High-speed railway station air conditioning system performance degradation perception and on-demand maintenance method and device

Abstract

The application provides a high-speed railway station air conditioning system performance degradation perception and on-demand maintenance method and device, relates to the technical field of high-speed railway station air conditioning systems, and comprises the following steps: determining a perception component of the high-speed railway station air conditioning system; based on the degradation mechanism of each perception component, corresponding equivalent degradation features are respectively constructed, the equivalent degradation features of each perception component are combined into a degradation equivalent feature vector; a health baseline model is established, the working condition and control features at the current time are input into the health baseline model, and an equivalent feature estimation value in a healthy state is generated; the residual features of the degradation equivalent feature vector and the equivalent feature estimation value in the healthy state are calculated, and the residual features are taken as performance degradation parameters of the high-speed railway station air conditioning system. The application can enable the high-speed railway station air conditioning system to run efficiently for a long time under an on-demand maintenance strategy, and can also reduce the maintenance cost of the high-speed railway station air conditioning system.

Description

Technical Field

[0001] This invention relates to the field of air conditioning system technology for high-speed railway stations, and more specifically, to a method and apparatus for sensing and maintaining the performance degradation of air conditioning systems in high-speed railway stations. Background Technology

[0002] High-speed railway station air conditioning systems operate under complex conditions of high humidity and heat disturbances, strong passenger flow fluctuations, and large-space air and water distribution losses. System performance degradation is easily masked by short-term compensation through control redundancy, leading to delayed maintenance response, increased energy consumption, and deterioration of comfort. Compared to conventional buildings, high-speed railway stations experience more frequent load fluctuations and more difficult-to-uniformly cover airflow, meaning that air conditioning system performance degradation is no longer a simple component issue but manifests as an overall efficiency loss through the coupling of the cooling source, water system, air system, and terminal environment. Currently, there is a lack of methods for detecting and condition-based maintenance of high-speed railway station air conditioning systems, making it difficult to implement effective maintenance strategies in a timely manner during system performance degradation. Summary of the Invention

[0003] The purpose of this invention is to provide a method and apparatus for sensing and maintaining the performance degradation of air conditioning systems in high-speed railway stations, in order to improve the aforementioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows: Firstly, this application provides a method for sensing the performance degradation of a high-speed railway station air conditioning system, including: Identify the sensing components of the high-speed railway station air conditioning system; Based on the degradation mechanism of each sensing component, corresponding equivalent degradation features are constructed respectively, and the equivalent degradation features of each sensing component are combined into a degradation equivalent feature vector. Establish a health baseline model, input the current operating conditions and control characteristics into the health baseline model, and generate equivalent characteristic estimates under the healthy state; Calculate the residual features between the degraded equivalent feature vector and the equivalent feature estimate under healthy conditions, and use the residual features as the performance degradation parameter of the high-speed railway station air conditioning system.

[0004] Secondly, this application also provides a method for condition-based maintenance of air conditioning systems in high-speed railway stations, including: A digital twin model of the air conditioning system of a high-speed railway station is established, and the digital twin model is used to simulate various operating conditions and various fault types to generate a multivariate time series dataset. Performance degradation parameters are constructed based on the data in the multivariate time series dataset, and running parameters are extracted from the multivariate time series dataset. The running parameters and the performance degradation parameters are used as candidate variables, and input features are selected from the candidate variables. A lifetime prediction model is constructed based on the input features, and the remaining lifetime is output using the lifetime prediction model. Using the remaining service life and operating condition information as the state space, the controllable adjustment parameters of the air conditioning system as the action space, and the weighted negative sum of the comprehensive costs as the cumulative reward, a reinforcement learning decision framework is constructed. The reinforcement learning decision framework is trained so that the trained decision framework selects actions from the action space based on the current state space to generate control and maintenance strategies, thereby maximizing the cumulative reward.

[0005] Thirdly, this application also provides a high-speed railway station air conditioning system performance degradation sensing device, comprising: The determination module is used to determine the sensing components of the high-speed railway station air conditioning system; The construction module is used to construct corresponding equivalent degradation features based on the degradation mechanism of each sensing component, and the equivalent degradation features of each sensing component are combined into a degradation equivalent feature vector. A module is established to build a health baseline model. The current operating conditions and control characteristics are input into the health baseline model to generate equivalent feature estimates under the health state. The calculation module is used to calculate the residual features between the degraded equivalent feature vector and the equivalent feature estimate under the healthy state, and to use the residual features as the performance degradation parameters of the high-speed railway station air conditioning system.

