Rotary machine health assessment and safety early warning method in sewage treatment environment

Abstract

The invention discloses a health assessment and safety early warning method for a rotating machine in a sewage treatment environment, and the method comprises the steps: obtaining or reading multi-source data of a monitored object from a historical database at a data side through a sensor; the parameters comprise bearing front end temperature, bearing rear end temperature, hydraulic oil temperature, vibration signals, drum rotating speed and differential speed, torque / load, motor current, voltage and optional environment quantity, and a training sample set is constructed. At a modeling side, two self-organizing incremental neural networks are adopted to respectively learn features of a healthy sample and a fault sample, a healthy sample set and a fault sample set are formed, and new fault data encountered in a deployment stage are used to perform incremental learning. According to the method, discrete grading is replaced by continuous quantitative grading, so that the uniformity of information resolution and cross-equipment / cross-working-condition calibers is remarkably improved; the double-stage GAN is introduced, the diversity of generated samples is remarkably improved through small sample enhancement facing rare faults, and it is ensured that the model learns a correct distribution rule from limited samples.

Description

Technical Field

[0001] This invention belongs to the field of industrial equipment condition monitoring and early warning technology, specifically relating to a method and system for quantitative assessment and early warning of the health status of rotating machinery. Background Technology

[0002] Wastewater treatment plant rotating machinery operates under harsh conditions of high speed, high solid load, and strong corrosion, resulting in a mean time between failures (MTBF) often shorter than the designed lifespan. Statistics from a wastewater treatment plant in East China show that in 2021, centrifuges at the plant had a total operating time of 2000 hours, with 27 failures, a failure rate of approximately 9.64%, indicating frequent failures and time-consuming repairs. Currently, manual periodic inspections and simple threshold alarms remain the mainstream equipment monitoring methods in wastewater treatment plants. This approach fails to fully utilize existing equipment monitoring data and requires highly experienced maintenance personnel. Furthermore, while plants regularly shut down rotating machinery for maintenance to mitigate safety risks, many failures go undetected. Frequent maintenance disrupts the stable operation of the entire wastewater treatment system and wastes human and material resources. Constructing a rotating machinery health monitoring system, utilizing data generated during equipment operation for health assessment, and using health scores for real-time quantitative analysis of equipment operating status, provides maintenance guidance to operators and triggers warnings when the health score falls below a threshold. This method can avoid over-maintenance, extend equipment lifespan, and effectively reduce operating costs.

[0003] While there is a wealth of research on the health monitoring of rotating machinery, there are still some shortcomings in wastewater treatment environments, hindering the practical application of related research findings. Firstly, wastewater treatment plants operate in harsh environments, are affected by seasonal factors, and generate diverse operational data characteristics. Furthermore, different types of faults exhibit different manifestations. Traditional methods based on feature statistics and deep learning tend to rely heavily on certain features, resulting in poor model generalization ability and high recognition rates only for certain types of faults. Secondly, current data-driven equipment health monitoring methods largely rely on the assumption of sufficient and balanced equipment status samples. However, real-world classification problems often involve class imbalance, where the number of samples in one class may far exceed those in other classes. In these cases, the classifier may tend to classify all test samples as belonging to the larger class while ignoring samples from the smaller class, leading to poor performance on the smaller class samples. The inherent characteristics of imbalanced datasets (absolute and relative lack of minority class data, data fragmentation, noise) and the limitations of traditional classification algorithms (inappropriate evaluation criteria and inappropriate inductive bias) are key constraints on the accurate and reliable classification of imbalanced datasets. Specifically, in the wastewater treatment environment, sample imbalance manifests in three aspects: 1. Because the normal operating time of equipment is much longer than the failure time, and the plant will quickly shut down the equipment after it shows abnormalities for safety reasons, the dataset of normally operating equipment is much larger than the dataset of failure equipment, resulting in a severe imbalance in sample distribution. In this case, mainstream neural network models may discard some small sample data and classify more samples as normal samples to improve the model's classification accuracy on the entire dataset, resulting in a low failure sample recognition rate. Traditional oversampling and undersampling techniques for solving the sample imbalance problem still have certain limitations. For example, oversampling can easily lead to overfitting of new samples, and undersampling can easily lose information. 2. The number of similar failure samples is too small, the information is incomplete, and it is easily affected by factors such as noise, making it difficult for the algorithm to obtain the true distribution pattern of this type of failure sample data from the limited samples. For example, bearing wear samples have little correlation with temperature features, but because the number of samples is small, the temperature distribution of failure samples is relatively concentrated. The model may regard this temperature distribution pattern as one of the failure features, thus making it difficult to identify similar failures with different temperature distributions. 3. Sample imbalance is not only reflected in the number of failure samples, but also in the types of failures. Rotating machinery failures are numerous and varied, but most failure categories are rare in actual production tasks. Training sample sets often fail to encompass all types of failures, further reducing the model's ability to identify specific types of failures. Finally, most existing research is based on publicly available datasets from equipment or component (e.g., bearings) tests within their respective fields. However, due to the confidentiality requirements of the wastewater treatment industry, obtaining relevant data from actual production processes is extremely difficult.The data obtained by equipment sensors in actual production environments are often of significantly different quality from the data obtained in laboratory environments (for example, the low sampling frequency makes frequency domain analysis methods ineffective), which is one of the reasons why it is difficult to apply related research results. Summary of the Invention

