Centrifugal pump set fault identification method and equipment

Abstract

The embodiment of the invention discloses a centrifugal pump set fault identification method and equipment, and the method comprises the steps: carrying out the operation simulation of a centrifugal pump set comprising a plurality of parts under different working conditions, and enabling the different working conditions to comprise a multi-source domain corresponding to a marked known working condition and a target domain corresponding to an unmarked target working condition; obtaining multi-source domain operation data and target domain operation data; processing the multi-source domain operation data and the target domain operation data respectively to generate a multi-source domain feature time sequence and a target domain feature time sequence; the multi-source domain feature time sequence and the target domain feature time sequence are input into a multi-source domain adaptive network model for training, a centrifugal pump set fault recognition model is obtained, the centrifugal pump set fault recognition model achieves extraction of public global fault features from long sequences of a multi-source domain and a target domain, and global time sequence association between the features is established; and carrying out fault identification on the target domain through the centrifugal pump set fault identification model.

Description

Technical Field

[0001] The present specification relates to the technical field of equipment fault diagnosis, and in particular to a method and equipment for identifying faults of a centrifugal pump group. Background Art

[0002] As an important production and operation equipment widely used in many fields, the health status of centrifugal pump units is crucial to the efficiency and safety of industrial production processes. Common centrifugal pump units are composed of multiple parts, including centrifugal pumps, motors, couplings, inverters, etc. During operation, these components are often in harsh environments such as continuous operation, load changes, high temperature and high speed, and are easily affected by factors such as mechanical vibration, fatigue wear, and flow corrosion, resulting in damage and failure of key components. These failures will not only affect the production efficiency of the pump unit and increase the risk of failure, but may also endanger the entire production chain, and even cause major safety accidents and threaten personal safety.

[0003] Therefore, in the actual generation process, in order to avoid the above risks, it is necessary to perform more accurate and timely fault identification on the centrifugal pump group. Summary of the invention

[0004] One or more embodiments of the present specification provide a centrifugal pump group fault identification method and device, which are used to solve the technical problems raised by the background technology.

[0005] One or more embodiments of this specification adopt the following technical solutions:

[0006] One or more embodiments of this specification provide a centrifugal pump group fault identification method, the method comprising:

[0007] Performing operation simulation on a centrifugal pump group including multiple components under different working conditions, wherein the different working conditions include multiple source domains corresponding to marked known working conditions and target domains corresponding to unmarked target working conditions;

[0008] Collect signals of the operation simulation through corresponding sensors to obtain multi-source domain operation data and target domain operation data;

[0009] The multi-source domain operation data and the target domain operation data are processed respectively to generate a multi-source domain feature time series and a target domain feature time series;

[0010] The multi-source domain feature time series and the target domain feature time series are input into the multi-source domain adaptive network model for training to obtain a centrifugal pump group fault recognition model. The centrifugal pump group fault recognition model extracts common global fault features from long sequences of the multi-source domain and the target domain by fusing the common deep feature extractor of the Informer encoder, establishes a global temporal association between features, and uses SWD to align the overall marginal distribution and class-level conditional distribution of domain-specific features, thereby reducing the domain offset between different working conditions, and aligns the outputs of domain-specific classifiers to reduce the recognition differences of different classifiers on the same target working condition data;

[0011] Fault identification is performed on the target domain using the centrifugal pump group fault identification model.

[0012] It should be noted that the embodiments of this specification have the following beneficial effects through the above scheme:

[0013] Improve adaptability: Through operation simulation and signal collection under different working conditions, the model can learn the characteristics of multiple source domains and target domains, thereby improving adaptability to various working conditions.

[0014] Enhanced generalization capability: The common deep feature extractor integrated with the Informer encoder can extract common global fault features and establish global temporal associations between features, which helps the model to better generalize to unlabeled target domain data. It also uses SWD to align the overall marginal distribution and class-level conditional distribution of domain-specific features, respectively, to reduce the domain offset between different working conditions, and to align the outputs of domain-specific classifiers, to reduce the recognition differences of different classifiers on the same target working condition data.

[0015] Improve fault identification accuracy: Feature extraction and association of long sequences in multiple source and target domains can capture fault features more comprehensively, thereby improving the accuracy of fault identification.

[0016] Real-time and reliability: The model can identify faults in the target domain in real time, discover potential faults in a timely manner, and improve the reliability and stability of the system.

[0017] Reduce risk: Accurate and timely fault identification can help avoid failures or reduce their impact, reducing risks in the production process.

[0018] Ensure production safety: Ensuring the stable operation of the centrifugal pump group helps to ensure the safety of the production process and avoid major safety accidents.

[0019] One or more embodiments of this specification provide a centrifugal pump group fault identification device, including:

[0020] at least one processor; and,

[0021] a memory communicatively connected to the at least one processor; wherein,

[0022] The memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to:

[0023] Performing operation simulation on a centrifugal pump group including multiple components under different working conditions, wherein the different working conditions include multiple source domains corresponding to marked known working conditions and target domains corresponding to unmarked target working conditions;

[0024] Collect signals of the operation simulation through corresponding sensors to obtain multi-source domain operation data and target domain operation data;

[0025] The multi-source domain operation data and the target domain operation data are processed respectively to generate a multi-source domain feature time series and a target domain feature time series;

[0026] The multi-source domain feature time series and the target domain feature time series are input into the multi-source domain adaptive network model for training to obtain a centrifugal pump group fault recognition model. The centrifugal pump group fault recognition model extracts common global fault features from long sequences of the multi-source domain and the target domain by fusing the common deep feature extractor of the Informer encoder, establishes a global temporal association between features, and uses SWD to align the overall marginal distribution and class-level conditional distribution of domain-specific features, thereby reducing the domain offset between different working conditions, and aligns the outputs of domain-specific classifiers to reduce the recognition differences of different classifiers on the same target working condition data;

[0027] Fault identification is performed on the target domain using the centrifugal pump group fault identification model.

[0028] At least one of the above technical solutions adopted in the embodiments of this specification can achieve the following beneficial effects:

[0029] Improve adaptability: Through operation simulation and signal collection under different working conditions, the model can learn the characteristics of multiple source domains and target domains, thereby improving adaptability to various working conditions.

[0030] Enhanced generalization capability: The common deep feature extractor integrated with the Informer encoder can extract common global fault features and establish global temporal associations between features, which helps the model to better generalize to unlabeled target domain data. It also uses SWD to align the overall marginal distribution and class-level conditional distribution of domain-specific features, respectively, to reduce the domain offset between different working conditions, and to align the outputs of domain-specific classifiers, to reduce the recognition differences of different classifiers on the same target working condition data.

