Centrifugal circulating pump group cooperative training method based on federal learning

Abstract

A centrifugal circulating pump group cooperative training method based on federated learning belongs to the technical field of artificial intelligence and industrial Internet of Things, and constructs a federated learning system composed of a central aggregation server and a plurality of edge clients, and realizes safe transmission of model parameters through gRPC encryption communication. The edge client carries out model training based on pump group operation data collected locally, and Gaussian noise is added by adopting a differential privacy technology to protect data privacy; the central server performs secure aggregation on the encrypted model weight update, generates an enhanced global model and distributes the enhanced global model to each client; and each node further realizes model personalized adaptation through local fine tuning, and realizes low-delay deployment by using model quantification and an RKNN inference engine. According to the method, on the premise of ensuring that the data is not delocalized, the generalization ability, the control precision and the system robustness of the model in the heterogeneous pipe network environment are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and industrial Internet of Things technology, and relates to a method for collaborative training, model updating and personalized deployment of centrifugal pump groups based on a federated learning framework. It is used to improve the adaptability of pump state reasoning models in different pipeline network environments under the premise of data privacy protection. Background Technology

[0002] Traditional single-end pump control systems suffer from data silos, making them difficult to adapt to the diversity of pipeline environments. Furthermore, the performance of neural network-based intelligent pump control technology is highly dependent on the completeness of training data; a single model lacks generalization ability across all scenarios, and obtaining a "full-condition" dataset covering all possible operating conditions and pipeline configurations is extremely difficult and costly in reality. The operating data of a single pump only reflects its characteristics under its specific pipeline environment, and models trained in this way often perform poorly in other diverse environments, indicating insufficient generalization ability. The traditional solution is to centrally upload the operating data of all pump groups to the cloud for unified training. However, this approach has serious drawbacks: 1) Data privacy and security risks: Pump operating data contains sensitive information such as user water usage habits and pipeline layout; direct upload may lead to privacy leaks. 2) High data transmission costs: Uploading large amounts of time-series data consumes significant bandwidth resources. 3) Compliance issues: Increasingly, data regulations restrict the cross-border and cross-system transmission of raw data. Federated learning, as a distributed machine learning technology, allows for collaborative training by exchanging model parameters rather than raw data without uploading local data, offering a potential solution to these problems. Summary of the Invention

[0003] To overcome the shortcomings of existing technologies, this invention proposes a collaborative training method for centrifugal circulating pump groups based on a federated learning architecture. This method constructs a collaborative computing system composed of a central aggregation server and distributed edge clients, realizing a closed-loop process from pump group data acquisition, distributed model training, secure aggregation to personalized model deployment. The system consists of a central aggregation server and edge clients deployed at N pump stations. The central server, based on a high-performance computing platform, runs a federated learning scheduling core and is primarily responsible for client management, training strategy execution, secure model aggregation, and encrypted communication management based on the gRPC protocol. The server never stores any raw data, fundamentally ensuring data privacy. The edge clients, based on the Rockchip RK3566 embedded platform, carry out local data acquisition, model training, and update functions. Specifically, they collect and store local pump group operating data (including speed, power, flow rate, head, etc.), with all data stored in a local embedded database to ensure data remains within the local domain. The system operation flow is as follows: Each edge client first obtains the initialized RNN baseline model from the server; then, based on private data in its local database, it trains the model and generates local weight updates; after training, the model weight parameters are uploaded to the server via a gRPC encrypted channel, and a differential transmission mechanism is used to significantly reduce communication overhead; the server applies a federated averaging algorithm to aggregate the collected weights, generating an enhanced global model; finally, the aggregated model is distributed to each edge node. Based on this, each edge client further incorporates local data to fine-tune the model, thereby obtaining a high-precision state reasoning model that combines global generalization ability with local environment adaptability. This invention, while ensuring end-to-end data privacy and security, achieves efficient collaborative training and rapid personalized adaptation of the pump swarm model, significantly improving the control accuracy and robustness of the system in heterogeneous deployment environments.