[0004] This invention constructs a health assessment and early warning system for rotating machinery in wastewater treatment environments. To address the aforementioned problems, this invention proposes an improved BI-GAN-SOINN (BI-GenerativeAdversarial Networks-Self-OrganizingIncremental) system with incremental learning capabilities, based on Self-Organizing Incremental Neural Network (SOINN) and Generative Adversarial Networks (GANs).

[0005] Neural Networks (NNN) device health assessment and early warning model.

[0006] This invention proposes using two self-organizing incremental learning neural networks (SOINNs) to learn the features of healthy and faulty samples respectively, avoiding over-reliance on certain local features. Furthermore, SOINN is improved to further enhance the model's generalization ability, recognition accuracy, training speed, and generation speed. In addition, SOINN's characteristics enable efficient online incremental learning, allowing the model to immediately perform incremental learning when encountering new categories of faulty samples, continuously improving its ability to identify low-probability faults in real-world production environments.

[0007] This invention addresses the problem of insufficient fault samples leading to model neglect and incomplete sample features. We propose using a generative adversarial network (GAN) to augment minority fault samples, ensuring overall balance in the training set and preventing these samples from being ignored. To overcome the issue of incomplete sample features and further improve the augmentation effect, we design a two-stage generative adversarial network (BAN) for sample augmentation, enhancing the diversity of generated samples and further improving model training performance.

[0008] This invention addresses the issue of missing rare samples by proposing a complete autonomous incremental learning mechanism for BI-GAN-SOINN. When the model encounters new types of fault samples during the deployment phase, it can efficiently and autonomously complete the enhancement and incremental learning of the new samples, enabling the model to effectively utilize new fault samples for unsupervised incremental learning in real-world working environments.

[0009] On the data side, this invention acquires multi-source data of the monitored object through sensors or reads it from historical databases, including but not limited to bearing front-end temperature, bearing rear-end temperature, hydraulic oil temperature, vibration signals, drum speed and differential speed, torque / load, motor current, voltage, and optional environmental quantities, thereby constructing a high-quality training sample set. Considering the inconsistent quality of on-site data (such as abnormal 0 / 1 values, out-of-order timestamps, missing measurements, and pulse interference), data governance is first performed, including time alignment, deduplication, outlier removal or replacement, resampling, and robust normalization. Subsequently, by combining mutual information / stability selection / sparse constraints and operating condition compensation, a set of key variables for specific rotating equipment is selected (for example, centrifuges use vibration, solid-phase end temperature, liquid-phase end temperature, and cooling water temperature for health assessment, consistent with the threshold alarm system currently used in the water plant), to reduce the model input dimension and computational complexity. After completing variable selection and governance, the samples are divided into health datasets and fault datasets based on maintenance records, inspection results, and expert review for subsequent model training and validation.

