Fault diagnosis method based on hydropower industry internet platform and dynamic fusion engine

Abstract

The invention belongs to the field of fault diagnosis, and discloses a fault diagnosis method based on a hydroelectric industry internet platform and a dynamic fusion engine, and the method comprises the steps: obtaining data for fault diagnosis through the hydroelectric industry internet platform, and extracting the feature data in the data; constructing a multi-strategy fault diagnosis model library, and fusing output results of different fault diagnosis models by using a dynamic fusion engine to obtain a fault diagnosis result; and optimizing and adjusting the dynamic fusion engine according to the user feedback information. Through collaborative innovation of three dimensions of data feature engineering, a model integration strategy and a dynamic optimization algorithm, an intelligent diagnosis system with an autonomous evolution capability is constructed, remarkable technical breakthrough is formed in the aspects of reliability, adaptability, interpretability and the like, the fault diagnosis accuracy is effectively improved, and the fault diagnosis efficiency is improved. The method is especially suitable for predictive maintenance scenes of complex electromechanical equipment in the hydropower industry.

Description

Technical Field

[0001] The present invention belongs to the field of fault diagnosis technology, and specifically relates to a fault diagnosis method based on a hydropower industrial Internet platform and a dynamic fusion engine. Background Art

[0002] In the hydropower industry, various faults are prone to occur due to the complexity of hydropower equipment and systems and the variability of the operating environment. Therefore, it is necessary to establish an efficient and accurate fault diagnosis system to ensure the normal operation and safety of hydropower stations.

[0003] However, existing fault diagnosis methods, particularly those based on physical models and rules, are typically trained and optimized for specific equipment or operating conditions. This significantly reduces the effectiveness of these methods when the equipment or environment changes. Furthermore, data-driven fault diagnosis methods, such as machine learning models, often suffer from overfitting during the training phase, resulting in poor performance in new or previously unseen fault scenarios. Although fault diagnosis is often combined with on-site inspections in practice, this is inefficient and highly dependent on the expertise of personnel, making it prone to diagnostic errors. Summary of the Invention

[0004] The purpose of the present invention is to address the above-mentioned problems and provide a fault diagnosis method based on a hydropower industrial Internet platform and a dynamic fusion engine. The method obtains data used for fault diagnosis through the hydropower industrial Internet platform, constructs a multi-strategy fault diagnosis model library including statistical models, machine learning models and deep learning models, and uses a dynamic fusion engine to fuse the output results of different fault diagnosis models to obtain fault diagnosis results; and optimizes and adjusts the dynamic fusion engine based on user feedback on the fault diagnosis results, thereby establishing a fault diagnosis system with strong adaptability, strong generalization ability and high accuracy. Relying on the high interconnectedness and scalability of the industrial Internet, the method can efficiently process large amounts of data and realize online real-time fault detection.

[0005] In order to achieve the above object, the technical solution provided by the present invention is: The fault diagnosis method based on the hydropower industrial Internet platform and the dynamic fusion engine includes the following steps: Step 1: Obtain data for fault diagnosis through the hydropower industrial Internet platform, preprocess the acquired data, and extract feature data; Step 2: Establish a multi-strategy fault diagnosis model library, use the characteristic data as input data, and use different fault diagnosis models to obtain preliminary fault diagnosis results; Step 3: Build a dynamic fusion engine based on the deep reinforcement learning model, and use the dynamic fusion engine to fuse the preliminary fault diagnosis results of different fault diagnosis models to obtain the final diagnosis result; Step 4: Integrate with the Industrial Internet platform to display the final diagnosis results to users and collect user feedback information; Step 5: Optimize and adjust the dynamic fusion engine based on user feedback.

[0006] Preferably, in step 1, the preprocessing includes: processing continuous data using the Z-score standardization method; processing categorical data using the Min-Max standardization method, and converting the data into the range of [0, 1].

[0007] Preferably, in step 1, the extracting of feature data specifically includes: using one-hot encoding to extract features for categorical data, and converting categorical variables into binary matrices; for continuous data or multidimensional sensor data, using a wavelet transform algorithm to extract features.

