Intelligent analysis and operation and maintenance method for full life cycle of centrifugal machine

Abstract

The invention discloses a centrifuge full life cycle intelligent analysis and operation and maintenance method. The method comprises the following steps: constructing a health scoring model which is based on an SAE-SOM neural network and comprises wavelet denoising and future parameter prediction functions; constructing an equipment life prediction model; a data set obtained from detection equipment is subjected to problem sample screening and normalization operation, and then a feedforward neural network is trained. And the hyper-parameter of the feedforward neural network is determined through a genetic algorithm. And a fault tracing model is constructed. The system is composed of a feature extraction network, a middle layer, a relation measurement network and a fault classification network. And establishing an equipment operation environment cross validation module. According to the invention, through mutual cooperation of the equipment health assessment module, the whole machine life estimation module, the fault alarm module, the early warning and traceability module, the operation environment cross validation module and the warning module, fusion analysis is carried out on key operation parameters of the centrifugal machine and external environment data; the method can be widely applied to water plants and other complex industrial scenes with high requirements for the operation reliability of key equipment.

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis technology for the operating status of industrial equipment. Specifically, it relates to a health assessment method and system for water plant centrifuges based on intelligent model evaluation. The system covers functional modules such as health assessment, overall lifespan prediction, fault alarm, early warning and tracing, predictive maintenance, and cross-validation of equipment operating environment. It falls within the scope of intelligent operation and maintenance management system for the entire life cycle of centrifuges. Background Technology

[0002] Centrifuges, as core equipment in sludge dewatering processes, are widely used in wastewater treatment plants to achieve solid-liquid separation, offering significant advantages such as high efficiency and small footprint. Their primary function is to separate water from solid matter through high-speed rotation, making them the most commonly used equipment for sludge dewatering in water plants. However, in actual operation, centrifuges often operate under harsh conditions such as high temperature, high load, and strong vibration. Furthermore, their operating status is affected by multiple factors including seasonal changes, load fluctuations, and environmental disturbances. These factors significantly increase the risk of equipment failure and maintenance pressure, posing a significant challenge to the long-term stable operation of the equipment.

[0003] Although a certain research foundation has been accumulated in the field of health status assessment and fault monitoring of industrial equipment, the existing technology still has obvious shortcomings in the specific application scenario of centrifuges and is difficult to meet actual needs.

[0004] First, most data-driven studies assume that the sample data is in good condition and abundant. However, in reality, water plant centrifuges often operate in complex and harsh environments, and the plant buildings where the equipment is located are subject to significant noise interference, resulting in poor data quality. Vibration signals, especially those closely related to health assessments, are affected not only by the inherent vibrations of the equipment itself but also by external interferences such as plant structure vibrations and fluid pulsations in pipelines. These external noises severely mask the inherent characteristics of the centrifuge vibration signals, making it difficult for existing models to achieve the expected results. Furthermore, existing sample data generally suffers from imbalance, with normal operation data far exceeding fault data, particularly with a severe shortage of samples for low-frequency fault types. This leads to low accuracy in anomaly identification, affecting the model's stability and robustness in anomaly detection tasks. Currently, many water plants still rely on alarm thresholds set manually based on experience, failing to provide early warnings based on trends in equipment condition changes. This results in delayed response to equipment failures, failing to meet the needs of predictive maintenance.

[0005] Furthermore, most fault diagnosis modules in existing industrial equipment health monitoring systems focus on identifying whether the equipment is in an abnormal state or simply classifying it into a certain fault type. However, in multivariable and complex equipment such as centrifuges, different fault types often have overlapping features, intersecting evolution paths, and complex influence chains, making it difficult for traditional classification models (such as rule-based models or shallow neural networks) to accurately determine the cause and direction of fault tracing. At the same time, due to the scarcity and incomplete labeling of fault samples in historical data, traditional supervised learning models struggle to cover all fault modes, especially showing weak identification capabilities for first-time occurrences or complex faults.

[0006] In traditional industrial equipment operation and maintenance systems, equipment repair and replacement plans are typically based on fixed cycles or rely on the experience and judgment of maintenance personnel. This results in significant delays and a lack of specificity, making it difficult to dynamically respond to changes in the equipment's condition. While some monitoring systems can identify and alert on the current operating status of equipment, they generally lack the ability to model and predict the remaining useful life (RUL). Traditional methods typically cannot effectively model and dynamically assess the overall service life of the machine based on historical data such as equipment operating trends and health score evolution trajectories, leading to a lack of scientific rigor and data support in maintenance plans. This often results in either over-maintenance or under-maintenance during equipment operation, increasing maintenance costs and threatening system stability and security, thus hindering the practical implementation of predictive maintenance concepts.