[0031] Improve fault identification accuracy: Feature extraction and association of long sequences in multiple source and target domains can capture fault features more comprehensively, thereby improving the accuracy of fault identification.

[0032] Real-time and reliability: The model can identify faults in the target domain in real time, discover potential faults in a timely manner, and improve the reliability and stability of the system.

[0033] Reduce risk: Accurate and timely fault identification can help avoid failures or reduce their impact, reducing risks in the production process.

[0034] Ensure production safety: Ensuring the stable operation of the centrifugal pump group helps to ensure the safety of the production process and avoid major safety accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art description. Obviously, the drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. In the drawings:

[0036] Figure 1 A schematic flow chart of a centrifugal pump group fault identification method provided for one or more embodiments of this specification;

[0037] Figure 2 An overall flow chart of a method based on global temporal association and multi-source domain adaptation provided in one or more embodiments of this specification;

[0038] Figure 3 A schematic diagram of a centrifugal pump assembly provided for one or more embodiments of this specification;

[0039] Figure 4 A schematic diagram of preprocessing operations provided for one or more embodiments of this specification;

[0040] Figure 5 A model architecture diagram provided for one or more embodiments of this specification;

[0041] Figure 6 A schematic diagram of the Informer encoder network structure provided for one or more embodiments of this specification;

[0042] Figure 7 A schematic diagram of a probabilistic sparse scaled dot product self-attention mechanism calculation structure provided for one or more embodiments of this specification;

[0043] Figure 8A schematic diagram of the overall training process of the centrifugal pump group fault identification model provided in one or more embodiments of this specification;

[0044] Fig. 9 A schematic diagram of the structure of a centrifugal pump group fault identification device provided in one or more embodiments of this specification. DETAILED DESCRIPTION

[0045] The embodiments of this specification provide a centrifugal pump group fault identification method and device.

[0046] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.

[0047] Figure 1 A schematic flow chart of a centrifugal pump group fault identification method provided in one or more embodiments of this specification, the process can be executed by a centrifugal pump group fault identification system. Certain input parameters or intermediate results in the process allow manual intervention and adjustment to help improve accuracy.

[0048] The method steps of the embodiment of this specification are as follows:

[0049] S102, performing operation simulation on a centrifugal pump group including multiple components under different working conditions, wherein the different working conditions include multiple source domains corresponding to marked known working conditions and target domains corresponding to unmarked target working conditions.

[0050] In the embodiments of this specification, the motor speed of the centrifugal pump group can be changed or the flow condition of the centrifugal pump can be changed by adjusting the frequency converter to simulate the operation of the centrifugal pump group fault containing multiple components under different conditions.

[0051] It should be noted that the above content can be implemented through the following specific implementation plans:

[0052] First, determine the different working conditions that need to be simulated, including the marked known working conditions (i.e., multi-source domain) and the unmarked target working conditions (i.e., target domain). Install and connect the centrifugal pump group, including the centrifugal pump, motor, coupling, inverter and other components. Set the parameters of the inverter and change the motor speed by adjusting the inverter. You can select different speed ranges and change modes to simulate different working conditions. At the same time, you can further simulate different operating conditions by adjusting the flow conditions of the centrifugal pump. You can use valves or other flow control devices to change the flow. During the operation simulation, use sensors to monitor the operating parameters of each component of the centrifugal pump group, such as vibration, temperature, pressure, etc. For known working conditions, you can mark the corresponding fault types and characteristics for subsequent fault identification and analysis. For unmarked target working conditions, you can record the operating data for subsequent fault diagnosis and analysis. During the simulation process, you can gradually change the working conditions and observe the response and changes of the centrifugal pump group to obtain more comprehensive operating data. Repeat the above steps and simulate different working conditions multiple times to increase the diversity and reliability of the data.

[0053] S104, collecting signals of the operation simulation through corresponding sensors to obtain multi-source domain operation data and target domain operation data.

[0054] S106, processing the multi-source domain operation data and the target domain operation data respectively to generate a multi-source domain feature time series and a target domain feature time series.

[0055] In the embodiments of this specification, regarding the above S106, the following specific implementation scheme can be adopted:

[0056] The multi-source domain operation data and the target domain operation data are respectively subjected to abnormal data point value identification and elimination, collection missing value filling and noise reduction processing to obtain preliminary processed data; the preliminary processed data is recorded as x raw1 , Among them, L T is the original signal length, Ch is the number of sensor channels; the multi-channel sensor data is segmented by sliding window, and the segmented segments contain data of multiple device rotation cycle lengths, and the total number of segmented segments is l m; Statistical features of each channel data of each segmented segment are extracted in the time domain, frequency domain and time-frequency domain respectively, wherein the extracted time domain features include the maximum value, average value, peak-to-peak value, distilled average value, variance, standard deviation, root mean square, crest factor, margin factor, pulse coefficient, shape coefficient, skewness and kurtosis; the extracted frequency domain features include spectral power, centroid frequency, mean square frequency, root mean square frequency and frequency standard deviation; the extracted time-frequency domain features include power spectrum entropy, singular spectrum entropy and wavelet packet energy entropy; the total number of extracted features of each channel data is recorded as ST. Each type of statistical features extracted from each channel data is sorted along the time series segmentation direction to obtain a dimension of The multi-domain feature long sequence is normalized by Z-score along the sequence direction; the preprocessed sequence and corresponding label under a single working condition parameter form a domain D, where x 1 is the long sequence of multi-channel and multi-domain features obtained after the above preprocessing operations, y 1 is the label value, L is the number of samples in the domain D, and the preprocessing sequences and labels corresponding to M working condition parameters are regarded as multiple source domains D S , Target domain D T There are K unlabeled samples u in memory k ,

[0057] S108, input the multi-source domain feature time series and the target domain feature time series into the multi-source domain adaptive network model for training, and obtain a centrifugal pump group fault recognition model. The centrifugal pump group fault recognition model extracts common global fault features from long sequences of the multi-source domain and the target domain by fusing the common deep feature extractor of the Informer encoder, establishes global temporal associations between features, and uses SWD to align the overall marginal distribution and class-level conditional distribution of domain-specific features, thereby reducing domain offsets between different working conditions, and aligns the outputs of domain-specific classifiers, thereby reducing the recognition differences of different classifiers on the same target working condition data.

[0058] S110, performing fault identification on the target domain by using the centrifugal pump group fault identification model.