[0010] On the modeling side, this invention employs two self-organizing incremental neural networks to learn the features of healthy and faulty samples respectively, forming healthy and faulty sample sets. New fault data encountered during the deployment phase is used for incremental learning. New faulty samples are categorized into existing and new faulty samples based on whether the pattern can correctly trigger an alarm. Existing faulty samples undergo sample augmentation using the SMOTE method, followed by incremental learning using SOINN2. For fault types that are appearing for the first time or have poor warning effects, this project proposes a two-stage generative adversarial network with transfer learning capabilities based on the GAN network structure to improve the limitations of traditional sample augmentation methods in terms of diversity. GAN1 is primarily responsible for transfer learning, comparing the preprocessed input data with the distribution of samples from normal device operation. Indicators exceeding the normal operating range are considered related to fault features, and the discriminator focuses on learning the intrinsic correlations of these data to ensure that the distribution of these fault-related indicators in the generated samples is consistent with the original faulty samples. Indicators within the normal operating range are considered unrelated to fault features, and the generator and discriminator respectively learn the distribution pattern of this parameter in the entire sample set, thereby ensuring the diversity of generated samples and avoiding the model learning incorrect distribution patterns due to insufficient samples of this type of fault. GAN2 is based on CGAN and learns from all samples (with labels) and new samples generated by GAN1. It then selectively generates new samples to avoid generating other types of faulty or healthy samples during the incremental process, thus ensuring the sample quality for subsequent incremental training of the SOINN scoring model. To improve training and inference efficiency, both GAN1 and GAN2 use lightweight structures similar to Fast-GAN.

[0011] On the scoring and decision-making side, this invention uses the distance from the current sample to the healthy prototype set and the faulty prototype set to obtain a continuous health score of 0–100 via a health scoring formula. Alarm thresholds can be set for different devices, with a default value of 25.

[0012] On the system and deployment side, this invention includes functional modules such as data acquisition and governance, key variable screening and operating condition compensation, dual SOINN prototype learning, dual-stage GAN sample augmentation, online incremental updates and model management, visualization, and edge-cloud collaboration. The system can be deployed on industrial control computers or edge platforms such as ARM to achieve low-latency online scoring, and uploads data that meets the confidentiality requirements of water plants to the cloud for viewing by an app via its built-in communication module, meeting the resource constraints and safety compliance requirements of industrial sites. It has already been deployed in a water plant production environment. This invention also provides a computer-readable storage medium storing program instructions, which, when executed on a processor, implement the above-mentioned method steps, thereby completing the entire process of rotating equipment health assessment and safety early warning.

[0013] Compared with existing technologies, this invention significantly improves information resolution and consistency of standards across devices and operating conditions by replacing discrete grading with continuous quantized scoring. The BI-SOINN scoring architecture effectively avoids problems such as poor generalization ability and the easy neglect of small sample data that are common in traditional data-driven algorithms. The introduction of a two-stage GAN, with small sample augmentation for rare faults, significantly improves the diversity of generated samples, ensuring that the model learns the correct distribution patterns from a limited number of samples. Relying on SOINN's efficient online incremental learning mechanism, the model can quickly absorb new or complex fault modes and remain stable, thereby effectively reducing unplanned downtime and maintenance costs while ensuring interpretability and deployability, and improving reliability and economy. The relevant results have been deployed in a wastewater treatment plant for health assessment of the horizontal screw centrifuge and have been validated on booster pumps and blowers. Attached Figure Description

[0014] 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.

[0015] Figure 1 Overall block diagram of the BI-GAN-SOINN algorithm.

[0016] Figure 2 Flowchart of the health assessment system.

[0017] Figure 3 SOINN runtime diagram.