[0006] Fourthly, this application also provides a condition-based maintenance device for a high-speed railway station air conditioning system, comprising: The twin model building module is used to build a digital twin model of the air conditioning system of a high-speed railway station. The digital twin model is used to simulate various operating conditions and various fault types to generate a multivariate time series dataset. The feature selection module is used to construct performance degradation parameters based on the data in the multivariate time series dataset, extract running parameters from the multivariate time series dataset, use the running parameters and the performance degradation parameters as candidate variables, and select input features from the candidate variables. The lifespan prediction module is used to construct a lifespan prediction model based on the input features and to output the remaining lifespan using the lifespan prediction model. The decision framework construction module is used to construct a reinforcement learning decision framework by taking the remaining service life and operating condition information as the state space, the controllable adjustment parameters of the air conditioning system as the action space, and the weighted negative sum of the comprehensive costs as the cumulative reward. The training module is used to train the reinforcement learning decision framework so that the trained decision framework selects actions from the action space according to the current state space to generate control and maintenance strategies, thereby maximizing the cumulative reward.

[0007] The beneficial effects of this invention are as follows: 1. This invention identifies the sensing components of a high-speed railway station air conditioning system and constructs equivalent degradation features. It transforms the physical degradation law of the sensing components into a unified degradation equivalent feature vector, and then combines it with a health baseline model to calculate residual features as performance degradation parameters. This allows the degradation state of the system to be separated from operating condition fluctuations and control redundancy, thus enabling the observation and quantification of the degree of degradation.

[0008] 2. This invention can accurately predict the remaining service life of an air conditioning system and combine the prediction results with operational condition information. Through a reinforcement learning decision-making framework, it autonomously optimizes the action space and maintenance strategies. While ensuring thermal comfort levels within the station, it significantly reduces the operating energy consumption of the air conditioning system and effectively slows down the performance degradation rate of sensing components such as compressors, fans, and water pumps, thereby extending the overall effective service life of the system.

[0009] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of the process for sensing the performance degradation of the air conditioning system in a high-speed railway station as described in this embodiment of the invention. Figure 2 This is a schematic diagram of the condition-based maintenance method for the air conditioning system of a high-speed railway station as described in this embodiment of the invention; Figure 3 This is a schematic diagram illustrating the Pearson correlation calculation results between candidate variables of the high-speed railway station air conditioning system and the proxy RUL as described in this embodiment of the invention. Figure 4 This is a schematic diagram of the mutual information ranking results between candidate variables and proxy RUL of the high-speed railway station air conditioning system in an embodiment of the present invention; Figure 5 This is a schematic diagram of the reinforcement learning interaction process described in an embodiment of the present invention; Figure 6This is a schematic diagram comparing the cumulative energy consumption of the air conditioning system under different strategies in an embodiment of the present invention; Figure 7 This is a comparison of air temperature changes inside the station under different strategies in the embodiments of the present invention; Figure 8 This is a schematic diagram of the performance degradation sensing device for the air conditioning system of a high-speed railway station as described in an embodiment of the present invention; Figure 9 This is a schematic diagram of the condition-based maintenance device for the air conditioning system of a high-speed railway station as described in an embodiment of the present invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0013] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance. Example 1

[0014] This embodiment provides a method for sensing the performance degradation of air conditioning systems in high-speed railway stations, which is applicable to the long-term continuous operation environment of air conditioning systems in high-speed railway stations.

[0015] See Figure 1 The figure illustrates this method: S1. Determine the sensing components of the high-speed railway station air conditioning system; Specifically, the sensing components in high-speed rail stations include compressors, heat exchangers (including evaporators and condensers), fans (including supply fans and return fans), water pumps, pipes (including air ducts and water pipes), and valves.

[0016] Based on the above embodiments, this method further includes: S2. Based on the degradation mechanism of each sensing component, construct corresponding equivalent degradation features respectively, and combine the equivalent degradation features of each sensing component into a degradation equivalent feature vector; Specifically, the equivalent degradation characteristics of compressors, fans, and water pumps are as follows: ; In the formula, express The comprehensive equivalent degradation characteristics at any given moment. The isentropic efficiency of the compressor. For fan efficiency, For water pump efficiency. This is the normalized weighted fusion operator. To improve the robustness of the representation, firstly... Scale alignment is performed in the healthy segment, and then the components are weighted and summed according to their sensitivity to system performance to form a normalized weighted fusion efficiency feature.

[0017] Specifically, the degradation of heat exchangers is mainly manifested in scaling and contamination of the heat exchange surface, leading to increased thermal resistance. The equivalent degradation characteristics of the heat exchanger are as follows: ; In the formula This refers to the equivalent degradation characteristics of the heat exchanger. This indicates the thermal resistance of the evaporator. SC ( t The ) represents the difference between the saturation temperature corresponding to the condensing pressure and the refrigerant temperature at the condenser outlet.

[0018] Specifically, the degradation of the pipeline typically manifests as an increase in pressure drop at the same flow rate or a decrease in available flow rate at the same driving capacity. Therefore, the equivalent resistance coefficient can be used to characterize the equivalent degradation features, specifically: ; In the formula, This refers to the instantaneous pressure drop across the two ends of the pipeline. This refers to the equivalent resistance coefficient of the pipeline system, i.e., the equivalent degradation characteristic. This represents the mass flow rate of the fluid inside the pipe.

