Multi-device flexible parameter self-tuning method based on transfer learning

Abstract

The invention discloses a multi-device flexible parameter self-tuning method based on transfer learning, and belongs to the technical field of multi-device parameter tuning, and the method comprises the following steps: S1, data preprocessing and feature screening; s2, aligning the feature space of the target equipment; s3, carrying out transfer learning model initialization and knowledge transfer; s4, performing online adaptive parameter optimization; and S5, multi-device parameter self-tuning circulation and model updating. According to the multi-device flexible parameter self-tuning method based on transfer learning, source device knowledge is reused through transfer learning, data dependence of target devices is reduced, and the cross-device transfer effect is improved; and convergence is accelerated and initial parameter sensitivity is reduced through an optimization strategy, so that efficient self-tuning of multi-device parameters is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of multi-device parameter setting, and particularly relates to a multi-device flexible parameter self-setting method based on transfer learning. BACKGROUND

[0002] Multi-device parameter setting refers to a technical process of determining and adjusting control parameters of devices in a scene with multiple devices, so as to realize accurate operation of the devices and improve performance (such as control accuracy and operation efficiency).

[0003] However, the prior art has the following disadvantages: a traditional data-driven method relies on a large amount of data of a target device, and has low setting accuracy when data accumulation is insufficient in an initial stage, and lacks effective use of common knowledge of the devices; a simple transfer learning method does not consider feature space differences of different devices, and has poor cross-device transfer effect and mismatched knowledge and the target device; and a single reinforcement learning method has slow convergence speed and sensitive performance to initial parameters.

[0004] Therefore, a new method is urgently needed. SUMMARY

[0005] The application aims to provide a multi-device flexible parameter self-setting method based on transfer learning, which reuses source device knowledge through transfer learning, reduces data dependence of a target device, improves cross-device transfer effect, and realizes efficient self-setting of multi-device parameters through optimization strategies to speed up convergence and reduce initial parameter sensitivity.

[0006] To achieve the above-mentioned purpose, the application provides a multi-device flexible parameter self-setting method based on transfer learning, comprising the following steps:

[0007] S1: Collect source device data in a privacy-protected manner through a federated learning driven distributed architecture to obtain multi-modal data; perform feature extraction, anomaly detection, denoising and normalization processing on the multi-modal data to generate uniform and high-quality source device feature vectors, thereby laying a foundation for cross-device knowledge transfer; wherein the multi-modal data includes vibration, temperature, image and text log of the source device; at the same time, collect initial data of a target device to obtain target device data; and generate target device features through feature extraction of the target device data;

[0008] S2: Map the target device features in S1 to source device features through GAN network composed of a generator and a discriminator through adversarial training to obtain aligned features aligned with the target device feature distribution;

[0009] S3, constructing a transfer learning model, generating initial parameters by combining pre-training and fine-tuning strategies using the multi-modal data processed in S1 and the target device data;

[0010] S4, constructing a state space according to the initial parameters in S3, combining device real-time running data, and outputting optimized parameters through a Q-learning algorithm; at the same time, accumulating the target device running data generated in the optimization process;

[0011] S5, according to the optimized parameters and target device running data output by S4, constructing a cyclic mechanism of data acquisition, feature extraction and alignment, parameter dynamic optimization and model iterative update; after triggering the update, the new data is fed back to S1 and S3.

[0012] Preferably, in S2, the structure of the generator is represented as: the input layer contains 100 nodes, the first hidden layer contains 256 nodes, the second hidden layer contains 128 nodes, and the output layer contains 100 nodes, using the Tanh activation function; the structure of the discriminator is represented as: the input layer contains 100 nodes, the first hidden layer contains 256 nodes, the second hidden layer contains 128 nodes, and the output layer contains 1 node, using the Sigmoid activation function.

[0013] Preferably, in S3, the structure of the transfer learning model is represented as: the first hidden layer contains 200 nodes, the second hidden layer contains 150 nodes, and the third output layer contains 3 nodes; each layer uses the ReLU activation function, and the input dimension of the model is 100.