[0007] Furthermore, the operating status of equipment can also affect its surrounding environment to some extent. For example, changes in equipment load, heat release, or mechanical vibration may cause fluctuations in environmental parameters such as local temperature, humidity, and particulate matter concentration. However, existing monitoring systems generally focus on the operating parameters of the equipment itself (such as vibration, temperature, and current), lacking a mechanism for sensing and synchronously analyzing the environmental conditions in which the equipment is located. Under complex operating conditions, there may be potential correlations between the operating status of equipment and environmental factors. For example, equipment overheating may lead to an increase in ambient air temperature, and some precursors to faults may also manifest as abnormal environmental fluctuations. Without environmental monitoring methods, the system cannot cross-validate the current state of the equipment through environmental signals or provide multi-dimensional support for future trends, thus limiting the accuracy and interpretability of health assessment results. Summary of the Invention

[0008] To address the technical problems existing in current equipment health assessment methods, such as sample imbalance, excessive sample noise, low fault diagnosis accuracy, weak fault tracing capabilities, imperfect alarm and early warning mechanisms, lack of overall machine life prediction capabilities, and lack of environmental condition cross-validation mechanisms, this invention proposes a centrifuge-specific equipment health assessment model, overall machine life prediction method, fault alarm, early warning and tracing mechanism, and equipment operating environment cross-validation method. A complete, multi-source data-driven intelligent analysis and operation and maintenance algorithm system for the entire lifecycle of centrifuges is constructed, including:

[0009] Firstly, this application constructs a health scoring model based on the SAE-SOM neural network, which includes wavelet denoising and future parameter prediction functions.

[0010] First, an initial dataset is acquired from the monitoring equipment. This dataset includes, but is not limited to, variables such as the bearing front end temperature, bearing rear end temperature, hydraulic oil temperature, vibration, and speed difference of the centrifuge. After filtering key variables and handling outliers in this initial dataset, wavelet decomposition denoising is used to optimize the data.

[0011] To better illustrate the future operation of the equipment, this application uses a TCN-GRU-based prediction model to predict the values ​​of several key variables one hour later. This prediction model employs TCN to extract features from multi-feature time-series data and automatically learns deep features from the time series through multi-layer convolutional operations. The feature fusion layer passes the features extracted by TCN to the GRU, which retains key information relevant to the prediction through a gating mechanism, thus outputting a predicted dataset of relevant parameters.

[0012] To address the challenge of evaluating equipment when fault samples are scarce or when there is no fault data, this application proposes an unsupervised SAE-SOM joint model. It leverages the feature extraction capabilities of SAE to compress the signal into a low-dimensional space, and utilizes the quantization error calculation properties of SOM to evaluate the mean quantization error (MQE) between the input sample and normal samples. During training, normal samples are used. First, the SAE structure is trained layer by layer until the model reconstruction error stabilizes and no longer decreases. Then, the SOM model is trained using the latent space features encoded by the SAE until convergence. At this point, the SAE-SOM model has learned the features of normal data.

[0013] In application, the data after wavelet denoising and the prediction dataset of TCN-GRU are used as the test dataset and input into the SAE model. The latent spatial features are output to the SOM model. The SOM calculates the MQE of the input features. Finally, the MQE is quantized to 0-100, which is the health assessment score. The health score is obtained by comparing the MQE between the test data and the normal data, thus obtaining the current health score and the future health score of the device, and avoiding the problem of imbalanced fault samples.

[0014] By comparing the current health score with the future health score, the device's status can be monitored in real time. When the current health score and the future health score fall below a certain value, alarms and warnings are issued respectively.

[0015] In the above implementation, after wavelet denoising, an SAE-SOM evaluation model is constructed to address the sample imbalance problem. A health baseline is built solely based on normal operating data. Stacked autoencoders extract data features, and a self-organizing neural network learns the normal data model. During evaluation, the maximum quantization error between the evaluated sample and the normal data is calculated and quantized as a health score, achieving unsupervised health assessment of the bearing. Simultaneously, a TCN is used to extract features from the current data, which are further processed by a GRU to output a dataset predicting various equipment parameters. This predicted dataset is then used as input to the SAE-SOM unsupervised health assessment model to obtain a future health score. Alarms and warnings are issued when the score falls below a certain value.

[0016] Secondly, this application also establishes a device life prediction model.

[0017] The dataset obtained from the detection equipment, after filtering and normalizing problematic samples, selects a certain number of samples with the best centrifuge conditions and a certain number of samples with the worst conditions, adds a remaining lifespan label, and trains a feedforward neural network (FNN). A genetic algorithm is then used to determine the hyperparameters of the feedforward neural network.

[0018] In the above implementation, when estimating equipment lifespan, sensor data obtained from the equipment to be predicted is used as input data and processed in the lifespan prediction model to obtain the current equipment lifespan prediction result. This truly achieves predictive maintenance of equipment, avoiding the phenomena of "over-maintenance" or "under-maintenance."

[0019] Thirdly, this application also constructs a fault tracing model.

[0020] First, the fault dataset is split into test samples and matching samples, which serve as the training set for the model.

[0021] The acquired equipment operation data is input into the trained model, and the fault tracing results are output after passing through the CNN-SNN model.

[0022] The model consists of four parts: a feature extraction network, an intermediate layer, a relation measurement network, and a fault classification network.

[0023] The feature extraction network mainly consists of three one-dimensional convolutional layers. By using different kernel sizes, it continuously compresses the number of features, achieving feature extraction from shallow to deep. The features output by the feature extraction network enter the intermediate layers for further transformation and integration.