[0004] The technical solution adopted by this invention to solve its technical problem is:

[0005] A collaborative training method for centrifugal circulating pump groups based on federated learning, the method comprising the following steps:

[0006] Step 1: Initialization of the federated learning system: In multiple heterogeneous pipe network environments, collect the status data of water pump operation to build an initial training set. The recurrent neural network model based on gated recurrent units is trained by the time series backpropagation algorithm. After training, a distributed learning architecture is built. The central server is configured with a federated learning coordination service based on the Flower framework, edge clients are configured, and differential privacy transmission encrypted communication based on the gRPC protocol is established.

[0007] Step 2, Collaborative Training Process: The central server initializes a lightweight recurrent neural network model (RNN) as the global baseline model and distributes it to each edge client through a secure channel. Each client uses locally collected pump group operation data for local training and updates the local model parameters using the stochastic gradient descent algorithm. During training, Gaussian noise is added through differential privacy to protect data privacy. After training is completed, the model weight update amount ΔW is encrypted to ensure that the original data does not go out of domain.

[0008] Step 3, Secure Aggregation and Model Update: The server uses a weighted average algorithm to securely aggregate the collected weight updates. After aggregation, the new model is distributed to all clients, completing one round of federated learning iteration.

[0009] Step 4: Personalized Model Deployment: After receiving the global model, each edge client fine-tunes and trains the model based on its local unique data to adapt the model to the specific characteristics of the pipeline network environment.

[0010] Step 5: Continuous optimization mechanism: Establish a model version management mechanism to record the contribution of each client and the model update history.

[0011] Furthermore, in step one, the forward propagation process is represented as follows:

[0012] h t =GRU(x t ,h t-1 (1);

[0013] Where x t = [Speed, Power, Flow Rate, Head] is the input feature sequence at time t, h t-1 It is in a hidden state;

[0014] The loss function uses mean squared error:

[0015]

[0016] Optimize model parameters using gradient descent algorithm:

[0017]

[0018] Preferably, in step one, the edge client is based on the Rockchip RK3566 platform and is equipped with a data acquisition module and a local storage unit.

[0019] Furthermore, in step two, the local model parameters are updated using the stochastic gradient descent algorithm:

[0020]

[0021] Where B represents mini-batch data and η represents the learning rate.

[0022] Furthermore, in step three, the secure aggregation calculation is as follows:

[0023]

[0024] Where n k Let n be the data volume of the k-th client, and n be the total data volume. The weight update amount is given for the k-th client, followed by updating the global model parameters.

[0025] In step four, an early stopping mechanism is used during fine-tuning training to prevent overfitting. Training stops when the validation set loss no longer decreases after three consecutive rounds. The optimized model is deployed to the local inference engine through the following steps: First, a model quantization tool is used to convert the FP32 model to INT8 precision, reducing the model size and improving inference speed while keeping the accuracy drop within an acceptable range. Finally, it is integrated into a real-time inference framework based on the RKNN inference engine to achieve inference latency less than a set threshold. After deployment, a continuous monitoring mechanism is established to collect the inference error metric Δe=|y in real time. pred -y actual When the moving average error exceeds the threshold, model recalibration is automatically triggered.

[0026] In step five, the formula for evaluating client contribution is:

[0027] c k =(n k ×ΔL k ) / t k (6);

[0028] Where n k Let ΔL be the data volume of the k-th client. k To represent the performance improvement of this client-side model, t k This is the last update time;

[0029] Configure a periodic federated training task to perform global aggregation at set intervals: An adaptive threshold mechanism is adopted to automatically trigger the retraining process when an anomaly is detected, ensuring continuous system optimization and stable operation.

[0030] Preferably, in step five, anomaly refers to a decrease in model performance ΔL > δ or a change in pipeline impedance characteristics exceeding 15%, where δ = μ + 2σ, μ is the average performance change, and σ is the standard deviation.