[0008] Preferably, in step 2, the multi-strategy fault diagnosis model library includes a Bayesian classifier, a logistic regression model, a support vector machine, a sparse autoencoder, a convolutional neural network and a recursive neural network.

[0009] Preferably, in step 3, the dynamic fusion engine uses a TwinDelayed Deep Deterministic Policy Gradient (TD3) algorithm to determine the fault diagnosis model selected in a specific scenario and the weights of the diagnosis results of each fault diagnosis model. The action of TD3 is to select the fault diagnosis model and the corresponding diagnosis result weights.

[0010] After performing the selected action, a reward is obtained to measure the diagnostic performance after selecting the action; during the TD3 algorithm training process, the reward corresponding to the action is determined by manual labeling; The TD3 algorithm uses a dual Q network to determine the reward of an action and a target policy network to calculate the next optimal action. Given a state and action, the policy network outputs the expected reward for this action in this state. The policy network is a neural network used to approximate the policy. Given a state, the policy network outputs an action or a probability distribution of an action. The policy network selects the action to maximize the reward predicted by the double Q network. The goal of the policy network is to select actions that maximize the Q value. During training, the output of the Q network is used to guide the update of the policy network. When the policy network selects a good action, that is, an action with a high Q value, the policy network is rewarded and strengthens this action; on the contrary, when the policy network selects a poor action, that is, an action with a low Q value, the policy network is punished.

[0011] In step 5, the weight adjustment factor of the diagnosis model is calculated based on the diagnosis evaluation index E and the user feedback score. ; In the formula is the weight adjustment factor of the diagnostic model with the largest weight, which is used to determine the adjustment range of the weight; F is the user feedback score, ranging from [1,5]; If E is less than the evaluation index threshold or F is less than 3, the diagnostic model weight is adjusted; the new diagnostic model weight The calculation formula is: ; in, is the current diagnostic model weight, and is the predetermined diagnostic model weight adjustment amount; Using new diagnostic model weights Update the TD3 algorithm and apply it in the next round of diagnosis tasks.

[0012] The system of the above fault diagnosis method includes: Data acquisition and transmission layer: Communicates with the hydropower industrial Internet platform to obtain data for fault diagnosis, pre-processes the acquired data, and extracts feature data; Fault diagnosis model library: This includes Bayesian classifiers, logistic regression models, support vector machines, sparse autoencoders, convolutional neural networks, and recurrent neural networks. It is used to provide different fault diagnosis models for the dynamic fusion engine. These fault diagnosis models generate preliminary fault diagnosis results based on data provided by the hydropower industrial internet platform. Data processing and dynamic fusion layer: Use the dynamic fusion engine to fuse the preliminary fault diagnosis results of different fault diagnosis models to obtain the final diagnosis results; User interaction and feedback layer: visualize the final diagnosis results through the user interface and collect user feedback information; optimize and adjust the dynamic fusion engine based on user feedback information.

[0013] Compared with the prior art, the present invention has the following beneficial effects: 1) Through collaborative innovation in three dimensions: data feature engineering, model integration strategy, and dynamic optimization algorithm, this invention constructs an intelligent diagnostic system with autonomous evolution capabilities. This system achieves significant technological breakthroughs in reliability, adaptability, and interpretability, effectively improving the accuracy of fault diagnosis. It is particularly suitable for predictive maintenance scenarios of complex electromechanical equipment in the hydropower industry.

[0014] 2) The present invention constructs a fault diagnosis model library that includes Bayesian classifiers, logistic regression models, machine learning models, and deep learning models, which comprehensively covers and effectively responds to various industrial failure modes. Through a multi-strategy parallel diagnosis mechanism, the fault diagnosis accuracy is greatly improved compared with a single diagnosis model, and the false alarm rate in complex fault scenarios is effectively reduced.

[0015] 3) The present invention adopts a deep reinforcement learning framework to construct a dynamic fusion engine and realizes online optimization of model weights through Q network, so that the present invention has working condition adaptability; the innovative design of weight adjustment factors organically combines objective evaluation indicators with subjective feedback, thereby improving optimization efficiency compared with traditional expert systems.