[0024] The intermediate layer adjusts the feature dimensions through convolution operations and reduces overfitting by using regularization, providing a more suitable feature representation for subsequent branch tasks (similarity calculation and classification).

[0025] After the output features are fed into the similarity measurement network, the convolutional layers capture the feature associations between samples and output the final similarity score, which quantifies the degree of similarity between the test sample and the paired sample.

[0026] The intermediate layer features are synchronously input into the fault classification network to further extract information related to the fault category. Finally, through the activation function, they are transformed into the probability distribution of each category to determine the fault category of the input sample.

[0027] In the above implementation, when diagnosing and tracing equipment faults, relevant data obtained from the equipment is used as input to the CNN-SNN model. After passing through four parts—feature extraction network, intermediate layer, relation metric network, and fault classification—the model outputs the probability distribution of different fault types. This solves the problem that traditional equipment evaluation models cannot determine the specific fault type of the equipment.

[0028] Fourthly, this application also establishes a module for cross-validation of the device operating environment.

[0029] Obtain relevant environmental indicators around the device, including but not limited to , , Temperature, humidity, etc.

[0030] After further filtering key variables and normalizing the initial dataset, an experimental dataset related to the indicator is established.

[0031] The experimental data were input into the GRU time series prediction model after three levels of pruning for training.

[0032] When performing the air quality index prediction task, the environmental monitoring dataset is input into the trained GRU model for processing to obtain the air quality index prediction results corresponding to the target environment.

[0033] In the above implementation method, when monitoring and predicting air quality indicators around the equipment, it is necessary to use the sensor data obtained from the equipment to be predicted as input data and substitute it into the model to achieve real-time monitoring and prediction of air quality indicators around the equipment. This solves the problem of lack of cross-validation of the equipment operating environment in traditional equipment health assessment methods.

[0034] In summary, this invention, through the collaborative efforts of modules for equipment health assessment, overall lifespan prediction, fault alarm, early warning and tracing, and cross-validation and alerts of the operating environment, integrates and analyzes key operating parameters and external environmental data of the centrifuge. This results in a comprehensive intelligent analysis and maintenance algorithm system for the entire lifecycle of a centrifuge, possessing capabilities for health assessment and lifespan prediction, fault alarm, early warning and tracing mechanisms, predictive maintenance, and cross-validation of the operating environment. This algorithm system can continuously learn and dynamically adapt under multiple operating conditions, significantly improving the accuracy of equipment status identification and the scientific nature of maintenance decisions. It can be widely applied in complex industrial scenarios such as water plants where the reliability of critical equipment is paramount. Attached Figure Description

[0035] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart of the centrifuge intelligent analysis algorithm system.

[0037] Figure 2 This is a flowchart of the noise reduction process for the equipment's characteristic operating indicators.

[0038] Figure 3 This is a diagram of the TCN-GRU feature variable prediction model.

[0039] Figure 4 This is a diagram of the SAE-SOM equipment health scoring model.

[0040] Figure 5 This is a diagram of the GA-FNN equipment lifespan prediction model.

[0041] Figure 6 This is a flowchart for CNN-SNN device fault diagnosis.

[0042] Figure 7 This is a diagram of the cross-validation model for the MGHP-GRU operating environment. Detailed Implementation

[0043] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of the embodiments of this application.

[0044] Firstly, this invention constructs a device health assessment model that can be deployed and run on a server terminal with computing capabilities. The server may include various types of electronic devices such as personal computers (PCs), tablets, smartphones, and personal digital assistants (PDAs). The model denoises the acquired parameters using wavelet denoising and constructs a hybrid model (TCN-GRU) based on a temporal convolutional network (TCN) and a gated recurrent unit (GRU) to achieve accurate prediction of device parameters. Finally, the model is processed by the SAE-SOM health assessment model to obtain the device's current health score and future health score.

[0045] Step 1: Obtain the initial dataset collected by the monitoring equipment. The initial dataset includes, but is not limited to, the bearing front end temperature, bearing rear end temperature, hydraulic oil temperature, vibration, and speed difference of the centrifuge.

[0046] Optionally, sensors can be pre-placed around the equipment requiring maintenance. These sensors collect data on multiple relevant variables of the equipment and can directly transmit the data to a server. The server then obtains the different relevant variables of the monitored equipment and constructs an initial dataset. Alternatively, the server can obtain data from the factory's database as the initial dataset. The initial dataset is then filtered to select key variables as the health assessment dataset.

[0047] Step 2: To improve data quality, perform data cleaning on the acquired health assessment dataset, including removing outliers.

[0048] Step 3: Denoise the dataset using the discrete wavelet denoising method, the process is as follows: Figure 2 As shown.

[0049] Formula (1) is the transform formula for discrete wavelets. The scale of the transform, i.e., the resolution control parameter in wavelet transform. The displacement position is used to control the translation of the wavelet basis. The original signal, Discrete wavelet basis functions are the core of wavelet transform, constructing sub-signals of different frequencies by shifting k and scaling j. Through this transform, the signal can be decomposed into wavelet coefficients at different scales and locations, representing the characteristics of the signal at each scale.