[0018] Figure 4 : GAN structure diagram.

[0019] Figure 5 : Structural diagram of a horizontal screw dewatering machine.

[0020] Figure 6 : Structure diagram of the booster pump.

[0021] Figure 7 : Blower structure diagram. Detailed Implementation

[0022] 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.

[0023] This invention constructs a health scoring model for rotating machinery, which 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 employs a dual self-organizing incremental neural network structure for feature extraction and state modeling, combined with an incremental learning mechanism to achieve continuous adaptive updates during operation. This method effectively solves the problems of insufficient generalization ability, scarce fault samples, incomplete type coverage, and difficulty in efficient updates found in traditional models, significantly improving the robustness and scalability of the health assessment model under complex operating conditions. Please refer to [link / reference]. Figure 2 , Figure 2 A flowchart illustrating the operation of a device health scoring model constructed for an embodiment of this application is provided. The method may include the following steps.

[0024] Step 1: The equipment hardware can connect to the water plant's database to obtain the required data. The initial dataset includes, but is not limited to, the bearing front end temperature, bearing rear end temperature, cooling water temperature, vibration, and speed difference collected by the equipment's sensors.

[0025] 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 uses this data to obtain the different relevant variables of the monitored equipment and construct an initial dataset. Alternatively, the server can retrieve historical data from the factory's database as the initial dataset.

[0026] Step 2: Filter the initial dataset, select key variables as the health score dataset, and divide it into a healthy dataset and a fault dataset for model training and testing.

[0027] Step 3: Use the health dataset and the fault dataset as two inputs to train the two neural networks SOINN1 and SOINN2 respectively, and learn their health features and fault features.

[0028] SOINN typically employs a two-layer structure. The first layer learns basic clusters, and the second layer reprocesses the output of the first layer to generate a more abstract representation. By calculating the similarity between samples through distance, it achieves efficient processing of high-dimensional, high-volume data.

[0029] The first layer of SOINN receives the raw data input and generates an initial topology representing nodes and edges. The second layer takes the learning result from the first layer as input and performs another learning iteration after each epoch (LT) to obtain a more concise and stable result. For a new input sample, SOINN determines whether to generate a new node by judging its distance to the two nearest neurons. Assuming the input data is ξ, and N represents the set of all nodes in the network, the neurons in set N closest to ξ are selected as the winners. Second winner The calculation method is as follows.

[0030] ;

[0031] ;

[0032] In the formula, This represents the weight of node i. If two nodes are connected by an edge, they are called adjacent nodes. and If no edges are connected, connect the two points and set the age of the edge to 0.

[0033] When node i has neighboring nodes, its activation threshold is... This is the Euclidean distance between node i and its farthest neighbor. If i has no neighbors, the threshold is defined as the minimum Euclidean distance between i and all other nodes in the network. The calculation formula is as follows:

[0034] ;

[0035] Among them It is the sum of the neighboring nodes of node i.

[0036] When the input signal ξ is The distance exceeds the corresponding threshold, i.e. ,or If the input signal is too far from the cluster where the winning node is located, then a new node should be created. Conversely, according to Equations 4 and 5 respectively... and its neighboring nodes To update the weights, in the formula and Manually set and In addition to the learning rate, the network also sets an age threshold; once the age threshold is exceeded... If an edge has not been activated for a long time, it should be deleted. Nodes without any connected edges should also be deleted.

[0037] ;

[0038] ;

[0039] A node insertion is also performed when the input information is a multiple of λ. SOINN determines whether this insertion is necessary by calculating the average error of each point in that region. The initial value is 0, which represents the distance between ξ and the nearest point. The cumulative value of the squared differences of the distances is calculated as shown in Equation 6.

[0040] ;

[0041] The learning algorithm for the second layer of SONN is largely the same as that for the first layer, with only slight differences in the calculation of the node threshold. After the first layer of learning is completed, a coarse clustering result has been obtained. The average distance between nodes in cluster C is used as the intra-cluster distance of that cluster. Using clustering , The minimum distance between any two nodes that do not belong to the same class is taken as the inter-class distance. As shown in Equations 7 and 8. The threshold of the second layer is a certain value greater than... But smaller than The value of .