[0019] Specifically, valve degradation often manifests as wear and jamming of the valve core and seat, control drift, and deviations in effective opening caused by condensation and icing. Its external manifestation is a change in the equivalent flow capacity under the same control command. Therefore, the equivalent flow coefficient is used to characterize the equivalent degradation features, specifically: ; In the formula, This refers to the mass flow rate of the fluid passing through the valve. The pressure difference across the valve. This is the valve's equivalent flow coefficient, i.e., its equivalent degradation characteristic.

[0020] The equivalent degradation features of each sensing component are combined to form a degradation equivalent feature vector: ; In the formula, Represents the degenerate equivalent eigenvector. These are the equivalent degradation features described above.

[0021] Based on the above embodiments, this method further includes: S3. Establish a health baseline model, input the current operating conditions and control characteristics into the health baseline model, and generate equivalent characteristic estimates under the healthy state; Specifically, step S3 includes: S31. Obtain historical operating data of the air conditioning system in a healthy state, wherein the historical operating data includes the equivalent degradation characteristics of each sensing component and the corresponding operating conditions and control characteristics; Specifically, when the air conditioning system is in a healthy state, a period of continuous historical data is collected; wherein, the operating conditions and control characteristics include at least the outdoor temperature and humidity, estimated indoor heat load, speed / valve position, and supply and return water temperature difference.

[0022] S32. A regression algorithm is used to train the historical operating data to construct a healthy baseline model with operating conditions and control characteristics as inputs and equivalent degradation characteristics as outputs; Specifically, the regression algorithm includes multiple linear regression, support vector regression, or random forest regression, etc. The regression algorithm is used to train the operating condition and control features and equivalent degradation features under healthy conditions, constructing a mapping relationship from operating condition and control features to equivalent degradation features: ; In the formula, This represents the estimated equivalent characteristics of the air conditioning system under current operating conditions and control characteristics, assuming it is in a healthy state. These are the characteristic vectors of operating conditions and control. This is a regression model.

[0023] S33. Collect the operating conditions and control characteristics at the current moment, input the operating conditions and control characteristics into the health baseline model, and obtain the equivalent characteristic estimate of the current health state.

[0024] Based on the above embodiments, this method further includes: S4. Calculate the residual features between the degraded equivalent feature vector and the equivalent feature estimate under healthy conditions, and use the residual features as the performance degradation parameter of the high-speed railway station air conditioning system; Specifically, the residual characteristics are: ; In the formula, This represents the residual characteristics, i.e., the performance degradation parameters. For degenerate equivalent feature vectors, This is an estimate of the equivalent characteristics under a healthy state.

[0025] When the air conditioning system is in good condition, the residuals mainly consist of measurement noise and modeling errors, and their values ​​fluctuate around zero. When the system experiences performance degradation, the residuals show a continuous positive or negative shift. Therefore, the residual characteristics can be used as performance degradation parameters for high-speed railway station air conditioning systems. Example 2

[0026] like Figure 2 As shown in the figure, this embodiment provides a condition-based maintenance method for the air conditioning system of a high-speed railway station, the method including: S1. Establish a digital twin model of the air conditioning system of a high-speed railway station, and use the digital twin model to simulate various operating conditions and various fault types to generate a multivariate time series dataset; Since it is difficult to obtain a large amount of fault data for actual high-speed railway station air conditioning systems, this embodiment constructs a digital twin model of the air conditioning system in simulation software to generate time-series data covering various operating conditions and fault types. The simulation software can be AMEsim. The digital twin model consists of five parts: station building and load, indoor terminals and chilled water side, chiller unit and cooling water side, control and data interface, and simulation output records.

[0027] In the station building and load module, the inputs are outdoor environmental parameters, changes in the number of people inside the station hall, and the heat load generated by equipment operation. The output is the station hall air temperature. T in The temperature serves as a feedback signal for subsequent control loops and a benchmark for comfort evaluation.

[0028] The indoor terminal units are coupled with the chilled water side modules and the building load modules. The heat exchange of the terminal equipment directly affects the temperature change in the station hall, while the heat load demand of the station hall is fed back to the evaporator inlet through the chilled water return temperature, forming a closed loop on the load side.

[0029] The refrigeration unit and the cooling water side module constitute a complete vapor compression refrigeration cycle, including a compressor, evaporator, condenser, and electronic expansion valve. The evaporator is responsible for cooling the chilled water, the condenser dissipates heat to the outdoor environment through the cooling water circuit, and the cooling water pump drives the cooling water to circulate between the condenser and the cooling tower.

[0030] The control and data interface module allows switching between rule-based control and reinforcement learning control strategies. In the default rule-based control mode, a PI (proportional-integral) controller is used to control the station hall temperature. T inThe compressor speed is adjusted based on the feedback signal; at the same time, a separate PI loop controls the opening of the electronic expansion valve to maintain the superheat at the evaporator outlet within the set range.