[0050]

[0051] The Daubechies4 wavelet was selected as the wavelet basis function to denoise the centrifuge data, and the soft threshold coefficient was set to 0.04. The maximum number of decomposition layers was calculated to be 10. Soft threshold denoising was applied to the decomposed signal layer by layer. In application, the number of wavelet decomposition layers was first determined. Assuming the length of the decomposed signal is L, the number of sub-signals N was calculated by formula (2). Next, the threshold of the detail signal under each decomposition layer was determined, and the detail signal was threshold quantized. The threshold-processed detail signal and the approximation signal were used to reconstruct the denoised signal.

[0052] ;

[0053] ;

[0054] ;

[0055] The soft threshold denoising calculation is shown in formula (3). This represents the wavelet coefficients after soft thresholding, where w represents the wavelet coefficient, and its absolute value is used in the calculation. , λ is the sign function of w, representing the positive or negative sign of w, and λ represents the threshold. When the absolute value of w is less than the threshold λ, the wavelet coefficients are set to 0; when the absolute value is greater than or equal to the threshold, the coefficients are subtracted from the threshold, thereby removing the noise part from the wavelet coefficients. The threshold λ is calculated by formula (4). The standard deviation of the noise is calculated by taking the standard deviation of the high-frequency coefficients after wavelet transform, where N is the signal length. It is the natural logarithm.

[0056] Step 4: Select a healthy sample set from the wavelet-denoised dataset and use it as the training samples for SAE-SOM. The model structure diagram is shown below. Figure 4 As shown.

[0057] An autoencoder (AE) is an unsupervised learning neural network consisting of an encoder and a decoder. The encoder comprises a hidden layer and an input layer, transforming the input vector through a nonlinear transformation. Extract latent features (n=1, 2, ..., N):

[0058] ;

[0059] ;

[0060] In formula (5) Here is the weight coefficient matrix of the encoder. This is the bias coefficient matrix of the encoder. This is the encoder activation function. In formula (6) This is the weight coefficient matrix of the decoder. This is the bias coefficient matrix of the decoder. is the activation function for the decoder.

[0061] A single Advanced Feature Extraction (SAE) is composed of multiple stacked Advanced Feature Extractions (AEs). After the first AE is trained, its features are input into the second AE; after the second AE is trained, its extracted features are used as input into the third AE. This process is repeated to train all AEs, thus achieving feature extraction from shallow to deep layers.

[0062] Self-Organizing Map (SOM) is an artificial neural network based on unsupervised learning. It achieves data visualization, clustering, and feature extraction by mapping high-dimensional input data to a low-dimensional space (usually a two-dimensional grid) while preserving the topological structure and similarity of the data.

[0063] In SOM, Euclidean distance is used to measure the similarity between input samples and neuron weight vectors. Its calculation formula is (7). For data sample vectors, For the first The weight vector of a neuron, where n is the feature dimension. The smaller the distance, the better the input sample matches the features of that neuron.

[0064] ;

[0065] After SOM training is completed, the Euclidean distance between the input sample and the BMU weight vector is used as the quantization error. Normal samples are usually mapped to high-density neuron regions, resulting in lower errors; abnormal samples may be mapped to sparse regions or far from all neuron weights, leading to significantly higher errors.

[0066] Step 5: The health dataset is processed by TCN-GRU to obtain a prediction dataset for different parameters of the operating equipment.

[0067] To address the non-stationarity and long-term dependency characteristics of centrifuge time-series data, this invention constructs a hybrid model (TCN-GRU) based on a temporal convolutional network (TCN) and a gated recurrent unit (GRU), the architecture of which is as follows: Figure 3As shown, this model achieves accurate prediction of equipment parameters through hierarchical feature extraction and temporal modeling.

[0068] The TCN-GRU combined model fully leverages the advantages of both TCN and GRU, effectively processing complex time series data and performing multi-feature prediction.

[0069] The feature extraction layer is responsible for extracting features from the input multi-feature time series data using the TCN. The TCN can automatically learn deep features from the time series data through multiple convolutional operations.

[0070] The feature fusion layer's role is to pass the features extracted by the TCN to the GRU. The GRU can further learn the temporal relationships between these features and perform feature fusion. Through the GRU's gating mechanism, noise information can be effectively filtered out while retaining key information relevant to the prediction.

[0071] Step 6: Use the acquired real-time operating parameters and predicted parameter dataset of the device as input to SAE-SOM. The SAE outputs latent spatial features to the SOM model. The SOM calculates the MQE of the input features, and finally quantizes the MQE to 0-100, which is the health assessment score. The final prediction model outputs the current health score and the predicted parameters for one hour later.

[0072] Step 7: When the current health score and the future health score are lower than a certain value, the relevant variables of the alarm and warning variables will be set to TRUE, and alarm warnings can be issued through external buzzers, indicator lights, etc.

[0073] For the second aspect, equipment lifespan estimation, please refer to [link / reference]. Figure 5 , Figure 5 The framework for estimating device lifespan constructed for this application.