[0059] In the embodiment of this specification, the overall network architecture of the multi-source domain adaptive network model may include a common feature extractor F 1 , a domain-specific feature distribution alignment module and a domain-specific classifier alignment module; wherein the domain-specific feature distribution alignment module includes multiple domain-specific feature extractors, and the domain-specific feature extractors are The domain-specific classifier alignment module includes a domain-specific recognition network corresponding to each domain-specific feature distribution alignment module. The domain-specific recognition network is The common feature extractor includes an Informer encoder. The common feature extractor and the domain-specific feature extractor map the multi-channel multi-domain feature sequences of the multi-source domains and the target domain to the same common deep feature space by sharing parameters. Each pair of common deep features of the source domain and the target domain generated is sent to the domain-specific feature distribution alignment module to determine the domain-invariant representations between the multi-source domains and the target domain. The overall distribution difference of the domain-specific features between the working conditions is measured based on the SWD. In the domain-specific classifier alignment module, each domain-specific recognition network performs supervised training on its own source domain samples. At the same time, the domain-specific classifier alignment module aligns the fault recognition probability outputs of the target domain in each classifier. In addition, the difference measurement of the class-level distribution of the domain-specific features is realized by combining the probability output of the target domain samples, the true labels of the source domain samples and the corresponding domain-specific features, so as to obtain the fault recognition result of the target domain.

[0060] Furthermore, the common feature extractor F 1 The Informer encoder is used as the backbone network. The Informer encoder includes a preset basic module. The Informer encoder adopts a probabilistic sparse scaled dot product self-attention mechanism in the multi-head attention layer, and a distillation layer network is added at the end of the preset basic module.

[0061] Furthermore, the scaled dot product self-attention mechanism formula is:

[0062]

[0063] Among them, Q, K, V are the query matrix, key matrix and value matrix of linear projection. The Q, K, V matrices of the first multi-head attention layer come from the linear projection of the model input after position encoding. W Q , W K , W V Q, K, and V correspond to the projection parameter matrices, respectively, and d is the characteristic dimension of the key matrix;

[0064] The probabilistic sparse scaled dot product self-attention mechanism is as follows:

[0065]

[0066] in, is the query matrix after filtering and replacement based on the maximum mean query vector sparsity measurement based on Kullback-Leibler divergence;

[0067] Set L Q , L K , L V The number of row vectors of Q, K, and V, respectively, where the number of random sampling is LQ LqCy K The query vector to key vector dot product pairs are used to replace all dot product pairs to measure the query vector sparsity. The number of selections after measurement and screening is specified as u, and the formula for u is:

[0068] u=c×lnL Q ;

[0069] Where c is the hyperparameter sampling factor, and c is a positive integer;

[0070] The maximum mean query vector sparsity measurement formula is:

[0071]

[0072] Among them, q i is the i-th row vector in Q, k s is the s-th row vector in K.

[0073] Furthermore, the common feature extractor F 1 The multi-channel and multi-domain feature long sequences of multiple source domains and target domains are used as input, and the input is mapped to the same common deep feature space through the parameter-sharing Informer encoder to perform adaptive fault feature recognition, and the extracted feature map is output after the flattening layer. For example,

[0074] Among them, the overall process of the common feature extractor is:

[0075] F 1 (x i )=Flatten(F informer_n (x i ));

[0076] In the formula, Flatten(·) is the flattening operation, F informer_n (·) is the calculation process of n Informer encoders;

[0077] The process of a single Informer encoder basic module is:

[0078]

[0079] Among them, Conv1d(.) is a one-dimensional convolution filter using the ELU activation function, MaxPool(.) is a distillation layer operation, [.] AB Including the multi-head probability scaling dot product self-attention mechanism and other necessary calculations, the formula of the multi-head probability scaling dot product self-attention mechanism is:

[0080] MH(Q,K,V)=Concat(Head 1 ,Head2 ,…,Head H )W O ;

[0081]

[0082] in, and W O is the projection parameter matrix, and H is the number of heads in the multi-head self-attention layer.

[0083] Furthermore, in the domain-specific feature distribution alignment module, each domain-specific feature extractor The parameters between them are not shared, and the public deep features of the target domain are Mapping obtains multiple domain-specific features. In order to mine the domain-invariant representation between known working conditions and target working conditions in a specific feature space and narrow the distribution differences in each specific feature space, the SWD distance is used to explicitly measure the overall marginal distribution and class-level conditional distribution of the features, and the mean of the measurements is used as the domain-specific overall distribution alignment loss. and domain-specific class-level distribution alignment loss Among them, the domain-specific overall distribution alignment loss The formula is:

[0084]

[0085] in, is the second-order SWD distance calculation formula approximated by the Monte Carlo method;

[0086] In the domain-specific feature space, SWD is used to measure the class centroid distance of each pair of source and target domain samples, and the class centroid distance is used as the domain-specific class-level distribution alignment loss. To minimize, The formula is:

[0087]

[0088] Among them, Rs i,c Indicates the source domain In by The fault type class centroid in the domain-specific feature space, Rt i,c is the target domain D in the corresponding domain-specific feature space T The centroid of the fault type class, Rs i,c The formula is:

[0089]

[0090] Among them, L c Source Domain The number of samples belonging to fault type c in Rt i,c The formula is:

[0091]

[0092] Among them, K c The target domain D T The number of samples judged as fault type c in Generate the corresponding category probability vector for the unlabeled target domain data u for the domain-specific recognition network The prediction confidence is quantified by entropy, and the formula is:

[0093]

[0094] Where H(·) is the cross entropy function;

[0095] The overall domain alignment loss function formula is:

[0096]

[0097] Among them, δ is a balance parameter that changes dynamically during the training process, and is minimized during the training process. The feature distributions between each pair of known working conditions and target working conditions are aligned, and each domain-specific fully connected network obtains the corresponding domain-invariant representation.

[0098] Furthermore, in the domain-specific classifier alignment module, the centrifugal pump group fault identification model is a multi-output network with non-shared parameters, which is composed of domain-specific fault classifiers, namely Each domain specific fault classifier is the Softmax classifier network, Receive the domain-specific domain-invariant representation of the corresponding source domain, identify the corresponding fault type, use the cross entropy loss function as the classification loss, optimize the network parameters, and the overall classification loss formula is:

[0099]

[0100] in, is the cross entropy function, Synchronously receive the domain-specific domain-invariant representations of the target domain samples, and predict the corresponding fault types. Differences are taken between the fault classification label probability outputs of the target domain samples on all classifiers to obtain the corresponding differences, and the absolute values ​​of the corresponding differences are used as the difference alignment loss.

[0101]

[0102] By minimizing Each domain-specific fault classifier outputs the fault type determination result for the same target domain sample, and the mean of each classifier output is calculated to identify the fault type of the target domain;

[0103] The overall loss function of the centrifugal pump group fault identification model based on multi-feature space adaptation is:

[0104]

[0105] Among them, β 1 , β 2 is the hyperparameter weight.