[0031] The simulation output recording module records the main variables generated during the simulation process at time steps, including the inlet and outlet pressures of the heat exchanger, the subcooling and superheat of the refrigerant, the refrigerant mass flow rate, the power of the compressor, fan and water pump, and the temperature difference between the chilled water supply and return water.

[0032] Specifically, fault modes and operating conditions are set for the digital twin model. During the operation of the digital twin model, several monitoring data are generated in chronological order. The monitoring data includes the equipment's operating parameters, operating conditions, and control characteristics.

[0033] Based on the above embodiments, this method further includes: S2. Construct performance degradation parameters based on the data in the multivariate time series dataset, extract running parameters from the multivariate time series dataset, use the running parameters and the performance degradation parameters as candidate variables, and select input features from the candidate variables; In this embodiment, based on the monitoring data, the degradation equivalent feature vector is calculated using the high-speed railway station air conditioning system performance degradation perception method described in Embodiment 1, and then the estimated value is obtained by calling the health baseline model. Finally, the difference is calculated to obtain the residual feature as the performance degradation parameter at the time point.

[0034] Specifically, step S2 includes: S21. Obtain the multivariate runtime time series in the multivariate time series dataset, and divide the multivariate runtime time series into training set, validation set and test set according to time order; Specifically, the first 70% of the data is used as the training set, the middle 15% as the validation set, and the last 15% as the test set.

[0035] S22. The training set is divided into multiple input sequences using a sliding window, wherein each input sequence corresponds to the remaining lifetime monitoring signal at the end of a window; Define the remaining lifetime of the agent as That is, a proxy RUL, which is constructed as follows: ; In the formula, This is the initial lifetime baseline value. For system uptime, For residual parameters.

[0036] Let the vector of monitoring data at time t be... ,in, Includes equipment operating parameters, working condition parameters, and performance degradation parameters. The vector dimension is .

[0037] A sliding window of length L is used to stack consecutive time points in the training set into an input sequence: ; Each input sequence The corresponding monitoring signal is the remaining lifetime of the agent at time t at the end of the window. This leads to a many-to-one sequence regression problem: ; In the formula, For the lifespan prediction model to be trained, These are the model parameters.

[0038] S23. The correlation between candidate variables and remaining useful life is analyzed using Pearson correlation coefficient and mutual information, and input features are selected based on the analysis results.

[0039] Because the recursive construction of the surrogate RUL inherently includes monotonic terms that degrade over time, there is a significant correlation between time and the surrogate RUL. If the time parameter is directly used as an input feature, the model may primarily rely on time information to fit the lifetime, thus weakening the learning of degradation signals during operation. Therefore, this embodiment uses Pearson correlation coefficient and mutual information to perform complementary analysis on the association between candidate variables and the surrogate RUL, thereby determining the final input feature set. The time parameter is not used as model input but only for sample indexing.

[0040] Specifically, the formula for calculating the sample Pearson correlation coefficient is as follows: ; In the formula, Indicate candidate variables and Pearson correlation coefficient, For the sample size, For sample index, The first The observed values ​​of two variables in a sample. These are the average values.

[0041] The formula for calculating mutual information is: ; In the formula, For the mutual information between variables X and Y, Let X be the joint probability distribution of variables X and Y. Let X and Y be the marginal probability distributions.

[0042] The Pearson correlation coefficient is used to measure the strength of linear association, and mutual information is used to measure the strength of nonlinear dependency. Combining both linear and nonlinear indicators, a value is selected that... Relevant variables with low redundancy among them constitute the final set of input features, such as... Figure 3 , Figure 4 As shown.

[0043] Based on the above embodiments, this method further includes: S3. Construct a lifetime prediction model based on the input features, and use the lifetime prediction model to output the remaining lifetime. Specifically, step S3 includes: S31. Construct a long short-term memory network and use the long short-term memory network to extract local fluctuation features in the input sequence; Specifically, the input sequence The input is a bidirectional long short-term memory network (BiLSTM). Preferably, the hidden layer dimension of the BiLSTM is set to 128, the number of layers is 1, and a bidirectional structure is adopted to capture the temporal dependencies of forward and backward directions.

[0044] Furthermore, performance degradation residual sequences are obtained from multivariate time series datasets. For performance degradation residual sequences Perform moving average decomposition to obtain low-frequency components. and high frequency components ,Right now .

[0045] S32. Construct a Transformer network and use the Transformer network to extract long-term degradation trend features from the input sequence; The model has a dimension of 128, 8 attention heads, and 3 encoder layers. The input sequence... The input is fed into the Transformer encoder to extract global dependencies and output a feature vector of long-term degradation trend.

[0046] Specifically, the predicted values ​​of high-frequency components Instead of being directly output from the network, it is reconstructed using a set of harmonic bases with preset periods. These preset periods include intraday periods (24 hours, 12 hours, 8 hours, 6 hours, 4 hours) and intraweek periods (168 hours, 84 hours). The harmonic base reconstruction formula is as follows: ; In the formula, For high-frequency components, For the number of cycles, For the first A preset cycle, and These are the coefficients for the adaptive output of the network.