[0074] Step 8: Use the initial dataset obtained in Step 1 as input to the GA-FNN lifetime prediction model.

[0075] Step 9: The GA-FNN algorithm will estimate the remaining lifespan of the target device based on the health score output by the health assessment model and the health assessment dataset in memory.

[0076] We implemented an automated hyperparameter optimization process for feedforward neural networks in Python by integrating the TensorFlow and DEAP libraries. We defined the network structure and data processing flow, and set a random seed to ensure repeatability. During the execution of the genetic algorithm, the fitness (i.e., model performance) of individuals is evaluated in each generation, and the population is continuously optimized through selection, crossover, and mutation operations. Finally, after reaching a predetermined number of iterations, the algorithm outputs the optimal hyperparameter combination on the training set and its corresponding fitness value, achieving automated tuning of neural network hyperparameters.

[0077] ;

[0078] This formula is the output expression of a three-layer feedforward neural network detection model, where the input is... This refers to the preprocessed input data, where g and f are activation functions; l is the number of output results; and the model's output Y is the final lifetime prediction result.

[0079] Thirdly, regarding fault tracing, this application designs a CNN-SNN fault diagnosis model. Figure 6 This refers to the network structure of the model.

[0080] Step 10: If the current health score is lower than the set threshold, it indicates that the device is malfunctioning.

[0081] Step 11: Obtain the data from the device over the past 15 minutes as input to the model, perform fault tracing using the fault tracing algorithm (CNN-SNN), save the results, update the fault record, and optionally upload to the cloud or save locally.

[0082] The CNN-SNN model consists of four parts: a feature extraction network, intermediate layers, a relation measurement network, and a fault classification network. The model's input samples are the test sample and its matching samples. During training, positive and negative samples are randomly matched, and similarity labels are generated: 1 indicates input samples of the same class, and 0 indicates input samples of different classes. The feature extraction network mainly consists of three one-dimensional convolutional layers. By varying the kernel size, the number of features is continuously compressed, achieving feature extraction from shallow to deep. The features output by the feature extraction network enter the intermediate layers for further transformation and integration. The intermediate layers adjust the feature dimensions through convolutional operations and use regularization to reduce overfitting, providing more suitable feature representations for subsequent branch tasks (similarity calculation and classification).

[0083] After the output features are fed into the similarity measurement network, convolutional layers capture the feature relationships between samples. The activation function maps the output to the interval [0, 1], outputting the final similarity score, which quantifies the similarity between the test sample and the paired sample. The similarity calculation is shown in Equation 9. Where... This represents the similarity between sample i and sample j. For feature mapping function, The output of the feature extraction network

[0084] ;

[0085] The intermediate layer features are synchronously input into the fault classification network to further extract information related to the fault category. Finally, through the activation function, they are transformed into the probability distribution of each category to determine the fault category of the input sample.

[0086] During training, sample pairs are constructed. and Similarity calculation and classification are performed, and the loss values ​​are accumulated. During the model testing phase, only one sample to be tested is input. After feature extraction, it is input into the classification network and no longer goes through the similarity branch. The mean squared error (MSE) is used as the loss function of the classification network, and its calculation is shown in formula (10).

[0087] ;

[0088] In the formula Represents classification loss, Indicates the true label of the sample. For predicted labels.

[0089] The feature extraction network is responsible for extracting feature information from the input samples; the relation metric network optimizes the training process through similarity information, promoting the clustering of features for similar samples and the dispersion of features for dissimilar samples. The classification network directly performs the final fault classification operation. In this architecture, backpropagation simultaneously optimizes both the metric network and the classification network. The model maintains a shallow structure to avoid the gradient vanishing problem in deeper layers and reduce overfitting.

[0090] Step 12: Use the acquired data as input to the fault tracing model, and save or upload the model's output, i.e., the fault tracing result, to the cloud.

[0091] Fourthly, in order to better synchronize and predict the state of the surrounding environment of the device, additional cross-validation is provided for the current state of the device.

[0092] Step 13: During the operation of the above equipment health assessment model, sensors collect data from multiple air quality-related variables of the equipment. The sensors can directly transmit the data to the server, allowing the server to obtain different relevant variables of the surrounding environment of the monitored equipment and construct an initial dataset. Alternatively, the server can also obtain historical data from the factory's database as the initial dataset.

[0093] Step 14: The server uses the MGHP-GRU algorithm to analyze the surrounding area of ​​the device using the acquired data. , Predicting temperature, humidity, etc., and saving or uploading the results to the cloud; the model structure is as follows: Figure 7 As shown.

[0094] The basic model structure consists of a GRU layer and a fully connected layer, where the number of neurons in the fully connected layer corresponds to the desired prediction time step. The basic model undergoes three levels of pruning: Group Lasso neuron pruning, gate pruning, and weight pruning, resulting in a smaller weight matrix with more zero values, enabling faster prediction.

[0095] Among them, Group Lasso neuron pruning is the coarsest pruning. Group Lasso is a structured regularization method. Its core idea is to treat each neuron of GRU as a group and drive the weight of the whole group to approach 0 through Lasso regularization, which facilitates the subsequent structured pruning of these low-importance neurons.