[0016] 4) This invention adopts a real-time closed-loop mechanism of "diagnosis-feedback-optimization" to establish a diagnostic credibility correction model through user feedback information; it uses an incremental weight update algorithm to achieve dynamic evolution of model parameters, which can ensure the continuous improvement of diagnostic accuracy.

[0017] 5) This invention adopts a differentiated preprocessing method for continuous data: Z-score normalization (preserving data distribution characteristics) and Min-Max normalization (ensuring consistency in numerical range) for categorical data, effectively solving the normalization problem of multi-source heterogeneous data in industrial scenarios. Multi-source and multi-dimensional data provides a strong guarantee for subsequent accurate fault diagnosis.

[0018] 6) Through deep integration with the Industrial Internet platform, the present invention has achieved technological breakthroughs and synergistic efficiency in the entire fault diagnosis process. Through docking with the Industrial Internet platform, the real-time data of equipment such as turbines, generators, and transformers are synchronized with the SCADA system and MIS system in seconds, thereby improving the coverage of diagnostic data; combined with the Industrial Internet platform, the fault diagnosis results are visualized and mapped to the digital twin objects of the Industrial Internet platform, so that operation and maintenance personnel can quickly locate the source of the fault; a breakthrough in the real-time online diagnosis is achieved, and the dynamic fusion engine supports 50ms-level diagnostic response and transmits the diagnostic results back to the Industrial Internet platform, which can meet the real-time decision-making and processing needs of key scenarios such as load shedding and overspeed protection of hydropower units. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The present invention will be further described below with reference to the accompanying drawings and examples.

[0020] Figure 1 Schematic diagram of a fault diagnosis system according to an embodiment of the present invention.

[0021] Figure 2 Schematic diagram of a fault diagnosis method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0022] like Figure 1 and Figure 2 As shown in FIG, the fault diagnosis method based on the hydropower industrial Internet platform and the dynamic fusion engine includes the following steps: Step 1: Obtain data for fault diagnosis through the hydropower industrial Internet platform, preprocess the acquired data, and extract feature data.

[0023] In one embodiment, data from a sensor cluster is acquired through an industrial internet platform. The sensor cluster includes water level sensors, water flow velocity sensors, water quality sensors, bearing temperature sensors, and current / voltage sensors. The sensors monitor the operating status of turbines, generators, and valves in a hydropower station in real time. The data acquisition frequency varies depending on the sensor type and actual needs. An edge computing device includes an industrial-grade embedded computing module / board, a local area network interface, and a data preprocessing module / software module. These modules filter and preprocess data, store short-term data, and directly analyze short-term data in the event of a network outage or analysis delay. A gateway device includes an industrial-grade network switch, a network firewall, and a data encryption module. The gateway device receives data from various edge devices and uploads it to a cloud or data center. The sensor cluster is connected to the industrial-grade embedded computing module / board via wired or wireless means. The edge computing device performs data preprocessing and filtering. The edge computing device connects to the gateway device via a deployed local area network and receives remote commands from the cloud or data center, including commands to open and close valves and adjust generator speed. The gateway device then uploads data to the cloud or data center, or receives control commands from the cloud and transmits them to the edge computing device or directly controls the sensor devices.

[0024] In the embodiment, the industrial Internet platform is connected to the SCADA system and the MIS system to synchronously obtain data from the SCADA system and the MIS system.

[0025] First, the data obtained from the Industrial Internet platform is decrypted and decompressed, followed by outliers. Outlier identification can be performed using different methods depending on the actual data. For example, for data such as temperature, current, voltage, and water pressure, threshold-based methods are used to identify outliers. Data above or below a certain threshold is empirically determined to be an outlier. For data such as water flow and power, a sliding window-based method is used. A time window, such as 5 minutes, is set and the data is observed. The average flow rate and standard deviation within the time window are calculated. If the current data deviates from the average by more than a certain number of standard deviations, it is marked as an outlier. For data such as pressure and humidity that should normally follow a normal distribution, statistical methods can be used to calculate the average and standard deviation of the data. Data points are compared to the average and standard deviation, and any data points that deviate significantly from the average are marked as outliers. For other data, such as multidimensional sensor data and complex system monitoring data, clustering algorithms such as K-means can be used to divide the data into different groups. Data points that fall far from all clusters are considered outliers. All outliers are marked to facilitate subsequent processing.