[0106] It should be noted that in the scenario of small sample variable working conditions, how to use global time series association and multi-source domain (unsupervised) adaptive methods to deal with the difficulties of multi-component fault diagnosis of centrifugal pump groups in new working conditions with only a small amount of unlabeled sensor data. Specifically, in the process of building a model that can realize fault diagnosis knowledge transfer, how to effectively establish global time series associations between multi-dimensional long sequence features to improve the model's adaptive feature extraction capabilities, and at the same time, in scenarios with large differences in multi-condition features, how to achieve edge distribution and conditional distribution alignment between multiple historical working conditions and target working conditions through effective feature measurement, thereby improving the accuracy of fault identification.

[0107] As an important production and operation equipment, centrifugal pump units are widely used in the fields of mineral mining, water supply and power generation, food and pharmaceutical, petrochemical, etc. Common centrifugal pump units are composed of centrifugal pumps, motors, couplings, frequency converters and other parts. In the operating state, the health status of key components in each part will directly affect the efficiency and safety of the industrial production process. Because during the operation process, the centrifugal pump unit is often in an environment of continuous operation, load changes, high temperature and high speed. The resulting mechanical vibration, fatigue wear, flow corrosion and other factors cause damage and failure of key components of various parts of the pump unit, such as various bearing failures in centrifugal pumps, impeller damage, various motor bearing failures, stator winding short circuit, coupling wear failure, equipment misalignment, etc. The various failures in these components may affect the production efficiency of the pump unit at the least, increase the risk of pump unit failure, or even endanger the entire production chain, resulting in major safety accidents such as personal safety. Therefore, accurate, timely and reliable fault diagnosis is one of the key means to avoid the above risks.

[0108] With the rapid development of the Internet of Things and time series data analysis technology, the existing data-driven, especially deep learning-based centrifugal pump group fault diagnosis methods have attracted widespread attention and achieved good results due to their adaptive feature extraction capabilities. However, most of the existing methods assume that the operating conditions of the centrifugal pump group are constant, and the conditions covered in the training and deployment stages are the same. Specifically, in the two stages of training and deployment, the equipment type, operating environment and operating parameters of the centrifugal pump group are the same, and the distribution of monitoring sensor data corresponding to these two stages is consistent. However, in actual production operation, the centrifugal pump group changes due to production rhythm or process requirements, and the equipment conditions (such as centrifugal pump flow, motor speed, etc.), and the distribution of monitored sensor data also changes accordingly, that is, domain shift occurs, resulting in a significant drop in the performance of deep learning methods that follow the independent and identically distributed data assumption when deployed in such scenarios. In addition, existing deep learning methods (such as LSTM, CNN, etc.) have limitations in the adaptive feature extraction of long time series, and cannot effectively establish long-distance time series associations between sequences.

[0109] In addition, as one of the important methods of transfer learning, unsupervised domain adaptation relaxes the independent and identically distributed restrictions in deep learning. It can use a small amount of unlabeled data under the working condition to be diagnosed and historical data under the known working condition to mine generalized features across working conditions and reduce the offset between different working conditions. However, most existing methods adopt the method of unsupervised domain adaptation in a single source domain, which easily ignores the transferable fault diagnosis knowledge between multiple historical working condition data. To address this problem, a small number of existing fault diagnosis methods adopt the method of unsupervised domain adaptation based on multi-source subdomains, and use the maximum mean discrepancy (MMD) to only align the marginal distribution between domains to reduce the domain offset and realize fault diagnosis under variable working conditions. In addition, there are large differences between multiple historical working conditions and target working conditions in actual operation, resulting in a large difference distribution in the feature space. There may be little or no overlap between feature distributions. In this case, it is difficult to obtain meaningful metrics using MMD.

[0110] After comprehensive analysis, the limitations of existing variable operating condition fault diagnosis methods are as follows:

[0111] 1) The existing variable operating condition fault diagnosis method only diagnoses the fault type of a single moving component (such as bearings, etc.), and does not perform comprehensive fault diagnosis for the faults of multiple moving components of a centrifugal pump group (such as bearings, motor stator winding short circuits, equipment misalignment, impeller damage, etc.).

[0112] 2) In the feature extraction process, the global temporal correlation of the monitoring data features of multiple measurement points is not considered;

[0113] 3) Existing variable operating condition fault diagnosis methods mostly adopt a single source domain unsupervised domain adaptation approach, without considering that the actual historical monitoring data may contain multiple operating conditions. If these data are only integrated into a single source domain for domain-invariant feature learning, the fault characteristics of the centrifugal pump group under different operating conditions are easily ignored, which in turn has a negative impact on the accuracy of fault identification.

[0114] 4) In the existing variable-condition fault diagnosis methods based on multi-source sub-domain adaptation, only the edge distribution alignment of inter-domain features is considered, and the fine-grained feature distribution alignment of fault categories between domains is ignored. At the same time, it does not consider how to effectively measure when the feature space distribution differences are large.

[0115] Biggest advantage: For the multi-component fault diagnosis of centrifugal pump groups under variable working conditions with small samples, a method based on global time series association and multi-source domain adaptation is proposed. This method can fully combine multiple historical working condition data for cross-working condition analysis under the premise of using only a small amount of unlabeled data in the target new working condition. The embodiment of this specification constructs a common feature extraction space through the Informer encoder, adaptively extracts common high-level fault features in long sequences, establishes global time series associations between features, and uses the Sliced ​​Wasserstein Distance (SWD) to align the domain-specific overall distribution and class-level distribution of high-level features of multiple historical working conditions and target new working conditions, thereby making the fault diagnosis results of the centrifugal pump group more accurate and the labeling cost lower.

[0116] The embodiments of this specification propose a method for multi-component fault diagnosis of a centrifugal pump group under variable operating conditions based on global time series association and multi-source domain adaptation. In order to solve the domain offset problem caused by differences in the distribution of sensor data of multiple historical operating conditions (called multi-source domains) and unlabeled data of target operating conditions (called target domains) under variable operating conditions, a centrifugal pump group fault identification model based on multi-feature space adaptation is constructed based on the idea of ​​multi-source sub-domain adaptation. Specifically, the model does not need to label the target domain data. By fusing the common deep feature extractor of the Informer encoder, it can extract common global fault features from long sequences of multi-source domains and target domains, and establish global time series associations between features. Next, the embodiments of this specification use SWD to align the overall marginal distribution and class-level conditional distribution of domain-specific features, respectively, to reduce the domain offset between different operating conditions, and to align the outputs of domain-specific classifiers, to reduce the recognition differences of different classifiers for the same target operating condition data, and to improve the accuracy and adaptability of the model's fault diagnosis results. The overall flow chart of the method based on global time series association and multi-source domain adaptation is shown below. Figure 2As shown in the figure, a multi-component fault simulation experiment under different working conditions is first carried out. In the experiment, multiple sensor measurement points are deployed to meet the multi-component monitoring needs of the centrifugal pump, and monitoring signals are collected using accelerometers, current clamps, etc. to obtain health and fault data of multiple historical working conditions. Then, the acquired data is preprocessed to generate a multi-source domain feature long time series. Then, the sequence is used as input for variable working condition model training, thereby achieving accurate fault identification under the target working condition.

