clustering algorithm based on local distribution is used to cluster the fault features of multiple samples in the training set to obtain the number of cluster centers, including: Initialize the cluster counter, where the initial value of the cluster counter is 0, and the count of the cluster counter is used to represent the number of cluster centers; Obtain the sample set at the q-th clustering and the cluster confidence corresponding to the cluster center at the (q-1)-th clustering, wherein when q is 1, the sample set at the q-th clustering is the training set, and the cluster confidence corresponding to the cluster center at the (q-1)-th clustering is the initial value. For any sample fault feature in the sample set during the q-th clustering, obtain the neighborhood fault features of the fault feature ... Using the neighborhood fault features and the cluster confidence corresponding to the cluster center at the (q-1)th clustering, the local density index corresponding to the fault features of any sample is calculated. Based on the local density index and the local distribution index, the clustering confidence of any sample fault feature is calculated, and after all sample fault features in the sample set at the qth clustering time have been traversed, the clustering confidence of each sample fault feature in the sample set is obtained. From the fault features of each sample in the sample set, select the fault feature with the highest clustering confidence as the cluster center for the qth clustering, and increment the clustering counter by 1. Determine whether the clustering stopping condition is met; If not, then from the sample set of the qth clustering, delete the cluster centers of the qth clustering to obtain the sample set of the (q+1)th clustering. Increment q by 1 and reacquire the sample set for the qth clustering iteration until the clustering stopping condition is met. The number of cluster centers is determined based on the total count of the clustering counter when the clustering stopping condition is met.

8. The method according to claim 6, characterized in that, Based on the fitness of each candidate individual, the prey set for the t-th iteration is determined, including: Find the maximum number of iterations and the maximum number of prey. Calculate the number of prey in the t-th iteration based on the maximum number of iterations, the maximum number of prey, and the number of iterations t. The candidate individuals are sorted in descending order of fitness to obtain the sorted sequence; Filter the top from the sorted sequence There are _n_ candidate individuals, which form the prey combination in the _t_th iteration, where _n_... Let be the number of prey in the t-th iteration.

9. The method according to claim 6, characterized in that, Calculate the hunting step size at the t-th iteration, and calculate the hunting intensity of each individual at the t-th iteration, including: The hunting intensity of any individual at the t-th iteration is calculated using the following formula. ； In the formula, This indicates the hunting intensity of any of the individuals mentioned. Let be the initial hunting intensity of any individual in the t-th iteration. This represents the hunting stamina coefficient of any individual. Decrease linearly from 1 to 0. Let represent the time influence coefficient and the location influence coefficient, respectively. Represents any of the individuals mentioned; Accordingly, based on the hunting step length, the hunting intensity of each individual at the t-th iteration, and their respective prey, the position of each individual at the t-th iteration is updated, including: For any individual at the t-th iteration, determine whether the hunting intensity of that individual is greater than the initial hunting intensity of that individual; If so, the hunting distance between any individual and its corresponding prey is calculated, and the position of any individual is updated based on the hunting distance and the hunting stride length to obtain the updated individual corresponding to any individual. Otherwise, a random jump update strategy is used to update the position of any individual to obtain the updated individual corresponding to any individual.

10. A multi-parameter fusion fault feature extraction and early warning system for reciprocating compressors, characterized in that, include: The acquisition unit is used to acquire vibration signals corresponding to different measuring points on the reciprocating compressor; The multi-scale feature extraction unit is used to splice multiple vibration signals to obtain a spliced ​​signal, and to extract multi-scale temporal features from the spliced ​​signal to obtain a multi-scale temporal depth feature vector of the reciprocating compressor. The physical prior feature extraction unit is used to extract time-frequency features from each vibration signal to obtain the time-frequency features of each vibration signal. Then, using a feature mapping network, nonlinear mapping is performed on each time-frequency feature to obtain the physical mapping features of each vibration signal. Based on the physical mapping features, the physical prior feature vector of the reciprocating compressor is generated. The reconstruction unit is used to map any vibration signal from multiple vibration signals to a discrete state space at different scales to obtain discrete state sequences at different scales, and to reconstruct the state vectors of each discrete state sequence to obtain reconstruction vectors at different scales. The composite feature extraction unit is used to extract the state arrangement and combination features of each reconstruction vector in the reconstruction space, so as to generate a composite feature vector corresponding to any vibration signal based on the state arrangement and combination features of each reconstruction vector in the reconstruction space, and after all vibration signals have been traversed, a composite correlation vector is formed by using each composite feature vector. In this case, any composite feature vector is used to characterize the local fluctuation degree and temporal evolution law of the corresponding vibration signal under multiple scales. The fault early warning unit is used to generate fault feature vectors by utilizing physical prior feature vectors, multi-scale time-series deep feature vectors, and composite correlation vectors, and inputs the fault feature vectors into the fault detection model to perform fault classification and early warning for reciprocating compressors based on the fault detection results.

Images