[0026] Next, the data is standardized to facilitate further processing by subsequent models, and different standardization processes are used for different types of data.

[0027] Continuous data, such as current, voltage, temperature, pressure, or other multidimensional sensor data, are processed using the Z-score normalization method. This can preserve the distribution of the data and enhance the fault detection capability of the model. The final output data will be centered on 0 with a standard deviation of 1. For categorical data, such as device switch status, device type, operating mode, etc., the Min-Max normalization method is used to convert the data into a fixed range such as [0,1].

[0028] Feature extraction algorithms, such as time series analysis and frequency domain analysis, are used to extract feature data from the real-time sensor data. Feature data can include average, median, standard deviation, trend information, frequency domain features, etc.

[0029] For categorical data, one-hot encoding is used for feature extraction, and the categorical variables are converted into binary matrices; for continuous data or multidimensional sensor data, wavelet transform algorithm is used for feature extraction.

[0030] Taking current data as an example, the specific process of extracting feature data using the wavelet transform algorithm includes: Assume that the selected wavelet mother function is , perform wavelet decomposition on the current data to obtain the wavelet coefficients of each layer and approximation coefficients , ; in, is the approximation coefficient at the Jth level, is the wavelet coefficient at the jth layer, J is the number of decomposition layers, and t is the time; The specific decomposition formula is: ; ; in, is the wavelet mother function after scaling and translation; Analyze the wavelet coefficients, calculate the energy, mean and standard deviation of each frequency band and use them as feature data. ; ; ; In the formula 、 、 represent the energy, mean, and standard deviation of the frequency band, respectively.

[0031] Step 2: Establish a multi-strategy fault diagnosis model library, use the characteristic data as input data, and use different fault diagnosis models to obtain preliminary fault diagnosis results.

[0032] In an embodiment, the fault diagnosis model library includes statistical models (Bayesian classifier, logistic regression, etc.), machine learning models (random forest, support vector machine, etc.), deep learning models (sparse autoencoder, convolutional neural network, recursive neural network, etc.) and hydropower knowledge base, which are used to diagnose various types and levels of faults.

[0033] The Bayesian classifier determines the state category by calculating the posterior probability. ; Where, is the prior probability, is the likelihood probability, Features The marginal probability of Indicates the fault type.

[0034] The Bayesian classifier calculates the probability of each fault type and selects the fault type with the highest probability as the prediction result.

[0035] The logistic regression model calculates the probability of failure based on the characteristic data used for fault diagnosis; The objective function of the logistic regression model is the cross entropy loss: ; ; in, is the objective function of the logistic regression model, is the class label of the i-th sample point, is the feature vector of the i-th sample point, m represents the number of sample points, Is the sigmoid function, which maps the model output to [0,1]. are the parameters of the logistic regression model.

[0036] The logistic regression model calculates the probability of a fault occurring and compares it with a threshold value. When the probability of a fault occurring exceeds the threshold value, it is determined that a fault exists.

[0037] The support vector machine performs classification through a hyperplane, and the objective function of the support vector machine is: ; The constraints are: ; in, is the normal vector of the hyperplane, is a slack variable, which indicates the distance between the sample point and the hyperplane. is the regularization parameter that controls the trade-off between the width of the interval and the number of misclassified points; is the class label of the i-th sample point, is the eigenvector of the i-th sample point, and b is the intercept of the hyperplane.

[0038] The sparse autoencoder predicts the probability of each fault category through the cascaded Softmax classification layer. The objective function of the sparse autoencoder is: ; in, is the mean square error of the sample data; 、 They are encoding and decoding functions, is the i-th input sample.

[0039] A convolutional neural network or a recurrent neural network predicts the probability of each fault class through cascaded classification layers.