[0117] by Figure 3 Taking the centrifugal pump group schematic diagram in as an example, in the fault simulation experiment, by adjusting the frequency converter to change the motor speed or change the centrifugal pump flow condition, multiple different centrifugal pump group operating parameters are obtained, and the number of operating parameters is recorded as M. Under each operating parameter, for typical fault types of centrifugal pump groups, such as motor bearing damage, motor stator winding short circuit, centrifugal pump bearing damage, centrifugal pump cavitation, impeller damage, loose anchor bolts, misalignment, etc., deployment is as follows Figure 3 The multiple vibration and current signal measurement points shown in the figure are used to carry out fault simulation experiments, and the multi-channel sensor signals are collected and annotated through the data acquisition module. Among them, the number of healthy and fault types of the centrifugal pump group is recorded as C.

[0118] After obtaining the above raw data samples, data preprocessing is carried out. First, the raw data is processed by identifying and eliminating abnormal data points, filling missing values, and reducing noise. Then, the data after preliminary processing is recorded as x raw1 , Where L T is the original signal length, Ch is the number of sensor channels, and the Figure 4 The preprocessing operation diagram is shown in the figure. First, the multi-channel sensor data is segmented by sliding window. The segmented segments contain data of multiple device rotation cycle lengths. The total number of segmented segments is l m . Then, statistical features are extracted in the time domain, frequency domain, and time-frequency domain for each channel data of each segmented segment, respectively. The extracted time domain features include but are not limited to the maximum value, average value, peak-to-peak value, distilled average value, variance, standard deviation, root mean square, crest factor, margin factor, pulse coefficient, shape coefficient, skewness and kurtosis, etc. The frequency domain features include but are not limited to spectral power, centroid frequency, mean square frequency, root mean square frequency, frequency standard deviation, etc. The time-frequency domain features include but are not limited to power spectrum entropy, singular spectrum entropy, wavelet packet energy entropy, etc. The total number of extracted features for each channel data is recorded as ST. Each type of statistical feature extracted from each channel data is sorted along the time series segmentation direction to obtain a dimension of The long sequence of multi-domain features. Since the extracted statistical feature values ​​have different ranges, in order to improve the convergence speed of network training and avoid gradient anomalies during training, the Z-score is used to normalize each feature along the sequence direction. Finally, the preprocessed sequence and the corresponding label under a single working condition parameter form a domain D, which is formally described as The above x l is the long sequence of multi-channel and multi-domain features obtained after the above preprocessing operations, y l is the label value, the number of label types is C, and L is the number of samples in domain D. The preprocessing sequences and labels corresponding to M working condition parameters are regarded as multiple source domains D S , formally described as Similarly, the data of the target working condition to be tested is also preprocessed according to the above description and regarded as the target domain D T .

[0119] After the data preprocessing operation is completed, multiple source domain data samples and target domain samples are sent to Figure 5 The model architecture diagram is shown in Figure 1. Figure 5 As shown in the figure, the overall network architecture consists of a common feature extractor F 1 , a domain-specific feature distribution alignment module (composed of multiple domain-specific feature extractors, denoted as and a domain-specific classifier alignment module (composed of each The corresponding domain-specific identification network is composed of Composition. The common feature extractor is mainly composed of the Informer encoder. In order to avoid the consumption of computing resources caused by repeated training of multiple networks, the feature extractor maps the multi-channel and multi-domain feature sequences of multiple source domains and target domains to the same common deep feature space through parameter sharing. Then, each pair of common deep features of the source domain and the target domain generated is sent to the domain-specific feature distribution alignment module to mine the domain-invariant representation between the historical working conditions and the target working conditions, and at the same time measure the overall distribution difference of the domain-specific features between the working conditions based on the SWD. In addition, in the domain-specific classifier alignment module, each domain-specific recognition network is supervised and trained for its own source domain samples. The module aligns the fault recognition probability outputs of the target domain unlabeled samples in each classifier. At the same time, the embodiment of this specification combines the probability output of the target domain samples, the true labels of the source domain samples and the corresponding domain-specific features to realize the difference measurement of the class-level distribution of the domain-specific features, so as to obtain a more reliable fault recognition result.

[0120] Common feature extractor F 1 Taking the Informer encoder as the backbone network, the Informer encoder network structure diagram is as follows Figure 6As shown in the figure, the whole is composed of basic modules. Compared with the Transformer encoder network architecture, in addition to the same residual and layer normalization, position-based feedforward layer operations, the Informer encoder also uses a probabilistic sparse scaled dot product self-attention mechanism in the multi-head attention layer, and adds a distillation layer network at the end of the basic module.

[0121] Among them, the schematic diagram of the calculation structure of the probability sparse scaling dot product self-attention mechanism is as follows Figure 7 As shown, the calculation formula of the scaled dot product self-attention mechanism is as follows:

[0122]

[0123] Where Q, K, and V are the query matrix, key matrix, and value matrix of the linear projection. It should be noted that the Q, K, and V matrices of the first multi-head attention layer come from the linear projection of the model input after position encoding. Q , W K , W V is the corresponding projection parameter matrix, and d is the characteristic dimension of the key matrix.

[0124] The probability sparse scaled dot product self-attention mechanism is calculated as follows:

[0125]

[0126] in, The query matrix is ​​filtered and replaced by the maximum-mean query vector sparsity measurement based on the Kullback-Leibler (KL) divergence. Q , L K , L V The number of row vectors of Q, K, and V, respectively, where the number of random sampling is L Q LqCy K The query vector-key vector dot product pairs are used to replace all dot product pairs to measure the sparsity of the query vector. The number of selections after measurement and screening is specified as u. The calculation formula for u is as follows:

[0127] u=c×lnL Q ;

[0128] Where c is the hyperparameter sampling factor, a positive integer.

[0129] The maximum-mean query vector sparsity measure is calculated as follows,

[0130]

[0131] In the formula, q i is the i-th row vector in Q, k s is the s-th row vector in K.