[0031] Compared with the prior art, the beneficial effects of this invention are as follows:

[0032] 1) End-to-end encrypted communication and differential privacy technology are employed to ensure that the original data always remains on the local edge device, with only encrypted model parameter updates transmitted. A secure communication channel is established through gRPC differential privacy, effectively preventing man-in-the-middle attacks and data leakage risks. Simultaneously, a Gaussian noise mechanism is added during local training to ensure that even if model parameters are intercepted, the original data cannot be inferred.

[0033] 2) Knowledge sharing and collaborative optimization were achieved through a federated learning framework. The weighted federated average algorithm was used to aggregate model updates from each edge node, where the weights were dynamically adjusted based on data quality, data volume, and client reliability, significantly improving the generalization ability of the global model.

[0034] 3) Low-latency inference was achieved through model lightweighting technology and edge computing architecture. By employing model quantization (FP32 to INT8), distillation, and pruning techniques, the model size was compressed to 1 / 4 of its original size, increasing inference speed by more than 3 times. Combined with the NPU acceleration capabilities of the RK3566 chip, an end-to-end inference latency of less than 10ms was achieved, fully meeting the requirements of real-time control.

[0035] 4) An automatic optimization mechanism based on dynamic thresholds is adopted. The system continuously monitors model performance indicators, including inference error, convergence status and generalization ability. When performance degradation (ΔL>δ) or changes in the operating environment are detected, the model retraining process is automatically triggered to ensure that the system always remains in the optimal state. Attached Figure Description

[0036] Figure 1 A schematic diagram of global training and aggregation for federated learning.

[0037] Figure 2 A flowchart for personalized fine-tuning and deployment inference at the edge. Detailed Implementation

[0038] The invention will now be further described with reference to the accompanying drawings.

[0039] Reference Figure 1 and Figure 2 A collaborative training method for centrifugal circulating pump groups based on federated learning, the method comprising the following steps:

[0040] Step 1: Federated Learning System Initialization: In multiple heterogeneous pipe network environments, collect pump operation status data to construct an initial training set. A recurrent neural network model based on a gated recurrent unit (GRU) is trained using a time-series backpropagation algorithm. Its forward propagation process is represented as follows:

[0041] h t =GRU(x t ,h t-1(1);

[0042] Where x t = [Speed, Power, Flow Rate, Head] is the input feature sequence at time t, h t-1 For the hidden state, the loss function uses mean squared error:

[0043]

[0044] Optimize model parameters using gradient descent algorithm:

[0045]

[0046] After training, a distributed learning architecture is constructed. The central server adopts a high-performance computing platform and is configured with federated learning coordination services based on the Flower framework. The edge clients are based on the Rockchip RK3566 platform, equipped with data acquisition modules and local storage units, and establish differential privacy transmission encrypted communication based on the gRPC protocol.

[0047] Step 2, Collaborative Training Process: The central server initializes a lightweight recurrent neural network (RNN) model as the global baseline model and distributes it to each edge client through a secure channel. Each client uses locally collected pump group operation data (speed, power, flow rate, head) for local training and updates its local model parameters using the stochastic gradient descent algorithm.

[0048]

[0049] Where B represents mini-batch data and η represents the learning rate. During training, Gaussian noise is added using differential privacy to protect data privacy. After training, the model weight update ΔW is encrypted to ensure that the original data does not go out of domain.

[0050] Step 3, Secure Aggregation and Model Update: The server uses a weighted average algorithm to securely aggregate the collected weight updates, calculated as follows:

[0051]

[0052] Where n k Let n be the data volume of the k-th client, and n be the total data volume. The weight update amount is given for the k-th client, followed by updating the global model parameters. Once aggregation is complete, the new model is distributed to all clients, completing one round of federated learning iteration.

[0053] Step 4: Personalized Model Deployment: After receiving the global model, each edge client fine-tunes the model based on its local data to adapt it to the specific characteristics of the network environment. An early stopping mechanism is used during fine-tuning to prevent overfitting; training stops when the validation set loss no longer decreases after three consecutive rounds. The optimized model is deployed to the local inference engine through the following steps: First, a model quantization tool is used to convert the FP32 model to INT8 precision, reducing the model size by 75% and improving inference speed while keeping the accuracy drop within an acceptable range. Finally, it is integrated into a real-time inference framework based on the RKNN inference engine, achieving an inference latency of less than 10ms. After deployment, a continuous monitoring mechanism is established to collect the inference error metric Δe=|y in real time. pred -y actual When the moving average error exceeds the threshold, model recalibration is automatically triggered.