[0040] The objective function of the recurrent neural network is the cross entropy loss: ; in, represents the recurrent neural network output at time step t, is the i-th sample, is the true label of the i-th sample at time step t, are the model parameters of the recurrent neural network, T is the length of the sample time series.

[0041] Step 3: Use the TD3 algorithm to build a dynamic fusion engine, and use the dynamic fusion engine to fuse the preliminary fault diagnosis results of different fault diagnosis models to obtain the final diagnosis result.

[0042] The action of TD3 is to select the fault diagnosis model and the corresponding diagnosis result weight; After performing the selected action, a reward is obtained to measure the diagnostic performance after selecting the action; during the TD3 algorithm training process, the reward corresponding to the action is determined by manual labeling; The TD3 algorithm uses a dual Q network to determine the reward of an action and a target policy network to calculate the next optimal action. Given a state and action, the policy network outputs the expected reward for this action in this state. The policy network is a neural network used to approximate the policy. Given a state, the policy network outputs an action or a probability distribution of an action. The policy network selects the action to maximize the reward predicted by the double Q network. The goal of the policy network is to select actions that maximize the Q value. During training, the output of the Q network is used to guide the update of the policy network. When the policy network selects a good action, that is, an action with a high Q value, the policy network is rewarded and strengthens this action; on the contrary, when the policy network selects a poor action, that is, an action with a low Q value, the policy network is punished.

[0043] The TD3 algorithm first calculates the Q value of the next state and selects the minimum value to update the Q value of the current state. ; in, is the current state, is the next state, Is in state The action to be performed is is the immediate reward for the action, is the discount factor that determines the importance of future rewards, is the estimated value of the target Q network, is the target strategy network; Indicates that the status and strategies Next action The cumulative reward after Next, the TD3 algorithm updates the policy network to maximize the expected Q value: ; In the formula Indicates that the optimal strategy is in state Next, choose the action that maximizes the Q value; And update the target policy network to ensure the stability of learning: ; in, are the parameters of the current policy network, are the parameters of the target policy network, is the update coefficient, a decimal between 0 and 1, used to control the update rate.

[0044] For example, consider the following sensor data: water level 75.2 meters, water velocity in the diversion pipe 2.3 m / s, generator current 100 A, voltage 400 V, turbine bearing temperature 50°C, and turbine vibration amplitude 0.002 mm.

[0045] The features derived from the aforementioned data are: water level change rate of 0.01 m / min; analysis of the standard deviation of the current data over the past minute yields a current stability index of 1.5; and a Fourier transform of the vibration data reveals a significant frequency peak at 20 Hz. Combining this with other information, we can construct the state vector: State = {water level: 75.2 m, water flow rate: 2.3 m / s, current: 100 A, voltage: 400 V, bearing temperature: 50°C, vibration amplitude: 0.002 mm, water level change rate: 0.01 m / min, current stability: 1.5, frequency domain characteristics: 20 Hz, time: 8:00 AM, operating mode: full load}.

[0046] This state vector contains all the information about the environment, allowing the dynamic fusion engine to make appropriate decisions or actions.

[0047] Step 4: In conjunction with the Industrial Internet platform, display the final diagnostic results to users and collect user feedback information.

[0048] In the embodiment, the fault diagnosis results are mapped to the digital twin objects of the industrial Internet platform to facilitate grassroots operation and maintenance personnel to quickly locate the source of the fault; and user feedback information on the fault diagnosis results is collected.

[0049] Step 5: Optimize and adjust the dynamic fusion engine based on user feedback.

[0050] The weight adjustment factor of the diagnostic model is calculated based on the diagnostic evaluation index E and the user feedback score. ; In the formula is the weight adjustment factor of the diagnostic model with the largest weight, which is used to determine the adjustment range of the weight; F is the user feedback score, ranging from [1,5]; If E is less than the evaluation index threshold or F is less than 3, the diagnostic model weight is adjusted; the new diagnostic model weight The calculation formula is: ; in, is the current diagnostic model weight, and is the predetermined diagnostic model weight adjustment amount; Using new diagnostic model weights Update the TD3 algorithm and apply it in the next round of diagnosis tasks.