[0132] Compared with the scaled dot product self-attention mechanism, the probabilistic sparse scaled dot product self-attention mechanism randomly samples the key value matrix K and uses the KL divergence to measure the sparsity of the query matrix Q, thereby screening out the main fault features in the long-time feature sequence. The above operation avoids the long-tail sparsity problem of the scaled dot product self-attention mechanism when processing long-time sequences (that is, only a few fault features with larger weights are included, and the others can be ignored), and the computational complexity is reduced from O(L 2 ) is reduced to O(LlogL).

[0133] In addition, the Informer encoder introduces a distillation layer at the end of the basic module to further reduce redundant features and improve the common feature extractor F 1 Ability to handle long trains of centrifugal pumps.

[0134] like Figure 5 As shown, the common feature extractor F 1 The multi-channel and multi-domain feature long sequences of multiple source domains and target domains are used as input, and the input is mapped to the same common deep feature space through the parameter-sharing Informer encoder to perform adaptive main fault feature identification, and finally the extracted feature map is output through the flattening layer. Take the example of the common feature extractor as follows:

[0135] F 1 (x i )=Flatten(F informer_n (x i ));

[0136] In the formula, Flatten(·) is the flattening operation, F informer_n (·) is the calculation process of n Informer encoders, where the calculation process of a single Informer encoder basic module is as follows,

[0137]

[0138] Among them, Conv1d(.) is a one-dimensional convolution filter using the ELU activation function, MaxPool(.) is a distillation layer operation, [.] AB It includes the multi-head probability scaling dot product self-attention mechanism and other necessary calculations. The calculation formula of the multi-head probability scaling dot product self-attention mechanism is:

[0139] MH(Q,K,V)=Concat(Head 1 ,Head 2 ,…,Head H )W O ;

[0140]

[0141] In the formula, and W O is the projection parameter matrix, and H is the number of heads in the multi-head self-attention layer.

[0142] After the above calculations, the common feature extractor F 1 Fully explore the temporal dependencies related to centrifugal pump faults. After flattening the layer output, obtain the high-level fault-related representation of each operating condition parameter in the common deep feature space.

[0143] In order to reduce the difficulty of directly aligning the common deep feature distributions of multiple working conditions, in the domain-specific feature distribution alignment module, each pair of common deep features of the source domain and the target domain are fed into their respective domain-specific fully connected networks. Extract domain-specific features and perform overall distribution and class-level alignment between distributions of domain-specific features.

[0144] Compared with metric functions such as MMD and Jensen-Shannon (JS), Wasserstein metric can mine geometric relationships in feature space and give meaningful metrics when measuring differences in feature distributions with little or no overlap. At the same time, using Wasserstein metric can avoid gradient vanishing and reduce mode collapse during model training. Given an arbitrary dimension r∈[2,+∞), is the L2 norm of the r-dimensional unit sphere, set at The hidden feature probability measure in is and The corresponding probability density functions are F t1 and F t2 , then the calculation formula of the p-order standard Wassertein metric is as follows,

[0145]

[0146] In the formula, Denotes all possible marginal distributions F t1 and F t2 The set of all possible joint distributions π of . Compared with the standard Wasserstein metric, SWD has lower computational cost and better scalability, especially in high-dimensional statistical inference scenarios, such as when measuring the distribution difference of the hidden feature space. Therefore, in scenarios where there are large feature differences in multiple working conditions, SWD can be measured more effectively than the above metric functions. In the actual calculation process of SWD, the Monte Carlo method can be used to obtain Upper uniform sampling For approximate calculation, the calculation formula is as follows:

[0147]

[0148] In the formula, represents the Radon operator.

[0149] In the domain-specific feature distribution alignment module, each The parameters between them are not shared, and the public deep features of the target domain are Mapping obtains multiple domain-specific features. In order to mine the domain-invariant representations between known working conditions and target working conditions in specific feature spaces and narrow their distribution differences in each specific feature space, the SWD distance is used to explicitly measure the overall marginal distribution and class-level conditional distribution of the features, and the mean of the measurements is used as the domain-specific overall distribution alignment loss. and domain-specific class-level distribution alignment loss where the domain-specific overall distribution alignment loss The calculation formula is as follows,

[0150]

[0151] In the formula, is the second-order SWD distance calculation formula approximated by the Monte Carlo method.

[0152] In a feature space with domain-invariant representation, the class centroid distances of source and target domain sample features should be close. Based on this premise, in a domain-specific feature space, the embodiments of this specification use SWD to measure the class centroid distance of each pair of source and target domain samples, and use this distance as the domain-specific class-level distribution alignment loss. To minimize, The specific calculation method is as follows:

[0153]

[0154] In the formula, Rs i,c Indicates the source domain In by The fault type class centroid in the domain-specific feature space, Rt i,c is the target domain D in the corresponding domain-specific feature space T The centroid of the fault type class. Among them, Rs i,c The calculation formula is as follows,

[0155]

[0156] Where, L c Source Domain The number of samples belonging to fault type c in Rt i,c The calculation formula is as follows,

[0157]

[0158] In the formula, K c The target domain D T The number of samples judged as fault type c in Generate the corresponding category probability vector for the unlabeled target domain data u for the domain-specific recognition network The prediction confidence of is quantified by entropy as follows:

[0159]

[0160] Where H(·) is the cross entropy function.

[0161] Then the overall domain alignment loss function is δ is a balance parameter that changes dynamically during the training process. The feature distributions between each pair of known working conditions and target working conditions are aligned, and each domain-specific fully connected network obtains the corresponding domain-invariant representation.

[0162] In the domain-specific classifier alignment module, the fault identification network is a multi-output network with non-shared parameters, which is composed of domain-specific fault classifiers, namely Each domain specific fault classifier For the Softmax classifier network, on the one hand, Receive the domain-specific domain-invariant representation of the corresponding source domain, identify the corresponding fault type, use the cross entropy loss function as the classification loss, optimize the network parameters, and the overall classification loss calculation formula is:

[0163]

[0164] In the formula is the cross entropy function. On the other hand, The domain-specific domain-invariant representations of the target domain samples are received synchronously, and the corresponding fault types are predicted. The outputs of the domain-specific fault classifiers for the target domain samples may be inconsistent, especially when the target domain samples are close to the fault classification boundary, the outputs of the classifiers may be significantly different. Therefore, this section subtracts the fault classification label probability outputs of the target domain samples on all classifiers to obtain the corresponding difference, and uses the absolute value of the corresponding difference as the difference alignment loss.

[0165] By minimizing Each domain-specific fault classifier outputs similar fault type judgment results for the same target domain sample. Finally, the proposed method obtains the mean of the outputs of each classifier to identify the fault type of the target domain sample.