[0042] ;

[0043] ;

[0044] Step 4: During actual deployment, when the system detects an update to the target data, it will save the data and perform a scoring. SOINN1 and SOINN2 will output... and The similarity between the current input sample of the device and its healthy and faulty states is quantified. The higher the similarity, the smaller the distance of the sample to that cluster in space, and the smaller the corresponding value. The formula for calculating the health score H is as follows:

[0045] ;

[0046] To further improve the accuracy of the algorithm and save computational resources, this project has made two improvements to the SOINN neural network:

[0047] 1. Vectorization and KDTree (k-dimensional tree) are used to improve the search speed of nearest neighbor nodes. In high-dimensional space, distance calculation is essentially a metric space search problem. KDTree is a binary space partitioning tree that reduces the complexity of nearest neighbor search from linear to logarithmic by recursively partitioning the hyperrectangular space. In actual computation, vectorization operations based on libraries such as NumPy utilize the CPU's SIMD (Single Instruction Multiple Data) instruction set to compute multiple distances in parallel, reducing loop overhead.

[0048] 2. The improved SOINN checks the number of nodes after each training iteration. If it exceeds a set value, it searches the entire network for the nearest node and merges them until the number of nodes drops to the set value. This improves node quality and reduces computational and storage overhead.

[0049] Step 5: If the health score falls below the set alarm threshold, the system will issue an alarm. Because the model is trained by classifying samples nearing fault conditions as fault samples, it can issue an alarm before the actual values ​​collected by the sensors reach the safety threshold. The sensitivity can be adjusted by modifying the alarm score threshold.

[0050] Step 6: The system automatically updates the relevant values ​​and charts on the visualization interface and uploads data that meets confidentiality requirements to the cloud via its built-in communication module for viewing on web pages and apps.

[0051] Step 7: When the model collects fault samples during field deployment, incremental training will be automatically performed. The algorithm classifies the fault into an existing fault type or a new fault type based on whether BI-SOINN correctly issues an alarm. Existing faults will be directly augmented using SMOTE, while new fault types will be augmented using a two-stage GAN. GAN1 first determines whether the parameters of the fault sample are within the normal distribution range. Parameters within the normal range are considered to be related to fault characteristics, and GAN1 will ensure that the distribution of the parameter in the generated samples is consistent with the distribution of the parameter in the full sample set. Indicators that are outside the normal operating range are considered to be related to fault characteristics, and the discriminator will focus on learning the intrinsic correlation of these parameters to ensure that the distribution of these fault-related indicators in the generated samples is consistent with the original fault samples. GAN2, based on CGAN, will learn all samples (labeled) and the new samples generated by GAN1, and then generate new samples in a targeted manner, thereby avoiding the generation of other types of fault samples or even healthy samples during the incremental process, ensuring the sample quality of subsequent incremental training of the SOINN scoring model. To improve training and inference efficiency, both GAN1 and GAN2 use a lightweight structure similar to Fast-GAN.

[0052] The training process of a generative adversarial network is an adversarial learning process, and its computation process is as follows:

[0053] ;

[0054] Theoretically, this maximal game in GAN training can converge, meaning the generator G and discriminator D reach Nash equilibrium, ultimately generating realistic data. However, in practice, the results are often less than ideal. Original GANs frequently suffer from vanishing gradients, training oscillations, and mode collapse. Since the inception of GANs, improvements to GAN functions, network structures, and training methods have been continuous. CGAN and Fast-GAN are commonly used GAN variants. CGAN, while retaining the original GAN ​​structure, introduces conditional information y into the generator G and discriminator D to constrain the generated samples. The training objective function is shown below:

[0055] ;

[0056] In the formula: D(x|y) represents the conditional information y introduced into the discriminator, and G(x|y) represents the conditional information y introduced into the generator. In CGAN, the conditional information can be in the form of a vector or a feature map. Fast-GAN mainly addresses the problems of slow training speed and difficulty in convergence of the original GAN, and significantly improves training efficiency and generation quality by simplifying the network structure and optimizing the training strategy.