[0047] Rewrite the above in vector form: ; in, This is the coefficient vector of the network prediction. is the harmonic basis vector at the current moment.

[0048] S33. The local fluctuation characteristics are fused with the long-term degradation trend characteristics to generate a predicted value of remaining useful life; Specifically, the local fluctuation feature vector output by BiLSTM and the long-term degradation trend feature vector output by Transformer are concatenated along the feature dimension to obtain a fused feature vector. This fused feature vector is then input into a fully connected layer, ultimately outputting the predicted remaining lifetime. ; In the formula, This is the predicted remaining useful life. This is the initial lifetime baseline value. For system uptime, To predict low-frequency components, This represents the estimated high-frequency components reconstructed using harmonic bases.

[0049] Specifically, the model is trained using the training set and validation set, and an early stopping mechanism is used during the training process: if the validation set loss does not improve for 12 consecutive rounds, the training is stopped and the model parameters with the minimum validation set loss are saved.

[0050] The loss function is a weighted average of low-frequency component loss and high-frequency component loss. The low-frequency component uses mean squared error (MSE) loss to track the drift of the degradation trend; the high-frequency component uses amplitude-weighted mean squared error, with larger amplitude moments assigned higher weights to enhance the learning of periods of violent fluctuations.

[0051] On the test set, the model performance was evaluated using mean absolute error (MAE), root mean square error (RMSE), and mean relative error (MAPE).

[0052] Based on the above embodiments, this method further includes: S4. Using the remaining service life and operating condition information as the state space, the controllable adjustment parameters of the air conditioning system as the action space, and the weighted negative sum of the comprehensive costs as the cumulative reward, a reinforcement learning decision framework is constructed. Specifically, the state space is used to describe the environment in which the reinforcement learning agent is located at time t. All necessary information. This embodiment organizes the state space into six types of parameters: (1) Deterioration / Health Category: Predicted Remaining Useful Life Entropy efficiency of compressors, etc., serve as direct basis for lifespan-friendly regulation; (2) Cyclic operating conditions: such as condenser side temperature / pressure, evaporator side inlet temperature, evaporator outlet pressure, subcooling, refrigerant mass flow rate, etc., used to characterize the operating point of the thermodynamic cycle; (3) Indoor environment: Indoor temperature, indoor relative humidity, etc., are the core observations for comfort targets; (4) Supply air side: supply air temperature, fan head, effective air supply rate, etc., which are key links connecting the action and the terminal response; (5) Energy load and external disturbances: outdoor temperature and humidity, number of passengers, heat load estimation, etc., to provide the external disturbance context required for strategy generalization; (6) Dynamic supplementary item: Action at the previous moment With key increments ; Combine all the above parameters into a multidimensional state vector. ,in, For the original observation subset, This is the predicted remaining useful life. The action in the previous moment. The difference is the subset of the original observations.

[0053] Specifically, the action space refers to the adjustment commands that the intelligent agent can output to the various actuators of the air conditioning system. In this embodiment, the controllable adjustment parameters are defined as a continuous action space, and the action numbers included are shown in Table 1 below: Table 1

[0054] To prevent the strategy output from being adjusted too drastically, upper and lower limits are set for each action variable, and the rate of change of action at adjacent time points is constrained to simulate the inertia of the actual actuator and reduce the additional stress on the equipment.

[0055] In this embodiment, the goal of reinforcement learning is to maximize the long-term cumulative reward. This embodiment defines the cumulative reward as the weighted negative sum of the overall cost, that is, to achieve multi-objective optimization by minimizing the overall cost: Specifically, define the comprehensive cost in one step. The comprehensive cost is a weighted sum of energy consumption cost, thermal comfort cost, and degradation cost. The opposite number as an instant reward ; The cumulative return is defined as the discount sum, specifically expressed as: ; In the formula, For a moment Cumulative returns As a discount factor, Discount factor of Power of 1 For a moment The instant reward received.

[0056] Furthermore, construct a Markov decision process: ; In the formula, Representing the state space, For a moment The system can be used for information representation in decision-making; Represents the action space, The control output of the intelligent agent; This represents the state transition law, used to describe the state transition of a system under the combined influence of actions and external disturbances. Evolved to The probability mechanism; This is an immediate reward function used to map the positive or negative consequences of the decision made in this step into scalar feedback; This is a discount factor used to weigh immediate returns against future returns.

[0057] The interactive process of reinforcement learning is as follows Figure 5 As shown. Since the action space is a continuous variable, this embodiment uses the Flexible Action Evaluation (SAC) algorithm for training. The SAC algorithm introduces a maximum entropy objective, maximizing the policy entropy while optimizing the cumulative reward, to enhance exploratory nature and improve robustness. Its objective function is: ; In the formula, Represents the mathematical expectation. Discount factor of Power of 1 For a moment Instant rewards received For temperature coefficient, For policy entropy, Let be the policy function, representing the state. The probability distribution of taking an action.