[0054] Step 5: Continuous Optimization Mechanism: Establish a model version management mechanism to record the contribution of each client and the model update history. The formula for evaluating client contribution is:

[0055] c k =(n k ×ΔL k ) / t k (6);

[0056] Where n k Let ΔL be the data volume of the k-th client. k To represent the performance improvement of this client-side model, t k This is the last update time. Set up a periodic federated training task to perform global aggregation every T = 24 hours: An adaptive threshold mechanism is adopted. When a model performance degradation ΔL>δ (where δ=μ+2σ, μ is the average performance change, and σ is the standard deviation) is detected or the operating environment changes significantly (such as the pipeline impedance characteristics change by more than 15%), the retraining process is automatically triggered to ensure continuous optimization and stable operation of the system.

[0057] In this embodiment, the central server initializes a lightweight GRU recurrent neural network model as the global baseline model (Step 1). The server distributes this global model to all edge clients participating in federated learning (deployed on the Rockchip RK3566 platform) via an encrypted communication channel based on gRPC. Each edge client, upon receiving the global model, initiates its local training process (Step 2). The training data originates from pump group operating data (speed, power, flow rate, head) collected locally and stored in an SQLite database. To strictly protect data privacy, differential privacy is introduced during training by adding Gaussian noise to gradient calculations to ensure the original data remains absolutely local. The client uses a stochastic gradient descent algorithm to update the local model parameters and calculates... After training is complete, the client encrypts the calculated model weight update ΔW and uploads it to the server.

[0058] After collecting a sufficient number of client updates, the central server initiates the secure aggregation process (step three). The server uses a weighted average algorithm to aggregate the encrypted updates, calculated as follows: Where n k Let n be the data volume of the k-th client, and n be the total data volume. The weight update amount is given for the k-th client, followed by updating the global model parameters. The updated global model will be redistributed to all clients, completing one round of federated learning iteration. A continuous optimization and evaluation mechanism (step five) is implemented at the end of the process. The system monitors for conditions such as model performance degradation (ΔL>δ) or significant changes in the operating environment. Once triggered, a new round of training will be automatically started to ensure continuous model optimization.

[0059] Reference Figure 2 After receiving the latest global model from the server, the edge client does not use it directly but enters the personalized fine-tuning stage (step four). In this stage, client-specific data reflecting its specific network environment is used to refine the model's training. An early stopping mechanism is used to monitor the training process; training automatically stops when the validation set loss no longer decreases after three consecutive rounds, effectively preventing overfitting. After fine-tuning, to meet the stringent requirements of embedded devices for computational efficiency and resource consumption, the model undergoes lightweighting and optimization. First, the model accuracy is converted from FP32 to INT8 using a quantization tool, reducing the model size by approximately 75% and significantly improving inference speed. The optimized model is integrated into an RKNN-based inference engine, achieving real-time state awareness and precise control, with an end-to-end inference latency of less than 10ms. After model deployment, the system establishes a continuous monitoring mechanism to collect the inference error metric Δe=|y in real time. pred -y actualWhen the moving average error exceeds the preset threshold, the system will automatically trigger the model recalibration process to ensure that the control accuracy is always maintained at a high level.

[0060] The embodiments described in this specification are merely examples of implementations of the inventive concept and are for illustrative purposes only. The scope of protection of this invention should not be considered limited to the specific forms described in these embodiments; rather, it extends to equivalent technical means conceived by those skilled in the art based on the inventive concept.