[0057] Step 8: After completing the sample increment, SOINN2 begins incremental training. Once incremental training is complete, the system will re-evaluate the fault sample. If the fault is correctly identified (score below the alarm threshold), the scoring ends. If the fault still cannot be correctly identified, the system returns to Step 7.

[0058] The online incremental update of the neural network model used in this application is a continuous learning technique that allows the model to be updated in real time as new data arrives. The characteristics of the SOINN neural network mean that incremental learning only requires the generation and merging of nodes according to a process, without retraining the entire model. This saves computational resources and effectively avoids the forgetting of existing knowledge. Incremental updates enable the model to continuously improve itself through new fault samples during runtime, and it can immediately learn rare fault features not present in the training set. This effectively solves the problems of insufficient model generalization ability, insufficient number of fault samples, and missing types.

Claims

1. A method for health assessment and safety early warning of rotating machinery in a wastewater treatment environment, characterized in that, The method for assessing the health and safety of rotating machinery in a wastewater treatment environment is used in a rotating machinery health scoring model, which is deployed on a server terminal with computing capabilities. The server includes a personal computer, tablet computer, smartphone, or personal digital assistant. The rotating machinery health scoring model uses a dual self-organizing incremental neural network structure for feature extraction and state modeling, and combines an incremental learning mechanism to achieve continuous adaptive updates of the rotating machinery health scoring model during operation.

2. The method for health assessment and safety early warning of rotating machinery in a wastewater treatment environment according to claim 1, characterized in that, Two self-organizing incremental learning neural networks are used to learn the features of healthy samples and faulty samples respectively, avoiding over-reliance on certain local features, and SOINN is improved.

3. The method for health assessment and safety early warning of rotating machinery in a wastewater treatment environment according to claim 1, characterized in that, Generative adversarial networks are used to augment minority fault samples, thereby balancing the overall training set and preventing minority samples from being ignored by the rotating machinery health scoring model.

4. The method for health assessment and safety early warning of rotating machinery in a wastewater treatment environment according to claim 1, characterized in that, To address the lack of some rare samples, a complete autonomous incremental learning mechanism of BI-GAN-SOINN is adopted. When the model encounters new types of fault samples during the deployment phase, it completes the enhancement and incremental learning of the new samples, enabling the rotating machinery health scoring model to effectively utilize new fault samples for unsupervised incremental learning in the actual working environment.

5. The method for health assessment and safety early warning of rotating machinery in a wastewater treatment environment according to claim 1, characterized in that, On the data side, multi-source data of the monitored object is acquired through sensors or read from historical databases, including bearing front end temperature, bearing rear end temperature, hydraulic oil temperature, vibration signal, drum speed and differential speed, torque / load, motor current, voltage and optional environmental parameters, thereby constructing a high-quality training sample set; First, time alignment, deduplication, outlier removal or replacement, resampling, and robust normalization data governance are performed; then, by combining mutual information / stability selection / sparse constraints and operating condition compensation, a set of key variables for specific rotating equipment is obtained. After variable screening and management, the samples were divided into healthy datasets and fault datasets based on maintenance records, inspection results, and expert review, for subsequent model training and validation.