[0051] Example 2 like Figure 1 As shown, a fault diagnosis system is provided based on the method of embodiment 1, including: Data acquisition and transmission layer: Communicates with the hydropower industrial Internet platform to obtain data for fault diagnosis, pre-processes the acquired data, and extracts feature data; Fault diagnosis model library: This includes Bayesian classifiers, logistic regression models, support vector machines, sparse autoencoders, convolutional neural networks, and recurrent neural networks. It is used to provide different fault diagnosis models for the dynamic fusion engine. These fault diagnosis models generate preliminary fault diagnosis results based on data provided by the hydropower industrial internet platform. Data processing and dynamic fusion layer: Use the dynamic fusion engine to fuse the preliminary fault diagnosis results of different fault diagnosis models to obtain the final diagnosis results; User interaction and feedback layer: visualize the final diagnosis results through the user interface and collect user feedback information; optimize and adjust the dynamic fusion engine based on user feedback information.

[0052] In actual operation, the response speed of dynamic fusion and optimization adjustments may be affected by various factors. Another embodiment of the present invention makes some improvements based on the above architecture. For example, data compression technology is used in the data and storage stages to reduce data latency, and parallel computing is used in the cloud or data center to accelerate model inference and weight adjustment.

[0053] Accelerating model reasoning specifically includes: first parallelizing the model. For deep learning models, especially deep networks such as convolutional neural networks or recurrent neural networks, model parallelization technology can be used to decompose the model into multiple parts. Let the original model be M, and decompose it into n sub-models M1, M2...Mn. Then perform cloud-based distributed reasoning, using the distributed computing power of the cloud to assign each sub-model to different computing nodes. These nodes can be virtual machines, containers, or dedicated GPU instances. For an input x, each sub-model Mi performs reasoning in parallel on its corresponding node and outputs the result Finally, the output results of all sub-models are aggregated to a central node, and the data format conversion or merging needs to be designed to obtain the final output of the model. It may be the probability distribution of a category. These distributions can be simply averaged or combined using other methods to obtain the final classification result.

Claims

1. A fault diagnosis method based on the hydropower industry Internet platform and a dynamic fusion engine is characterized by: The following steps are involved: Step 1: Obtain data for fault diagnosis through the hydropower industrial Internet platform, preprocess the acquired data, and extract feature data; Step 2: Build a fault diagnosis model library, use the characteristic data as input data, and use different fault diagnosis models to obtain preliminary fault diagnosis results; Step 3: Use the dynamic fusion engine to fuse the preliminary fault diagnosis results of different fault diagnosis models to obtain the final diagnosis result; Step 4: Display the final diagnosis results to the user and collect user feedback; Step 5: Optimize and adjust the dynamic fusion engine based on user feedback.

2. The fault diagnosis method according to claim 1, characterized in that: In step 1, extracting the characteristic data specifically includes: For categorical data, one-hot encoding is used for feature extraction, converting categorical variables into binary matrices; for continuous data or multidimensional sensor data, wavelet transform algorithm is used for feature extraction; Taking current data as an example, the specific process of extracting feature data using the wavelet transform algorithm includes: Assume that the selected wavelet mother function is , perform wavelet decomposition on the current data to obtain the wavelet coefficients of each layer and approximation coefficients , ； in, is the approximation coefficient at the Jth level, is the wavelet coefficient at the jth layer, J is the number of decomposition layers, and t is the time; The specific decomposition formula is: ； ； in, is the wavelet mother function after scaling and translation; Analyze the wavelet coefficients, calculate the energy, mean and standard deviation of each frequency band and use them as feature data. ； ； ； In the formula 、 、 represent the energy, mean, and standard deviation of the frequency band, respectively.

3. The fault diagnosis method according to claim 2, characterized in that: In step 2, the fault diagnosis model library includes a Bayesian classifier, a logistic regression model, a support vector machine, a sparse autoencoder, and a recursive neural network.