[0014] Preferably, in S4, the Q-learning algorithm includes defining a state space, an action space and a reward function, wherein the state space includes PID parameters, vibration, temperature, control accuracy and running efficiency.

[0015] Preferably, the calculation formula of the reward function is:

[0016] Q(s t ,a t )=Q(s t ,a t )+α[r t +γmax a Q(s t+1 ,a)-Q(s t -a t )];

[0017] Wherein, α is the learning rate, set to 0.1; γ is the discount factor, set to 0.9; r t is the immediate reward obtained by the current action; max a Q(st+1 a) is the maximum expected reward in the next state.

[0018] The application also provides a multi-device flexible parameter self-tuning system based on transfer learning, comprising:

[0019] A multi-modal data extraction module is configured to obtain multi-modal data, perform feature extraction, anomaly detection, denoising and normalization processing on the multi-modal data, and generate a unified and high-quality source device feature vector, thereby laying a foundation for cross-device knowledge transfer.

[0020] A cross-device feature space alignment module is connected to the multi-modal data extraction module and configured to construct a GAN network composed of a generator and a discriminator, map the target device feature to the source device feature through adversarial training, and obtain an aligned feature that is aligned with the target device feature distribution.

[0021] A transfer learning model module is connected to the cross-device feature space alignment module and configured to generate initial parameters by using the processed multi-modal data and the target device data through a strategy combining pre-training and fine-tuning.

[0022] A parameter adaptive optimization module is connected to the transfer learning model module and configured to construct a state space by combining device real-time running data, output optimized parameters through a Q-learning algorithm, and accumulate target device running data generated in the optimization process.

[0023] A multi-device parameter tuning closed-loop iteration module is connected to the parameter adaptive optimization module and configured to construct a cyclic mechanism of data acquisition, feature extraction and alignment, parameter dynamic optimization and model iteration update according to the optimized parameters and the target device running data.

[0024] Therefore, the application adopts the above-mentioned multi-device flexible parameter self-tuning method based on transfer learning.

[0025] (1) The application reuses source device historical data through transfer learning, and the target device only needs a small amount of initial data to complete parameter optimization, without relying on a large amount of target device data, thereby solving the problem of low precision caused by insufficient data in the initial stage.

[0026] (2) The application realizes the feature space alignment of the target device and the source device through a GAN (Generative Adversarial Network), combines GTN gradient domain migration regularization technology to retain the common knowledge of the source device and the characteristics of the target device, improves the cross-device migration effect, and solves the problem of low migration efficiency caused by hardware and environmental differences;

[0027] (3) The application constructs a Q-learning reinforcement learning feedback mechanism to form a "prediction-verification-optimization" closed loop with fast convergence speed; the real-time reward function dynamically adjusts the parameters to reduce the sensitivity to the initial parameters, solving the problems of slow convergence and large initial parameter influence.

[0028] The technical solutions of the application will be further described in detail below with the aid of the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 The flowchart of the embodiment of the application is shown in the figure.

[0030] Figure 2 The flowchart of the parameter adaptive optimization module of the embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0031] To make the purpose, technical solutions and advantages of the embodiments of the application clearer, the technical solutions in the embodiments of the application will be described clearly and completely below with the aid of the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the application, rather than all the embodiments of the application. All other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the application. Unless otherwise defined, the technical terms or scientific terms used in the application should be understood as the usual meanings understood by those of ordinary skill in the art.

[0032] Embodiment one

[0033] As shown in the figure, the application is a multi-device flexible parameter self-tuning method based on transfer learning, which comprises the following steps: Figure 1

[0034] S1, realize the privacy-protected collection of source device data through a federated learning driven distributed architecture, obtain multi-modal data, perform feature extraction, anomaly detection, denoising and normalization processing on the multi-modal data, generate uniform and high-quality source device feature vectors, and lay a foundation for cross-device knowledge transfer; wherein the multi-modal data includes vibration, temperature, image, text log and the like of the source device; at the same time, collect the initial data of the target device to obtain target device data; the target device data is generated into target device features after feature extraction; ​