6. The method for health assessment and safety early warning of rotating machinery in a wastewater treatment environment according to claim 1, characterized in that, On the modeling side, two self-organizing incremental neural networks are used to learn the features of healthy samples and fault samples respectively, forming a set of healthy samples and a set of fault samples, and incremental learning is performed using new fault data encountered during the deployment phase. New fault samples are categorized into existing fault samples and new fault samples based on whether the pattern can correctly issue an alarm. Existing fault samples will be augmented using the SMOTE method, followed by incremental learning using SOINN2. For fault types that are appearing for the first time or whose early warning effect is poor, a two-stage generative adversarial network with transfer learning function is proposed based on the GAN network structure. Among them, GAN1 is responsible for transfer learning. During runtime, it first compares the distribution of the preprocessed input data with the distribution of the normal operating samples of the device. Indicators that are outside the normal operating range are considered to be related to fault features. The discriminator will focus on learning the intrinsic correlation of these data to ensure that the distribution of these fault-related indicators in the generated samples is consistent with the original fault samples. Indicators within the normal operating range are considered irrelevant to fault characteristics. The generator and discriminator will learn the distribution pattern of this parameter in the full sample set, thereby ensuring the diversity of generated samples and avoiding the model learning the wrong distribution pattern due to too few samples of this type of fault. GAN2 is based on CGAN and will learn all samples and new samples generated by GAN1, and then generate new samples in a targeted manner to ensure the sample quality of incremental training of the subsequent SOINN scoring model. Both GAN1 and GAN2 use the lightweight structure of Fast-GAN.

7. The method for health assessment and safety early warning of rotating machinery in a wastewater treatment environment according to claim 1, characterized in that, On the scoring and decision-making side, a continuous health score of 0–100 is obtained based on the distance of the current sample to the healthy prototype set and the faulty prototype set through the health scoring formula.

8. The method for health assessment and safety early warning of rotating machinery in a wastewater treatment environment according to claim 1, characterized in that, On the deployment side, it includes modules for data collection and governance, key variable screening and operating condition compensation, dual SOINN prototype learning, dual-stage GAN sample augmentation, online incremental update and model management, visualization and edge-cloud collaboration. Deployed on industrial control computers or ARM edge platforms to achieve low-latency online scoring, and uploads data that meets the water plant's confidentiality requirements to the cloud via its built-in communication module for viewing by the APP.