4. The fault diagnosis method according to claim 3, characterized in that: In step 2, the Bayesian classifier determines the state category by calculating the posterior probability. ； Where, is the prior probability, is the likelihood probability, Features The marginal probability of Indicates the fault category; The Bayesian classifier calculates the probability of each fault type and selects the fault type with the highest probability as the prediction result.

5. The fault diagnosis method according to claim 3, characterized in that: In step 2, the logistic regression model calculates the fault probability based on the characteristic data used for fault diagnosis; The objective function of the logistic regression model is the cross entropy loss: ； ； in, is the objective function of the logistic regression model, is the class label of the i-th sample point, is the feature vector of the i-th sample point, m represents the number of sample points, is the sigmoid function, are the parameters of the logistic regression model; The logistic regression model calculates the probability of a fault occurring and compares it with a threshold value. When the probability of a fault occurring exceeds the threshold value, it is determined that a fault exists.

6. The fault diagnosis method according to claim 3, characterized in that: In step 2, the support vector machine performs classification through a hyperplane, and the objective function of the support vector machine is: ； The constraints are: ； in, is the normal vector of the hyperplane, is a slack variable, which indicates the distance between the sample point and the hyperplane. is the regularization parameter; is the class label of the i-th sample point, is the eigenvector of the i-th sample point, b is the intercept of the hyperplane, and m represents the number of sample points.

7. The fault diagnosis method according to claim 3, characterized in that: In step 2, the sparse autoencoder predicts the probability of each fault category through the cascaded Softmax classification layer. The objective function of the sparse autoencoder is: ； in, is the mean square error of the sample data; 、 They are encoding and decoding functions respectively, m represents the number of sample points, is the eigenvector of the i-th sample point.

8. The fault diagnosis method according to claim 1, characterized in that: In step 3, the dynamic fusion engine uses the double-delayed deep deterministic policy gradient algorithm TD3 to determine the fault diagnosis model selected in the specific scenario and the weight of the diagnosis results of each fault diagnosis model. The action of TD3 is to select the fault diagnosis model and the corresponding diagnosis result weight; After performing the selected action, a reward is obtained to measure the diagnostic performance after selecting the action; during the TD3 algorithm training process, the reward corresponding to the action is determined by manual labeling; The TD3 algorithm uses a dual Q network to determine the reward of an action and a target policy network to calculate the next optimal action. Given a state and action, the policy network outputs the expected reward for this action in this state. The goal of the policy network is to select actions that maximize the Q value. During training, the output of the Q network is used to guide the update of the policy network. When the policy network selects a good action, that is, an action with a high Q value, the policy network is rewarded and strengthens this action; on the contrary, when the policy network selects a poor action, that is, an action with a low Q value, the policy network is punished.

9. The fault diagnosis method according to claim 8, characterized in that: The TD3 algorithm first calculates the Q value of the next state and selects the minimum value to update the Q value of the current state. ； in, is the current state, is the next state, Is in state The action to be performed is is the immediate reward for the action, is the discount factor that determines the importance of future rewards, is the estimated value of the target Q network, is the target strategy network; Indicates that the status and strategies Next action The cumulative reward after Next, the TD3 algorithm updates the policy network to maximize the expected Q value: ； In the formula Indicates that the optimal strategy is in state Next, choose the action that maximizes the Q value; And update the target policy network to ensure the stability of learning: ； in, are the parameters of the current policy network, are the parameters of the target policy network, is the update coefficient.

10. The fault diagnosis method according to any one of claims 1 to 9, characterized in that: In step 5, the weight adjustment factor of the diagnosis model is calculated based on the diagnosis evaluation index E and the user feedback score. ； In the formula is the weight adjustment factor of the diagnostic model with the largest weight, which is used to determine the adjustment range of the weight; F is the user feedback score, ranging from [1,5]; If E is less than the evaluation index threshold or F is less than 3, the diagnostic model weight is adjusted; the new diagnostic model weight The calculation formula is: ； in, is the current diagnostic model weight, and is the predetermined diagnostic model weight adjustment amount; Using new diagnostic model weights Update the TD3 algorithm and apply it in the next round of diagnosis tasks.

Images