[0058] Thus, the condition-based maintenance problem of high-speed railway station air conditioning systems has been fully formalized into an optimization problem that can be solved within a reinforcement learning framework.

[0059] Based on the above embodiments, this method further includes: S5. Train the reinforcement learning decision framework so that the trained decision framework selects actions from the action space according to the current state space to generate control and maintenance strategies, thereby maximizing the cumulative reward; Specifically, step S5 includes: S51. Initialize the policy network, value network, and experience replay pool of the reinforcement learning decision framework; The value network includes a soft Q-function and a soft state value function, as shown below: ; In the formula, It is a soft Q-function, representing the state. Take action below Then, according to the strategy The expected cumulative reward that can be obtained from the interaction For the current action; To follow the strategy Calculate the expectation of the joint trajectory of the environmental state transition. For instant reward function, As a discount factor, For the next state Soft state value function; For state The soft state value function, Indicates the state Below, regarding the action By strategy Calculate the expected value of the distribution. For temperature coefficient, This is the strategy function.

[0060] S52. In each training round, select an action from the action space according to the current state space, execute the action, and then update the environment state; Specifically, the training hyperparameter settings are shown in Table 2: Table 2

[0061] In each training round, the agent selects an action from the action space based on the current state information, and after executing the action, the digital twin model updates the environment state.

[0062] S53. Obtain immediate feedback from the environment, combine the current state space, the selected action, the immediate feedback, and the updated state space into an experience sample, and store it in the experience replay pool; Specifically, an immediate reward is returned after an action is performed, thus forming an experience sample. ,in, These represent the current state, the selected action, the immediate reward, and the next state, respectively. The agent stores experience samples in the experience replay pool.

[0063] S54. Randomly sample experience samples from the experience replay pool, and update the parameters of the policy network and the value network using the Flexible Action Evaluation Algorithm (SAC) until the cumulative reward converges; Among them, the SAC algorithm is a deep reinforcement learning algorithm for continuous action space. It encourages the agent to achieve a balance between exploration and exploitation by maximizing the weighted sum of cumulative reward and policy entropy.

[0064] This embodiment uses the SAC algorithm to update the parameters of the policy network and the value network until the cumulative reward is reached. Convergence. Once training is complete, the optimal policy can be obtained. .

[0065] The trained decision framework outputs a sequence of control actions based on the current state to achieve real-time control of the air conditioning system. At the same time, when the predicted remaining service life is lower than a preset threshold, the framework can trigger a maintenance reminder. The threshold can be set by maintenance personnel according to actual needs.

[0066] The trained decision-making framework is deployed into a digital twin model, using real high-speed railway station outdoor environmental parameters and passenger flow data as input sequences for closed-loop simulation evaluation. Compared with traditional PID rule control, the reinforcement learning control strategy proposed in this embodiment significantly reduces cumulative energy consumption throughout the entire cycle while improving thermal comfort.

[0067] like Figure 6 As shown, under maintenance-free conditions, the cumulative energy consumption of the reinforcement learning control strategy is 4.22 × 10⁻⁶. 4 kWh, compared to 5.45 × 10 kWh for PID rule control. 4 kWh decreased by 22.6%.

[0068] like Figure 7 As shown, the average comfort deviation over the entire lifecycle of the reinforcement learning control strategy is 0.82, while the average comfort deviation under PID rule control is 0.89, a reduction of 7.9%. Example 3

[0069] like Figure 8 As shown, this embodiment provides a performance degradation sensing device for a high-speed railway station air conditioning system, comprising: The determination module is used to determine the sensing components of the high-speed railway station air conditioning system; The construction module is used to construct corresponding equivalent degradation features based on the degradation mechanism of each sensing component, and the equivalent degradation features of each sensing component are combined into a degradation equivalent feature vector. A module is established to build a health baseline model. The current operating conditions and control characteristics are input into the health baseline model to generate equivalent feature estimates under the health state. The calculation module is used to calculate the residual features between the degraded equivalent feature vector and the equivalent feature estimate under the healthy state, and to use the residual features as the performance degradation parameters of the high-speed railway station air conditioning system.

[0070] Based on the above embodiments, the establishment module includes: The acquisition unit is used to acquire historical operating data of the air conditioning system in a healthy state. The historical operating data includes the equivalent degradation characteristics of each sensing component and the corresponding operating conditions and control characteristics. The training unit is used to train the historical operating data using a regression algorithm to construct a healthy baseline model with operating conditions and control characteristics as inputs and equivalent degradation characteristics as outputs. The data acquisition unit is used to acquire the current operating conditions and control characteristics, input the operating conditions and control characteristics into the health baseline model, and obtain the equivalent characteristic estimate of the current health state. It should be noted that the specific methods by which each module performs its operation in the apparatus described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated upon here. Example 4