9. The method for health assessment and safety early warning of rotating machinery in a wastewater treatment environment according to claim 1, characterized in that, The implementation process of this system includes the following steps: Step 1: The equipment hardware can connect to the water plant database to obtain the required data. The initial dataset obtained includes, but is not limited to, the bearing front end temperature, bearing rear end temperature, cooling water temperature, vibration, and speed difference collected by the equipment sensors. Sensors are placed around the equipment that needs maintenance in advance. The sensors collect data from multiple relevant variables of the equipment and transmit the data directly to the server. The server then obtains the different relevant variables of the equipment under test and forms an initial dataset. Alternatively, the server can obtain historical data from the factory's database as the initial dataset. Step 2: Filter the initial dataset, select key variables as the health score dataset, and divide it into a healthy dataset and a fault dataset for model training and testing; Step 3: Use the health dataset and the fault dataset as two inputs to train the two neural networks SOINN1 and SOINN2 respectively, and learn their health features and fault features. SOINN employs a two-layer structure: the first layer learns basic clustering, and the second layer further processes the output of the first layer. The first layer of SOINN receives the raw data input and generates an initial topology representing nodes and edges. The second layer takes the learning result from the first layer as input and performs another learning iteration after each epoch (LT) to obtain a more concise and stable result. For a new input sample, SOINN determines whether to generate a new node by judging its distance to the two nearest neurons. Let the input data be ξ, and N represent the set of all nodes in the network. The neurons in set N that are closest to ξ are selected as the winners. Second winner The calculation method is as follows: ； ； In the formula, This represents the weight of node i; if two nodes are connected by an edge, then these two nodes are called adjacent nodes; if... and If there are no connected edges, connect the two points and set the age of the edge to 0. When node i has neighboring nodes, its activation threshold is... The threshold is the Euclidean distance between i and the farthest neighboring node i; if i has no neighbors, the threshold is defined as the minimum Euclidean distance between i and other nodes in the network; the calculation formula is as follows: ； in, The sum of the neighboring nodes of node i; When the input signal ξ is The distance exceeds the corresponding threshold, i.e. ,or If the input signal is far from the cluster where the winning node is located, a new node is created; otherwise, according to Equations 4 and 5 respectively... and its neighboring nodes To update the weights, in the formula and Manually set and In addition to the learning rate, the network also sets an age threshold; once the age threshold is exceeded... If an edge has not been activated for a long time, it should be deleted; nodes without any edge connections should also be deleted. ； ； When the amount of input information is an integer multiple of λ, a node insertion is also performed; SOINN determines whether this insertion is necessary by calculating the average error of each point in the region; error The initial value is 0, which represents the distance between ξ and the nearest point. The cumulative value of the squared differences of the distances is calculated as shown in Equation 6; ； The learning algorithm for the second layer of SONN is the same as that for the first layer. After the first layer is completed, a coarse clustering result has been obtained. The average distance between nodes in cluster C is used as the intra-cluster distance of that cluster. Using clustering , The minimum distance between any two nodes that do not belong to the same class is taken as the inter-class distance. As shown in Equations 7 and 8; the threshold of the second layer is a certain value greater than... But smaller than The value; ； ； Step 4: During actual deployment, when the system detects an update to the target data, it will save the data and score it; SOINN1 and SOINN2 will output respectively. and The similarity between the current input sample of the device and its healthy and faulty states is quantified. The higher the similarity, the smaller the distance of the sample to the cluster in space, and the smaller the corresponding value. The formula for calculating the health score H is as follows: ； Step 5: If the health score is lower than the set alarm threshold, the system will issue an alarm. Because the model is trained by classifying samples that are close to the fault state as fault samples, the model can issue an alarm before the actual value collected by the sensor reaches the safety threshold. This can be adjusted by modifying the alarm score threshold. Step 6: The system automatically updates the relevant values ​​and charts on the visualization interface and uploads data that meets confidentiality requirements to the cloud via its built-in communication module for viewing on web pages and apps; Step 7: When the model collects fault samples during field deployment, it will automatically perform incremental training. Based on whether BI-SOINN correctly issues an alarm, the fault will be classified as an existing fault type or a new fault type. Existing faults will be directly augmented using SMOTE, while new fault types will be augmented using a two-stage GAN. GAN1 first determines whether each parameter of the fault sample is within the normal distribution range. Parameters within the normal range are considered related to fault characteristics, and GAN1 will ensure that the distribution of this parameter in the generated samples is consistent with the distribution of this parameter in the entire sample set. Indicators exceeding the normal operating range are considered related to fault characteristics, and the discriminator will focus on learning the intrinsic correlation of these parameters to ensure that the distribution of these fault-related indicators in the generated samples is consistent with the original fault samples. GAN2, based on CGAN, learns from all samples and the new samples generated by GAN1, and then generates new samples accordingly. Both GAN1 and GAN2 use a lightweight structure similar to Fast-GAN. The training process of a generative adversarial network is an adversarial learning process, and its computation process is as follows: ； Since the introduction of GANs, improvements to GAN functions, network structures, and training methods have continued. Among them, CGAN and Fast-GAN are commonly used GAN variants. CGAN, while retaining the original GAN ​​structure, introduces conditional information y into the generator G and discriminator D to constrain the generation of samples. The training objective function is shown below: ； In the formula: D(x|y) refers to the conditional information y introduced in the discriminator, and G(x|y) refers to the conditional information y introduced in the generator; in CGAN, the conditional information can be in the form of a vector or a feature map. Step 8: After completing the sample increment, SOINN2 begins incremental training; after the incremental training is completed, the system will re-score the faulty sample. If the fault can be correctly identified, the scoring ends; if the fault still cannot be correctly identified, it jumps back to step 7.

10. The method for health assessment and safety early warning of rotating machinery in a wastewater treatment environment according to claim 9, characterized in that, Two improvements were made to the SOINN neural network: The first is to improve the search speed of nearest neighbor nodes by using vectorization and KDTree; Secondly, the improved SOINN checks the number of nodes after each training iteration.

Images