[0096] The formula for calculating the number of Divine Generals after pruning is as follows:

[0097] ;

[0098] —The number of neurons after neuron pruning;

[0099] NPR—Neuron pruning rate;

[0100] Group Lasso's gated pruning is largely the same as Group Lasso-based neuron pruning methods, differing only in pruning granularity and content. Since some neurons have overlapping expressive power for temporal features, redundant neurons can be removed through neuron pruning. Short-term fluctuations (e.g., hourly changes) may be controlled by the reset gate, while long-term trends (e.g., daily changes) may be controlled by the update gate. By filtering the importance of gates, non-dominant gated branches can be selectively closed. A gate mask matrix of the same size as the weight matrix is ​​constructed; locations requiring gated pruning have a weight of 0, while locations not requiring gated pruning have a weight of 1. The formula is as follows:

[0101] ;

[0102] Weight pruning, as an unstructured pruning method, complements the previous structured pruning. It locates and removes redundant weights by evaluating parameter-by-parameter significance (e.g., based on L1 norm or gradient magnitude), thereby optimizing model sparsity at a micro level. The weight pruning process is largely the same as gated pruning, with regularization and importance ranking during training, and the use of mask matrices during model reconstruction and fine-tuning. The only difference is that gated pruning uses a gated parameter set. The L2 norm is used for sorting, while weight pruning sorts the L1 norm of each weight.

[0103] GRU is a highly effective variant of LSTM networks. It has a simpler structure than LSTM and performs well, making it a very popular network choice. As a variant of LSTM, GRU can also solve the long dependency problem in RNN networks. LSTM introduces three gate functions: input gate, forget gate, and output gate to control the input, memory, and output values. The GRU model, however, only has update and reset gates. The main formulas include: It is a weight matrix.

[0104] Update Gate:

[0105] ;

[0106] The update gate determines whether the memory from the previous step should be passed to the current step. Among other things, It is the hidden state from the previous step. σ is the input for the current step, and σ is the Sigmoid activation function.

[0107] Reset Door:

[0108] ;

[0109] Candidate hidden state:

[0110] ;

[0111] The candidate hidden state is a new candidate value that takes into account the reset gate. Here, tanh is the hyperbolic tangent activation function.

[0112] Hidden status update:

[0113] ;

[0114] The final hidden state is obtained by combining the updated gate and the candidate hidden state.

[0115] These gating mechanisms allow GRU to mitigate the vanishing gradient problem while handling long-term dependencies. GRU's design makes it relatively easy to train on some sequence modeling tasks.

[0116] Step 15: Package the model's output, including equipment health score, overall lifespan estimate, fault tracing information, and relevant data from cross-validation of the operating environment, together with the real-time operating status parameters collected by the current equipment (e.g., vibration, water temperature, speed difference, etc.) and the real-time status parameters of the surrounding environment (e.g., temperature, humidity, etc.), and send it to the local display module for visualization via the communication interface. Simultaneously, it can be linked with a mobile app to transmit the above information to a remote terminal, enabling remote real-time monitoring and management of the equipment's operating status.

[0117] This invention proposes a health assessment model, a lifespan prediction method, a fault tracing mechanism, and a cross-validation system for centrifuges based on multi-source operational data. It achieves real-time equipment status assessment, early warning of abnormal trends, and cross-validation with the operating environment. Through quantitative scoring of operating status, dynamic prediction of remaining lifespan, and fault alarm, warning, and tracing, this algorithm system significantly improves the intelligence level of equipment operating status management and avoids unplanned downtime events.

Claims

1. A method for intelligent analysis and operation and maintenance of centrifuges throughout their entire lifecycle, characterized in that: A device health assessment model is constructed and deployed on a server terminal with computing capabilities. The server includes a personal computer, tablet computer, smartphone, or personal digital assistant. The acquired parameters are denoised using wavelet denoising. A hybrid model TCN-GRU based on temporal convolutional network (TCN) and gated recurrent unit (GRU) is constructed to achieve accurate prediction of device parameters. Finally, the current health score and future health score of the device are obtained through the SAE-SOM health assessment model.

2. The intelligent analysis and operation and maintenance method for the entire life cycle of a centrifuge according to claim 1, characterized in that: The method includes the following steps: Step 1: Obtain the initial dataset collected by the monitoring equipment; the initial dataset includes the bearing front end temperature, bearing rear end temperature, hydraulic oil temperature, vibration, and speed difference of the centrifuge. Step 2: To improve data quality, perform data cleaning on the acquired health assessment dataset and remove outliers; Step 3: Denoise the dataset using the discrete wavelet denoising method; Formula (1) is the transform formula for discrete wavelets. The scale of the transform, i.e., the resolution control parameter in wavelet transform. The displacement position is used to control the translation of the wavelet basis. The original signal, Discrete wavelet basis functions are the core of wavelet transform. They construct sub-signals of different frequencies by shifting k and scaling j. Through the transform, the signal is decomposed into wavelet coefficients at different scales and locations, representing the characteristics of the signal at each scale. ； The Daubechies4 wavelet was selected as the wavelet basis function to denoise the centrifuge data, and the soft threshold coefficient was set to 0.