[0166] In general, the overall loss function of the centrifugal pump fault identification model based on multi-feature space adaptation is as follows:

[0167]

[0168] In the formula, β 1 , β 2 is the hyperparameter weight.

[0169] Combining the above process, the overall training process diagram of the centrifugal pump group fault identification model is as follows: Figure 8 As shown. First, multi-channel vibration and current sensor data are collected synchronously during the processing, and the high-frequency raw data is converted into a long sequence of multi-channel and multi-domain statistical features through data preprocessing. M sequence data of known working condition parameters are selected as source domain input, and a small amount of unlabeled sequence data of target working condition parameters are selected as target domain input. Through the setting of the above network parameters and hyperparameters, a multi-source domain adaptive network is constructed for training. During the training process, each input passes through the common feature extractor, the domain-specific feature distribution alignment module, and the domain-specific classifier alignment module in turn, and the domain-specific overall distribution alignment loss is calculated. Domain-specific class-level distribution alignment loss Classification Loss and output difference alignment loss By minimizing each loss function, the network model can learn domain-specific invariant representations under each known operating condition parameter and target parameter, while each fault classifier improves the ability to identify the fault type of the target parameter sequence data. After the final training is completed, the model can accurately identify the fault type of other sequence data within the target parameter.

[0170] Fig. 9 A schematic diagram of a centrifugal pump group fault identification device provided for one or more embodiments of this specification, characterized in that it includes:

[0171] at least one processor; and,

[0172] a memory communicatively connected to the at least one processor; wherein,

[0173] The memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to:

[0174] Performing operation simulation on a centrifugal pump group including multiple components under different working conditions, wherein the different working conditions include multiple source domains corresponding to marked known working conditions and target domains corresponding to unmarked target working conditions;

[0175] Collect signals of the operation simulation through corresponding sensors to obtain multi-source domain operation data and target domain operation data;

[0176] The multi-source domain operation data and the target domain operation data are processed respectively to generate a multi-source domain feature time series and a target domain feature time series;

[0177] The multi-source domain feature time series and the target domain feature time series are input into the multi-source domain adaptive network model for training to obtain a centrifugal pump group fault recognition model. The centrifugal pump group fault recognition model extracts common global fault features from long sequences of the multi-source domain and the target domain by fusing the common deep feature extractor of the Informer encoder, establishes global temporal associations between features, and uses SWD to align the overall marginal distribution and class-level conditional distribution of domain-specific features, thereby reducing domain offsets between different working conditions, and aligns domain-specific classifier outputs to reduce recognition differences between different classifiers for the same target working condition data;

[0178] Fault identification is performed on the target domain using the centrifugal pump group fault identification model.

[0179] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

Claims

1. A centrifugal pump group fault identification method, characterized in that: include: Performing operation simulation on a centrifugal pump group including multiple components under different working conditions, wherein the different working conditions include multiple source domains corresponding to marked known working conditions and target domains corresponding to unmarked target working conditions; Collect signals of the operation simulation through corresponding sensors to obtain multi-source domain operation data and target domain operation data; The multi-source domain operation data and the target domain operation data are processed respectively to generate a multi-source domain feature time series and a target domain feature time series; The multi-source domain feature time series and the target domain feature time series are input into the multi-source domain adaptive network model for training to obtain a centrifugal pump group fault recognition model. The centrifugal pump group fault recognition model extracts common global fault features from long sequences of the multi-source domain and the target domain by fusing the common deep feature extractor of the Informer encoder, establishes a global temporal association between features, and uses SWD to align the overall marginal distribution and class-level conditional distribution of domain-specific features, thereby reducing the domain offset between different working conditions, and aligns the outputs of domain-specific classifiers to reduce the recognition differences of different classifiers on the same target working condition data; Fault identification is performed on the target domain using the centrifugal pump group fault identification model.

2. The method according to claim 1, characterized in that: The operation simulation of the centrifugal pump group failure including multiple components under different working conditions includes: By adjusting the frequency converter to change the motor speed of the centrifugal pump group or change the flow condition of the centrifugal pump, the operation simulation of the centrifugal pump group containing multiple components under different working conditions can be carried out.

3. The method according to claim 1, characterized in that The processing of the multi-source domain operation data and the target domain operation data respectively to generate a multi-source domain feature time series and a target domain feature time series includes: Respectively performing abnormal data point value identification and elimination, collection missing value filling and noise reduction processing on the multi-source domain operation data and the target domain operation data to obtain preliminary processed data; The preliminary processed data is recorded as Among them, L T is the original signal length, Ch is the number of sensor channels; The multi-channel sensor data is segmented by sliding windows. The segmented segments contain data of multiple device rotation cycle lengths. The total number of segmented segments is l m ; Statistical features of each channel data of each segmented segment are extracted in the time domain, frequency domain and time-frequency domain. The extracted time domain features include the maximum value, average value, peak-to-peak value, distilled average value, variance, standard deviation, root mean square, crest factor, margin factor, pulse coefficient, shape coefficient, skewness and kurtosis. The extracted frequency domain features include spectral power, centroid frequency, mean square frequency, root mean square frequency and frequency standard deviation. The extracted time-frequency domain features include power spectrum entropy, singular spectrum entropy and wavelet packet energy entropy. The total number of extracted features of each channel data is recorded as ST. Each type of statistical features extracted from each channel data is sorted along the time series segmentation direction to obtain a dimension of The long sequence of multi-domain features is normalized along the sequence direction using Z-score; The preprocessed sequence and corresponding label under a single working condition parameter form a domain D, where: x l is the long sequence of multi-channel and multi-domain features obtained after the above preprocessing operations, y l is the label value, L is the number of samples in the domain D, and the preprocessing sequences and labels corresponding to M working condition parameters are regarded as multiple source domains D S , Target domain D T There are K unlabeled samples u in memory k , 4. The method according to claim 3, characterized in that The overall network architecture of the multi-source domain adaptive network model includes a common feature extractor F1, a domain-specific feature distribution alignment module and a domain-specific classifier alignment module; wherein, The domain-specific feature distribution alignment module includes multiple domain-specific feature extractors. The domain-specific classifier alignment module includes a domain-specific recognition network corresponding to each domain-specific feature distribution alignment module. The domain-specific recognition network is The common feature extractor includes an Informer encoder. The common feature extractor maps the multi-channel multi-domain feature sequences of the multi-source domains and the target domain to the same common deep feature space by means of parameter sharing. Each pair of common deep features of the source domain and the target domain generated is sent to a domain-specific feature distribution alignment module to determine the domain-invariant representations between the multi-source domains and the target domain. The overall distribution difference of the domain-specific features between the working conditions is measured based on the SWD. In the domain-specific classifier alignment module, each domain-specific recognition network performs supervised training on its own source domain samples. The domain-specific classifier alignment module aligns the fault recognition probability outputs of the target domain in each classifier, and combines the probability output of the target domain samples, the true labels of the source domain samples and the corresponding domain-specific features to achieve the difference measurement of the class-level distribution of the domain-specific features, thereby obtaining the fault recognition result of the target domain.