Claims

1. A collaborative training method for centrifugal circulating pump groups based on federated learning, characterized in that, The method includes the following steps: Step 1: Initialization of the federated learning system: In multiple heterogeneous pipe network environments, collect the status data of water pump operation to build an initial training set. The recurrent neural network model based on gated recurrent units is trained by the time series backpropagation algorithm. After training, a distributed learning architecture is built. The central server is configured with a federated learning coordination service based on the Flower framework, edge clients are configured, and differential privacy transmission encrypted communication based on the gRPC protocol is established. Step 2, Collaborative Training Process: The central server initializes a lightweight recurrent neural network model (RNN) as the global baseline model and distributes it to each edge client through a secure channel. Each client uses locally collected pump group operation data for local training and updates the local model parameters using the stochastic gradient descent algorithm. During training, Gaussian noise is added through differential privacy to protect data privacy. After training is completed, the model weight update amount ΔW is encrypted to ensure that the original data does not go out of domain. Step 3, Secure Aggregation and Model Update: The server uses a weighted average algorithm to securely aggregate the collected weight updates. After aggregation, the new model is distributed to all clients, completing one round of federated learning iteration. Step 4: Personalized Model Deployment: After receiving the global model, each edge client fine-tunes and trains the model based on its local unique data to adapt the model to the specific characteristics of the pipeline network environment. Step 5: Continuous optimization mechanism: Establish a model version management mechanism to record the contribution of each client and the model update history.

2. The centrifugal circulating pump group collaborative training method based on federated learning as described in claim 1, characterized in that, In step one, the forward propagation process is represented as follows: h t =GRU(x t ,h t-1 ) (1); Where x t = [Speed, Power, Flow Rate, Head] is the input feature sequence at time t, h t-1 It is in a hidden state; The loss function uses mean squared error: Optimize model parameters using gradient descent algorithm:

3. The method for collaborative training of centrifugal circulating pump groups based on federated learning as described in claim 1 or 2, characterized in that, In step one, the edge client is based on the Rockchip RK3566 platform and is equipped with a data acquisition module and a local storage unit.

4. The centrifugal circulating pump group collaborative training method based on federated learning as described in claim 3, characterized in that, In step two, the stochastic gradient descent algorithm is used to update the local model parameters: Where B represents mini-batch data and η represents the learning rate.

5. The centrifugal circulating pump group collaborative training method based on federated learning as described in claim 4, characterized in that, In step three, the secure aggregation calculation is as follows: Where n k Let n be the data volume of the k-th client, and n be the total data volume. The weight update amount is given for the k-th client, followed by updating the global model parameters.

6. The method for collaborative training of centrifugal circulating pump groups based on federated learning as described in claim 5, characterized in that, In step four, an early stopping mechanism is used during fine-tuning training to prevent overfitting. Training stops when the validation set loss no longer decreases after three consecutive rounds. The optimized model is deployed to the local inference engine through the following steps: First, the FP32 model is converted to INT8 precision using a model quantization tool to reduce the model size and improve inference speed, while keeping the accuracy drop within an acceptable range. Finally, it is integrated into a real-time inference framework based on the RKNN inference engine to achieve inference latency less than a set threshold. After deployment, the system establishes a continuous monitoring mechanism to collect the inference error metric Δe=|y in real time. pred -y actual When the moving average error exceeds the threshold, model recalibration is automatically triggered.

7. The centrifugal circulating pump group collaborative training method based on federated learning as described in claim 5, characterized in that, In step five, the formula for evaluating client contribution is: c k =(n k ×ΔL k ) / t k (6); Where n k Let ΔL be the data volume of the k-th client. k To represent the performance improvement of this client-side model, t k This is the last update time; Configure a periodic federated training task to perform global aggregation at set intervals: An adaptive threshold mechanism is adopted to automatically trigger the retraining process when an anomaly is detected, ensuring continuous system optimization and stable operation.

8. The method for collaborative training of centrifugal circulating pump groups based on federated learning as described in claim 7, characterized in that, In step five, an anomaly refers to a decrease in model performance ΔL > δ or a change in pipeline impedance characteristics exceeding 15%, where δ = μ + 2σ, μ is the average performance change, and σ is the standard deviation.

Images