04. The maximum number of decomposition layers was calculated to be 10. The soft threshold was used to denoise the decomposed signal layer by layer. When applying the decomposition, the number of wavelet decomposition layers was first determined. Assuming the length of the decomposed signal is L, the number of decomposed sub-signals N was calculated by formula (2). Secondly, the threshold of the detail signal under each decomposition layer was determined, and the detail signal was threshold quantized. The threshold-processed detail signal and the approximation signal were used to reconstruct the denoised signal. ； ； ； The soft threshold denoising calculation is shown in formula (3). This represents the wavelet coefficients after soft thresholding, where w represents the wavelet coefficient, and its absolute value is used in the calculation. , λ is the sign function of w, indicating the sign of w, and λ represents the threshold. When the absolute value of w is less than the threshold λ, the wavelet coefficients are set to 0; when the absolute value is greater than or equal to the threshold, the coefficients are subtracted from the threshold, thereby removing the noise part in the wavelet coefficients. The threshold λ is calculated using formula (4). The standard deviation of the noise is calculated by taking the standard deviation of the high-frequency coefficients after wavelet transform, where N is the signal length. It is the natural logarithm; Step 4: Select a healthy sample set from the wavelet-denoised dataset and use it as the training sample for SAE-SOM; An autoencoder (AE) is an unsupervised learning neural network consisting of an encoder and a decoder. The encoder comprises a hidden layer and an input layer, transforming the input vector through a nonlinear transformation. Extract latent features n=1, 2, ..., N: ； ； In formula (5) Here is the weight coefficient matrix of the encoder. This is the bias coefficient matrix of the encoder. For encoder activation function; in formula (6) This is the weight coefficient matrix of the decoder. This is the bias coefficient matrix of the decoder. is the activation function for the decoder; SAE is composed of multiple stacked AEs. After the first AE is trained, the features are input into the second AE. After the second AE is trained, the extracted features are used as the input of the third AE; and so on, training of all AEs is completed, thereby realizing feature extraction from shallow to deep layers. Self-organizing maps (SOMs) are artificial neural networks based on unsupervised learning. SOMs use Euclidean distance to measure the similarity between input samples and neuron weight vectors, calculated using formula (7). For data sample vectors, For the first The weight vector of each neuron, where n is the feature dimension; the smaller the distance, the better the input sample matches the features of that neuron. ； After the self-organizing map (SOM) is trained, the Euclidean distance between the input sample and the BMU weight vector is used as the quantization error. Step 5: Process the health dataset to obtain a prediction dataset for different parameters of the operating equipment using TCN-GRU; To address the non-stationarity and long-term dependency characteristics of centrifuge time-series data, a hybrid model TCN-GRU, based on Temporal Convolutional Network (TCN) and Gated Recurrent Unit (GRU), is constructed. Through hierarchical feature extraction and temporal modeling, accurate prediction of equipment parameters is achieved. The combined TCN-GRU model fully leverages the advantages of both TCN and GRU, effectively processing complex time-series data and performing multi-feature prediction. The feature extraction layer utilizes TCN to extract features from the input multi-feature time-series data. TCN automatically learns deep features in the time series through multi-layer convolutional operations. The function of the feature fusion layer is to pass the features extracted by TCN to GRU; Step 6: Use the acquired real-time operating parameters of the device and the predicted parameter dataset as input to SAE-SOM. The potential spatial features output by SAE are fed into the SOM model. The SOM model calculates the MQE of the input features. Finally, the MQE is quantized to 0-100, which is the health assessment score. The final prediction model outputs the current health score and the future prediction parameters after 1 hour. Step 7: When the current health score and the future health score fall below a certain value, the relevant variables of the alarm and warning variables will be set to TRUE, and an alarm warning will be issued through an external buzzer and indicator light. Step 8: Use the initial dataset obtained in Step 1 as input to the GA-FNN lifetime prediction model; Step 9: The GA-FNN algorithm will estimate the remaining lifespan of the target device based on the health score output by the health assessment model and the health assessment dataset in memory; Step 10: If the current health score is lower than the set threshold, it indicates that the device is malfunctioning; Step 11: Use the past data obtained from the device as input to the model, perform fault tracing through the CNN-SNN fault tracing algorithm, save the results, update the fault records, and choose to upload to the cloud or save locally. Step 12: Use the acquired data as input to the fault tracing model, and save or upload the output of the fault tracing model, i.e., the fault tracing result, to the cloud. Step 13: During the operation of the above equipment health assessment model, sensors collect data from multiple air quality-related variables of the equipment. The sensors directly transmit the data to the server so that the server can obtain different relevant variables of the surrounding environment of the tested equipment and form an initial dataset; the server obtains historical data from the factory's database as the initial dataset. Step 14: The server uses the MGHP-GRU algorithm to analyze the surrounding area of ​​the device using the acquired data. , Predict temperature and humidity, and save or upload the results to the cloud. Step 15: Package the model's output, including equipment health score, estimated overall lifespan, fault tracing information, and relevant data from cross-validation of the operating environment, together with the real-time operating status parameters collected by the current equipment and the real-time status parameters of the environment. Then, send the packaged data to the local display module for visualization via the communication interface. It can also be linked with a mobile app to transmit information synchronously to a remote terminal, enabling remote real-time monitoring and management of the equipment's operating status.