[0035] In this step, first, a federated learning driven distributed collection architecture is adopted, a lightweight collection module is deployed at the source device end, and cross-device feature collaborative extraction is realized through a homomorphic encryption protocol to avoid raw data transmission and protect data privacy;

[0036] Then, a three-dimensional sampling weight model is constructed based on hardware similarity (weight 0.4), task relevance (0.3), and data timeliness (0.3), and the sampling weight is determined through softmax normalization;

[0037] Subsequently, a cross-modal feature encoder is constructed, in which a CNN-Transformer hybrid network is used to extract 100-dimensional time-frequency domain features from the vibration signal, MobileNetV3 is used to extract spatial features from the image data, and BERT embedding layer is used to convert the text log into a semantic vector. Finally, the multi-modal features are fused through an attention mechanism to generate a unified 100-dimensional feature vector;

[0038] Then, an isolated forest and local outlier factor hybrid framework is used for anomaly detection, and a three-layer deep denoising autoencoder (DAE) is used for feature denoising. The loss function is set as MSE and KL divergence, and the iteration round is 5000 rounds;

[0039] Then, cross-device normalization is performed. First, the features are mapped to the standard normal distribution quantile space through quantile normalization, and then the GAN driven adversarial domain adaptation is used to realize feature distribution alignment, with an iteration round of 5000 rounds;

[0040] S2, through the construction of a GAN network composed of a generator and a discriminator, the target device features in S1 are mapped to the source device features through adversarial training, obtaining aligned features that are aligned with the target device feature distribution, solving the problem of feature mismatch caused by hardware and environmental differences;

[0041] In this step, a 3-layer fully connected network generator G and discriminator D are first constructed. The generator G is a 3-layer fully connected network, and the structure is represented as: the input layer contains 100 nodes, the first hidden layer contains 256 nodes, the second hidden layer contains 128 nodes, and the output layer contains 100 nodes, using the Tanh activation function. The discriminator D is a 3-layer fully connected network, and the structure is represented as: the input layer contains 100 nodes, the first hidden layer contains 256 nodes, the second hidden layer contains 128 nodes, and the output layer contains 1 node, using the Sigmoid activation function;

[0042] Then, through adversarial training, the generator G maps the target device features X t to G(X t which is close to the source device features, while the discriminator D distinguishes between the real source device features X s and the generated G(Xt );

[0043] The Adam optimizer was used for training, with a learning rate of 0.0002 to control the step size of parameter updates. The exponential decay rate of the first-order moment estimation was beta1 = 0.5, and the training iterations were 5000. Through 5000 rounds of adversarial training between the generator and the discriminator, the generator was able to identify the target device features X. t Convert to an aligned feature X consistent with the feature distribution of the source device. t' Ultimately, this achieves feature space alignment between the target device and the source device;

[0044] This addresses the issue of feature distribution mismatch caused by differences in device hardware and operating environments, providing a foundation for subsequent cross-device knowledge transfer.

[0045] S3. Construct a transfer learning model. Use the processed multimodal data and target device data from S1 to generate initial parameters through a strategy that combines pre-training and fine-tuning, so as to achieve efficient transfer of knowledge from the source device to the target device.

[0046] In this step, a three-layer fully connected network transfer learning model is first constructed. The structure of the transfer learning model is as follows: the first hidden layer contains 200 nodes, the second hidden layer contains 150 nodes, and the third output layer contains 3 nodes. Each layer uses the ReLU activation function, and the input dimension of the model is 100.

[0047] Next, a strategy combining pre-training and fine-tuning was adopted: first, the aligned source device data was used for 200 rounds of pre-training with mean squared error loss, and then a small amount of aligned target device data was used for fine-tuning for 50 rounds to generate initial parameters;

[0048] S4, such as Figure 2 As shown, based on the initial parameters in S3, a state space is constructed by combining the real-time operating data of the equipment. The optimized parameters are output through the Q-learning algorithm to realize online dynamic optimization of equipment parameters and improve adaptability to changes in operating conditions. At the same time, the target equipment operating data generated during the optimization process is accumulated.