[0071] like Figure 9 As shown, this embodiment provides a condition-based maintenance device for a high-speed railway station air conditioning system, including: The twin model building module is used to build a digital twin model of the air conditioning system of a high-speed railway station. The digital twin model is used to simulate various operating conditions and various fault types to generate a multivariate time series dataset. The feature selection module is used to construct performance degradation parameters based on the data in the multivariate time series dataset, extract running parameters from the multivariate time series dataset, use the running parameters and the performance degradation parameters as candidate variables, and select input features from the candidate variables. The lifespan prediction module is used to construct a lifespan prediction model based on the input features and to output the remaining lifespan using the lifespan prediction model. The decision framework construction module is used to construct a reinforcement learning decision framework by taking the remaining service life and operating condition information as the state space, the controllable adjustment parameters of the air conditioning system as the action space, and the weighted negative sum of the comprehensive costs as the cumulative reward. The training module is used to train the reinforcement learning decision framework so that the trained decision framework selects actions from the action space according to the current state space to generate control and maintenance strategies, thereby maximizing the cumulative reward.

[0072] Based on the above embodiments, the feature filtering module includes: A partitioning unit is used to obtain the multivariate runtime time series in the multivariate time series dataset and divide the multivariate runtime time series into training set, validation set and test set according to time order. A sliding window unit is used to divide the training set into multiple input sequences using a sliding window, wherein each input sequence corresponds to a remaining lifetime monitoring signal at the end of a window; The analysis and screening unit is used to analyze the correlation between candidate variables and remaining useful life using Pearson correlation coefficient and mutual information, and to screen input features based on the analysis results.

[0073] Based on the above embodiments, the lifetime prediction module includes: Long Short-Term Memory (LSTM) units are used to construct LSM networks and extract local fluctuation features from input sequences using these networks. Transformer unit, used to construct Transformer network, and use the Transformer network to extract long-term degradation trend features from input sequence; The fusion unit is used to fuse the local fluctuation features with the long-term degradation trend features to generate a predicted value of remaining useful life.

[0074] Based on the above embodiments, the training module includes: An initialization unit is used to initialize the policy network, value network, and experience replay pool of the reinforcement learning decision framework. An action execution unit is used to select an action from the action space according to the current state space in each training round, execute the action, and then update the environment state. An experience storage unit is used to obtain immediate feedback from the environment, combine the current state space, the selected action, the immediate feedback, and the updated state space into an experience sample, and store it in the experience replay pool. The parameter update unit is used to randomly sample experience samples from the experience replay pool and update the parameters of the policy network and the value network using a flexible action evaluation algorithm until the cumulative reward converges.

[0075] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0076] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims. A health baseline model is established, and the current operating conditions and control characteristics are input into the health baseline model to generate equivalent characteristic estimates under healthy conditions.

Claims

1. A method for sensing the performance degradation of a high-speed railway station air conditioning system, characterized in that, include: Identify the sensing components of the high-speed railway station air conditioning system; Based on the degradation mechanism of each sensing component, corresponding equivalent degradation features are constructed respectively, and the equivalent degradation features of each sensing component are combined into a degradation equivalent feature vector. Establish a health baseline model, input the current operating conditions and control characteristics into the health baseline model, and generate equivalent characteristic estimates under the healthy state; Equivalent feature estimate Calculate the residual features between the degraded equivalent feature vector and the equivalent feature estimate under healthy conditions, and use the residual features as the performance degradation parameter of the high-speed railway station air conditioning system.

2. The method for sensing performance degradation of a high-speed railway station air conditioning system according to claim 1, characterized in that, The current operating conditions and control characteristics are input into the health baseline model to generate equivalent feature estimates under the healthy state, including: Acquire historical operating data of the air conditioning system in a healthy state, the historical operating data including the equivalent degradation characteristics of each sensing component and the corresponding operating conditions and control characteristics; A regression algorithm is used to train the historical operating data to construct a healthy baseline model with operating conditions and control characteristics as inputs and equivalent degradation characteristics as outputs. The operating conditions and control characteristics at the current moment are collected, and the operating conditions and control characteristics are input into the health baseline model to obtain the equivalent characteristic estimate of the health state at the current moment.

3. A method for condition-based maintenance of a high-speed railway station air conditioning system, characterized in that, include: A digital twin model of the air conditioning system of a high-speed railway station is established, and the digital twin model is used to simulate various operating conditions and various fault types to generate a multivariate time series dataset. Performance degradation parameters are constructed based on the data in the multivariate time series dataset, and running parameters are extracted from the multivariate time series dataset. The running parameters and the performance degradation parameters are used as candidate variables, and input features are selected from the candidate variables. A lifetime prediction model is constructed based on the input features, and the remaining lifetime is output using the lifetime prediction model. Using the remaining service life and operating condition information as the state space, the controllable adjustment parameters of the air conditioning system as the action space, and the weighted negative sum of the comprehensive costs as the cumulative reward, a reinforcement learning decision framework is constructed. The reinforcement learning decision framework is trained so that the trained decision framework selects actions from the action space based on the current state space to generate control and maintenance strategies, thereby maximizing the cumulative reward.