3. The intelligent analysis and operation and maintenance method for the entire life cycle of a centrifuge according to claim 1, characterized in that: Sensors are placed around the equipment that needs maintenance in advance. The sensors collect data from multiple relevant variables of the equipment. The sensors can directly transmit the data to the server so that the server can obtain different relevant variables of the equipment under test and form an initial dataset; or, the server can obtain data from the factory's database as the initial dataset. The initial dataset is then filtered to select key variables as the health assessment dataset.

4. The intelligent analysis and operation and maintenance method for the entire life cycle of a centrifuge according to claim 1, characterized in that: The hyperparameter optimization process of the feedforward neural network defines the network structure and data processing flow, and sets a random seed to ensure repeatability. During the execution of the genetic algorithm, the fitness of individuals is evaluated in each generation, and the population is continuously optimized through selection, crossover and mutation operations. Finally, after reaching the predetermined number of iterations, the hyperparameter combination that performs best on the training set and its corresponding fitness value are output, realizing the automated tuning of the neural network hyperparameters. ； Equation (8) is the output expression of the three-layer feedforward neural network detection model, where the input is directed to the network. This refers to the preprocessed input data, where g and f are activation functions; l is the number of output results; and the output Y of the three-layer feedforward neural network detection model is the final lifetime prediction result.

5. The intelligent analysis and operation and maintenance method for the entire life cycle of a centrifuge according to claim 1, characterized in that: The CNN-SNN model consists of four parts: a feature extraction network, intermediate layers, a relation measurement network, and a fault classification network. The model's input samples are the test sample and its matching samples. During training, positive and negative samples are randomly matched, and similarity labels are generated: 1 indicates input samples of the same class, and 0 indicates input samples of different classes. The feature extraction network consists of three layers of one-dimensional convolutional structures. By varying the kernel size, the number of features is continuously compressed, achieving feature extraction from shallow to deep. The features output by the feature extraction network enter the intermediate layers for transformation and integration. The intermediate layers adjust the feature dimensions through convolutional operations and use regularization to reduce overfitting, providing suitable feature representations for subsequent branch tasks. After the output features are fed into the similarity measurement network, convolutional layers capture the feature relationships between samples. The activation function maps the output to the interval [0, 1], outputting the final similarity score, which quantifies the similarity between the test sample and the paired sample. The similarity calculation is shown in Equation 9; where... This represents the similarity between sample i and sample j. For feature mapping function, The output of the feature extraction network; ； The intermediate layer features are synchronously input into the fault classification network to extract information related to the fault category. Through the activation function, they are transformed into the probability distribution of each category to determine the fault category of the input sample. During training, sample pairs are constructed. and Perform similarity calculation and classification, and accumulate the loss values; During the model testing phase, only one sample to be tested is input. After feature extraction, it is input into the classification network without going through the similarity branch. The mean squared error (MSE) is used as the loss function of the classification network, and its calculation is shown in formula (10). ； In the formula Represents classification loss. Indicates the true label of the sample. For predicted labels.

6. The intelligent analysis and operation and maintenance method for the entire life cycle of a centrifuge according to claim 5, characterized in that: The feature extraction network is responsible for extracting feature information from the input samples; the relation measurement network optimizes the training process through similarity information, making features of similar samples more clustered and features of dissimilar samples more dispersed. The classification network then directly performs the final fault classification operation; The basic structure of the model consists of a GRU layer and a fully connected layer, where the number of neurons in the fully connected layer is the time step to be predicted. After three levels of pruning—group Lasso neuron pruning, gated pruning, and weight pruning—a smaller weight matrix with more zero values ​​is obtained, enabling faster prediction tasks. Group Lasso neuron pruning is the coarsest-grained pruning. Group Lasso is a structured regularization method that treats each neuron of the GRU as a group and drives the weights of the entire group to approach 0 through Lasso regularization. The formula for calculating the number of Divine Generals after pruning is as follows: ； —The number of neurons after neuron pruning; NPR—Neuron pruning rate; By constructing a gated mask matrix of the same size as the weight matrix, the positions that require gated pruning have a weight of 0, and the positions that do not require gated pruning have a weight of 1; the formula is as follows: ； LSTM introduces three gate functions: the input gate, the forget gate, and the output gate, to control the input, memory, and output values; while the GRU model only has update and reset gates, including: It is a weight matrix; Update Gate: ； The update gate determines whether the memory from the previous step should be passed to the current step; among them, It is the hidden state from the previous step. σ is the input of the current step, and σ is the Sigmoid activation function; Reset Door: ； Candidate hidden state: ； The candidate hidden state is a new candidate value that takes into account the reset gate; where tanh is the hyperbolic tangent activation function; Hidden status update: ； The final hidden state is obtained by combining the updated gate and the candidate hidden state.

Images