[0049] In this step, a 20-dimensional state space S is first defined, which contains key information about the equipment operation, such as PID parameters, data collected by sensors (such as vibration, temperature, etc.), and performance indicators (such as control accuracy, operating efficiency, etc.), to describe the current operating state of the equipment; and during the parameter optimization process, the operating data of the target equipment is continuously collected and accumulated.

[0050] The action space A is defined to include 9 parameter adjustment actions, specifically adjusting the proportional (P), integral (I), and derivative (D) parameters of the PID by ±5%, or keeping the current parameters unchanged, which are the specific operations that the agent can take.

[0051] Define the reward function R to evaluate the pros and cons of actions, give +10 rewards when the device running accuracy meets the standard, and give-5 penalties when the running appears abnormal, so as to guide the algorithm to learn to the optimization goal;

[0052] Then, based on the Bellman equation, the action value function Q(s, a) (Q value) is updated iteratively, Q(s, a) represents the expected cumulative reward obtained after taking action a in state s, and the calculation formula is as follows:

[0053] Q(s t ,a t )=Q(s t ,a t )+α[r t +γmax a Q(s t+1 ,a)-Q(s t -a t )];

[0054] Wherein, α is the learning rate, set to 0.1, used to control the Q value update amplitude; γ is the discount factor, set to 0.9, used to measure the weight of future rewards, emphasizing recent rewards while taking into account long-term gains; r t is the immediate reward obtained by the current action; max a Q(s t+1 ,a) is the maximum expected reward in the next state;

[0055] This embodiment adopts the greedy strategy with initial ε=0.9 and decay rate 0.995, to balance the attempt of new actions to find better strategies and the selection of known high reward actions;

[0056] Through obtaining the current device state S, selecting actions from A based on the ε-greedy strategy, and updating the Q value table according to the device performance feedback R using the Bellman equation, the closed loop continues to iterate, so that the algorithm gradually learns the optimal parameter adjustment strategy, realizes the real-time response of device parameters to working condition changes, and finally improves the control accuracy and running stability;

[0057] S5, according to the optimized parameters and target device running data output by S4, a cycle mechanism of data acquisition, feature extraction and alignment, parameter dynamic optimization and model iterative update is constructed, so that the multi-device parameter self-tuning process can be continuously iteratively optimized, and then the adaptability of the system to different devices and various working conditions is enhanced;

[0058] In this step, the initial data collected by the target device is first transmitted to the feature extraction module of S1, and 100-dimensional features are extracted through CNN; the extracted features are transmitted to S2, and the feature space is aligned through GAN to generate aligned features X t' ; X t'The migration learning model of S3 is transmitted, and the model generates initial parameters based on this; the initial parameters and the real-time state of the equipment jointly construct the state space S of S4, which is iteratively optimized by Q-learning until the performance meets the standard, and new data generated in the optimization process is continuously accumulated;

[0059] When the continuously accumulated new data reaches a threshold, model updating is triggered: the new data is retransmitted to S1 for feature extraction and screening, and to S3 for fine-tuning of the migration model again, forming a cycle of data acquisition, feature extraction and alignment, dynamic optimization of parameters, and iterative updating of the model, realizing the continuous circulation and reuse of data in the whole process.

[0060] Therefore, the application adopts the above-mentioned multi-device flexible parameter self-tuning method based on migration learning, which reuses the knowledge of the source device through migration learning, reduces the dependence on target device data, and improves the cross-device migration effect; and through the optimization strategy, it accelerates the convergence and reduces the sensitivity of the initial parameters, realizing efficient self-tuning of multi-device parameters.