4. The method for condition-based maintenance of high-speed railway station air conditioning systems according to claim 3, characterized in that, Performance degradation parameters are constructed based on the data in the multivariate time-series dataset, and running parameters are extracted from the multivariate time-series dataset. The running parameters and the performance degradation parameters are used as candidate variables, and input features are selected from the candidate variables, including: Obtain the multivariate runtime time series from the multivariate time series dataset, and divide the multivariate runtime time series into training set, validation set and test set according to time order; The training set is divided into multiple input sequences using a sliding window, wherein each input sequence corresponds to the remaining lifetime monitoring signal at the end of a window; Pearson correlation coefficient and mutual information were used to analyze the association between candidate variables and remaining useful life, and input features were selected based on the analysis results.

5. The method for condition-based maintenance of high-speed railway station air conditioning systems according to claim 4, characterized in that, A lifetime prediction model is constructed based on the input features, and the remaining lifetime is output using the lifetime prediction model, including: Construct a long short-term memory network and use the long short-term memory network to extract local fluctuation features from the input sequence; Construct a Transformer network and use the Transformer network to extract long-term degradation trend features from the input sequence; The local fluctuation characteristics are fused with the long-term degradation trend characteristics to generate a predicted value for the remaining useful life.

6. The method for condition-based maintenance of high-speed railway station air conditioning systems according to claim 3, characterized in that, Training the reinforcement learning decision framework includes: Initialize the policy network, value network, and experience replay pool of the reinforcement learning decision framework; In each training round, an action is selected from the action space based on the current state space, and the environment state is updated after the action is executed. Obtain immediate feedback from the environment, combine the current state space, the selected action, the immediate feedback, and the updated state space into an experience sample, and store it in the experience replay pool. Experience samples are randomly sampled from the experience replay pool, and the parameters of the policy network and the value network are updated using a flexible action evaluation algorithm until the cumulative reward converges.

7. A performance degradation sensing device for a high-speed railway station air conditioning system, characterized in that, include: The determination module is used to determine the sensing components of the high-speed railway station air conditioning system; The construction module is used to construct corresponding equivalent degradation features based on the degradation mechanism of each sensing component, and the equivalent degradation features of each sensing component are combined into a degradation equivalent feature vector. A module is established to build a health baseline model. The current operating conditions and control characteristics are input into the health baseline model to generate equivalent feature estimates under the health state. The calculation module is used to calculate the residual features between the degraded equivalent feature vector and the equivalent feature estimate under the healthy state, and to use the residual features as the performance degradation parameters of the high-speed railway station air conditioning system.

8. The high-speed railway station air conditioning system performance degradation sensing device according to claim 7, characterized in that, The establishment module includes: The acquisition unit is used to acquire historical operating data of the air conditioning system in a healthy state. The historical operating data includes the equivalent degradation characteristics of each sensing component and the corresponding operating conditions and control characteristics. The training unit is used to train the historical operating data using a regression algorithm to construct a healthy baseline model with operating conditions and control characteristics as inputs and equivalent degradation characteristics as outputs. The acquisition unit is used to acquire the operating conditions and control characteristics at the current moment, input the operating conditions and control characteristics into the health baseline model, and obtain the equivalent feature estimate of the health state at the current moment.

9. A condition-based maintenance device for a high-speed railway station air conditioning system, characterized in that, include: The twin model building module is used to build a digital twin model of the air conditioning system of a high-speed railway station. The digital twin model is used to simulate various operating conditions and various fault types to generate a multivariate time series dataset. The feature selection module is used to construct performance degradation parameters based on the data in the multivariate time series dataset, extract running parameters from the multivariate time series dataset, use the running parameters and the performance degradation parameters as candidate variables, and select input features from the candidate variables. The lifespan prediction module is used to construct a lifespan prediction model based on the input features and to output the remaining lifespan using the lifespan prediction model. The decision framework construction module is used to construct a reinforcement learning decision framework by taking the remaining service life and operating condition information as the state space, the controllable adjustment parameters of the air conditioning system as the action space, and the weighted negative sum of the comprehensive costs as the cumulative reward. The training module is used to train the reinforcement learning decision framework so that the trained decision framework selects actions from the action space according to the current state space to generate control and maintenance strategies, thereby maximizing the cumulative reward.

10. The condition-based maintenance device for the air conditioning system of a high-speed railway station according to claim 9, characterized in that, The feature filtering module includes: A partitioning unit is used to obtain the multivariate runtime time series in the multivariate time series dataset and divide the multivariate runtime time series into training set, validation set and test set according to time order. A sliding window unit is used to divide the training set into multiple input sequences using a sliding window, wherein each input sequence corresponds to a remaining lifetime monitoring signal at the end of a window; The analysis and screening unit is used to analyze the correlation between candidate variables and remaining useful life using Pearson correlation coefficient and mutual information, and to screen input features based on the analysis results.

Images