5. The method according to claim 4, characterized in that The common feature extractor F1 uses the Informer encoder as the backbone network. The Informer encoder includes a preset basic module. The Informer encoder adopts a probabilistic sparse scaled dot product self-attention mechanism in the multi-head attention layer, and adds a distillation layer network at the end of the preset basic module.

6. The method according to claim 5, characterized in that The scaled dot product self-attention mechanism formula is: Among them, Q, K, V are the query matrix, key matrix and value matrix of linear projection. The Q, K, V matrices of the first multi-head attention layer come from the linear projection of the model input after position encoding. W Q , W K , W V Q, K, and V correspond to the projection parameter matrices, respectively, and d is the characteristic dimension of the key matrix; The probabilistic sparse scaled dot product self-attention mechanism is as follows: in, is the query matrix after filtering and replacement based on the maximum mean query vector sparsity measurement based on Kullback-Leibler divergence; Set L Q , L K , L V The number of row vectors of Q, K, and V, respectively, where the number of random sampling is L Q LqCy K The query vector to key vector dot product pairs are used to replace all dot product pairs to measure the sparsity of the query vector. The number of selections after measurement and screening is specified as u, and the formula for u is: u=c×lnL Q ; Where c is the hyperparameter sampling factor, and c is a positive integer; The maximum mean query vector sparsity measurement formula is: Among them, q i is the i-th row vector in Q, k s is the s-th row vector in K.

7. The method according to claim 6, characterized in that The common feature extractor F1 takes a long sequence of multi-channel and multi-domain features from multiple source domains and target domains as input, maps the input to the same common deep feature space through a parameter-sharing Informer encoder, performs adaptive fault feature recognition, and outputs the extracted feature map through a flattening layer; Among them, the overall process of the common feature extractor is: F1(x i )=Flatten(F informer_n (x i )); In the formula, Flatten(·) is the flattening operation, F informer_n (·) is the calculation process of n Informer encoders; The process of a single Informer encoder basic module is: Among them, Conv1d(.) is a one-dimensional convolution filter using the ELU activation function, MaxPool(.) is a distillation layer operation, [.] AB Including the multi-head probability scaling dot product self-attention mechanism and other necessary calculations, the formula of the multi-head probability scaling dot product self-attention mechanism is: MH(Q,K,V)=Concat(Head1,Head2,…,Head H )W O ; in, and W O is the projection parameter matrix, and H is the number of heads in the multi-head self-attention layer.

8. The method according to claim 7, characterized in that The method further comprises: In the domain-specific feature distribution alignment module, each domain-specific feature extractor The parameters between them are not shared, and the public deep features of the target domain are Mapping obtains multiple domain-specific features. In order to mine the domain-invariant representation between known working conditions and target working conditions in a specific feature space and narrow the distribution differences in each specific feature space, the SWD distance is used to explicitly measure the overall marginal distribution and class-level conditional distribution of the features, and the mean of the measurements is used as the domain-specific overall distribution alignment loss. and domain-specific class-level distribution alignment loss Among them, the domain-specific overall distribution alignment loss The formula is: in, is the second-order SWD distance calculation formula approximated by the Monte Carlo method; In the domain-specific feature space, SWD is used to measure the class centroid distance of each pair of source and target domain samples, and the class centroid distance is used as the domain-specific class-level distribution alignment loss. To minimize, The formula is: Among them, Rs i,c Indicates the source domain In by The fault type class centroid in the domain-specific feature space, Rt i,c is the target domain D in the corresponding domain-specific feature space T The centroid of the fault type class, Rs i,c The formula is: Among them, L c Source Domain The number of samples belonging to fault type c in Rt i,c The formula is: Among them, K c The target domain D T The number of samples judged as fault type c in Generate the corresponding category probability vector for the unlabeled target domain data u for the domain-specific recognition network The prediction confidence is quantified by entropy, and the formula is: Where H(·) is the cross entropy function; The overall domain alignment loss function formula is: Among them, δ is a balance parameter that changes dynamically during the training process, and is minimized during the training process. The feature distributions between each pair of known working conditions and target working conditions are aligned, and each domain-specific fully connected network obtains the corresponding domain-invariant representation.

9. The method according to claim 8, characterized in that The method further comprises: In the domain-specific classifier alignment module, the centrifugal pump group fault identification model is a multi-output network with non-shared parameters, which is composed of domain-specific fault classifiers, namely Each domain specific fault classifier is the Softmax classifier network, Receive the domain-specific domain-invariant representation of the corresponding source domain, identify the corresponding fault type, use the cross entropy loss function as the classification loss, optimize the network parameters, and the overall classification loss formula is: in, is the cross entropy function, Synchronously receive the domain-specific domain-invariant representations of the target domain samples, and predict the corresponding fault types. Differences are taken between the fault classification label probability outputs of the target domain samples on all classifiers to obtain the corresponding differences, and the absolute values ​​of the corresponding differences are used as the difference alignment loss. By minimizing Each domain-specific fault classifier outputs the fault type determination result for the same target domain sample, and the mean of each classifier output is calculated to identify the fault type of the target domain; The overall loss function of the centrifugal pump group fault identification model based on multi-feature space adaptation is: Among them, β1 and β2 are hyperparameter weights.

10. A centrifugal pump group fault identification device, characterized in that: include: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to: Performing operation simulation on a centrifugal pump group including multiple components under different working conditions, wherein the different working conditions include multiple source domains corresponding to marked known working conditions and target domains corresponding to unmarked target working conditions; Collect signals of the operation simulation through corresponding sensors to obtain multi-source domain operation data and target domain operation data; The multi-source domain operation data and the target domain operation data are processed respectively to generate a multi-source domain feature time series and a target domain feature time series; The multi-source domain feature time series and the target domain feature time series are input into the multi-source domain adaptive network model for training to obtain a centrifugal pump group fault recognition model. The centrifugal pump group fault recognition model extracts common global fault features from long sequences of the multi-source domain and the target domain by fusing the common deep feature extractor of the Informer encoder, establishes a global temporal association between features, and uses SWD to align the overall marginal distribution and class-level conditional distribution of domain-specific features, thereby reducing the domain offset between different working conditions, and aligns the outputs of domain-specific classifiers to reduce the recognition differences of different classifiers on the same target working condition data; Fault identification is performed on the target domain using the centrifugal pump group fault identification model.

Images