[0061] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0062] Finally, it should be pointed out that: the above embodiments are only used to illustrate the technical solutions of the present application and not to limit them, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can still be modified or replaced by equivalents, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A multi-device flexible parameter self-tuning method based on transfer learning, characterized in that, The method comprises the following steps: S1, privacy protection collection of source device data is realized through a federated learning driven distributed architecture, and multi-modal data is obtained; feature extraction, anomaly detection, denoising and normalization processing are performed on the multi-modal data, and a unified and high-quality source device feature vector is generated, thereby laying a foundation for cross-device knowledge transfer; wherein the multi-modal data comprises vibration, temperature, image and text log of the source device; at the same time, initial data of a target device is collected, and target device data is obtained; the target device data is subjected to feature extraction to generate target device features; S2, a GAN network composed of a generator and a discriminator is constructed, and the target device features in S1 are mapped to the source device features through adversarial training, thereby obtaining aligned features aligned with the target device feature distribution; S3, a transfer learning model is constructed, and an initial parameter is generated by using the multi-modal data processed in S1 and the target device data through a pre-training and fine-tuning combined strategy; S4, according to the initial parameter in S3, a state space is constructed in combination with real-time running data of the device, and an optimized parameter is output through a Q-learning algorithm; at the same time, target device running data generated in the optimization process is accumulated; S5, according to the optimized parameter and the target device running data output by S4, a cyclic mechanism of data collection, feature extraction and alignment, parameter dynamic optimization and model iterative updating is constructed; after triggering the update, new data is fed back to S1 and S3.

2. The method of claim 1, wherein, In S2, the structure of the generator is represented as: the input layer comprises 100 nodes, the first hidden layer comprises 256 nodes, the second hidden layer comprises 128 nodes, and the output layer comprises 100 nodes, and a Tanh activation function is used; the structure of the discriminator is represented as: the input layer comprises 100 nodes, the first hidden layer comprises 256 nodes, the second hidden layer comprises 128 nodes, and the output layer comprises 1 node, and a Sigmoid activation function is used.

3. The method of claim 1, wherein, In S3, the structure of the transfer learning model is represented as: the first hidden layer comprises 200 nodes, the second hidden layer comprises 150 nodes, and the third output layer comprises 3 nodes; ReLU activation functions are used in each layer, and the input dimension of the model is 100.

4. The method of claim 1, wherein, In S4, the Q-learning algorithm comprises defining a state space, an action space and a reward function, wherein the state space comprises PID parameters, vibration, temperature, control accuracy and running efficiency.

5. The method of claim 4, wherein, The calculation formula of the reward function is: Q(s t ,a t ) = Q(s t ,a t ) + a[r t + g max a Q(s t+1 ,a) - Q(s t ,a t )] where a is the learning rate, set to 0.1; g is the discount factor, set to 0.9; r t is the immediate reward obtained for the current action; max a Q(s t+1 ,a) is the maximum expected reward in the next state.

6. A multi-device flexible parameter self-tuning system based on transfer learning, applied to the multi-device flexible parameter self-tuning method based on transfer learning in any one of claims 1-5, characterized in that, It comprises: a multi-modal data extraction module for obtaining multi-modal data; feature extraction, anomaly detection, denoising and normalization processing are performed on the multi-modal data, and a unified and high-quality source device feature vector is generated, thereby laying a foundation for cross-device knowledge transfer; wherein the multi-modal data comprises vibration, temperature, image and text log of the source device; at the same time, initial data of a target device is collected, and target device data is obtained; the target device data is subjected to feature extraction to generate target device features; The cross-device feature space alignment module is connected with the multi-modal data extraction module, is configured to construct a GAN network composed of a generator and a discriminator, and is configured to map target device features to source device features through adversarial training to obtain aligned features aligned with a target device feature distribution; The transfer learning model module is connected with the cross-device feature space alignment module, is configured to generate initial parameters by using processed multi-modal data and target device data through a strategy combining pre-training and fine-tuning, and is configured to output the initial parameters. The parameter adaptive optimization module is connected with the transfer learning model module, is configured to construct a state space by combining device real-time running data, and is configured to output optimized parameters through a Q-learning algorithm; meanwhile, the parameter adaptive optimization module is configured to accumulate target device running data generated in the optimization process. The multi-device parameter setting closed-loop iteration module is connected with the parameter adaptive optimization module, is configured to construct a cyclic mechanism of data acquisition, feature extraction and alignment, parameter dynamic optimization, and model iteration update according to the optimized parameters and the target device running data, and is configured to feed new data to the multi-modal data extraction module and the transfer learning model module after triggering update.

Images