A rail transit equipment fault prediction method based on dynamic evolution double weight mechanism

Abstract

The application relates to the field of intelligent operation and maintenance of rail transit equipment, and discloses a rail transit equipment fault prediction method based on a dynamic evolution double-weight mechanism, which comprises the following steps: firstly, collecting multi-dimensional operation parameters of equipment to construct a monitoring data set containing dynamic and static parameters; secondly, determining the static weights of the parameters based on expert experience; then, calculating statistical characteristics of the parameters by using historical data, constructing a feature vector, and generating dynamic weights by using a trainable attention model; next, weighting and fusing the static and dynamic weights to obtain an equipment health index, and then realizing fault early warning; finally, realizing iterative updating of the weights by using a triggering mechanism to continuously optimize the evaluation model.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance of rail transit equipment, and in particular to the intelligent fault prediction function of rail transit equipment. Background Technology

[0002] With the rapid development of urban rail transit and the gradual expansion of the network scale in various regions, the number of equipment in various specialties within the network has also increased significantly, leading to a surge in equipment inspection / maintenance workload. Metro electromechanical equipment mainly consists of systems such as ventilation and air conditioning, water supply and drainage, low-voltage power distribution and lighting, environmental and equipment monitoring, escalators and platform screen doors, etc., and is crucial for ensuring the safe operation of the metro and providing passenger services. Metro electromechanical equipment involves many specialties, spans a wide range between them, has a large number of pieces of equipment, a large maintenance workload, is geographically dispersed, has many professional interfaces, operates in a harsh environment, and directly faces passengers. It is closely related to operational safety, passenger order, passenger service, and emergency rescue, making the task heavy and the responsibility significant.

[0003] Existing statistics show that electromechanical equipment failure is one of the major causes of subway traffic accidents and malfunctions, accounting for over 70% of all accidents and malfunctions. As the total number of subway electromechanical equipment increases, the number of devices experiencing failures also rises. Currently, the maintenance of subway electromechanical equipment mainly adopts a combined approach of planned maintenance, condition-based maintenance, and fault-based maintenance, with planned maintenance as the primary method and condition-based and fault-based maintenance as supplementary methods. However, the current operation and maintenance of electromechanical equipment faces the following difficulties: Poor visibility of actual equipment conditions hinders timely reflection of status changes, making it difficult to identify safety hazards and posing significant operational risks. There is a lack of clear guidelines for determining when and where repairs are needed, as well as identifying the most urgent repair areas, leading to a lack of emphasis on importance and urgency, resulting in a heavy workload and low efficiency. Planned inspections and maintenance require substantial investment of manpower, equipment, and spare parts, resulting in high costs. Equipment fault information cannot be obtained promptly and accurately; a large amount of data lacks specificity and fails to reflect the real-time status of the equipment.

[0004] Therefore, the existing market urgently needs to change the above situation, conduct accurate "real-time analysis" and "comprehensive evaluation" of equipment status, realize intelligent operation and maintenance of electromechanical equipment, and realize the transformation from traditional diagnosis based on phenomena and indicators to prediction based on intelligent systems. Electromechanical equipment can serve the safe operation of enterprises more effectively and reliably in networked operation and maintenance, and carry out full life cycle management of electromechanical equipment to achieve the project goal of lowest cost and highest overall efficiency throughout the entire equipment life cycle. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, the purpose of this invention is to utilize massive amounts of operation and maintenance data to predict equipment failures, thereby reducing maintenance costs, minimizing losses caused by equipment failures, and lowering the operating costs of rail transit.

[0006] To achieve the above objectives, the technical solution adopted by this invention is as follows: a fault prediction method for rail transit equipment based on a dynamic evolution dual-weight mechanism, comprising the following steps: Step 1: Collect multi-dimensional operating parameters of the equipment to determine the set of monitoring data that affect the health status of the equipment, including dynamic parameters and static parameters; Step 2: Determine the static weights based on the contribution of each parameter to the equipment function and prior knowledge; Step 3: Based on the equipment's historical operating data and maintenance records, calculate multidimensional statistical features for each monitoring parameter and construct a statistical feature vector; then, through a trainable attention model, obtain the dynamic weights of different statistical features. Step 4: Calculate the device health score by weighting and fusing the static and dynamic weights. Step 5: Compare the calculated health index with the warning threshold to complete the fault warning; Step 6: Set up a trigger mechanism to iteratively update the weights.

[0007] In a preferred embodiment of the present invention, the set of monitoring data X affecting the health status of the equipment in step one is determined as follows: , Where xi is the i-th evaluation parameter and n is the total number of parameters.

[0008] In a preferred embodiment of the present invention, in step two, static weights are determined based on the Delphi method according to the contribution of each parameter to the device function and prior knowledge. for: , in, S represents static.

[0009] In a preferred embodiment of the present invention, in step three, the statistical feature vector consists of three indicators: historical volatility, threshold deviation, and fault correlation, specifically formulated as follows: ; in This indicates the fluctuation characteristics of the normalized parameters. This represents the frequency characteristics of abnormal parameters after normalization. This represents the correlation characteristics between the normalized parameters and historical failure events; The i-th feature is the historical fluctuation feature. for: ; Where i represents the i-th feature, k represents the k-th historical sampling point, and m represents the number of samples within the sampling window. This represents the mean of the i-th feature within the historical window; After normalization:

[0010] The frequency characteristics of threshold deviation are:

[0011] in, This represents the normal interval for the i-th feature; After normalization:

[0012] The correlation coefficient of the fault correlation is:

[0013] in, Represents the time series of fault events. , This represents the mean value within the k-th historical window; After normalization: .

[0014] In a preferred embodiment of the present invention, in step three, a trainable attention model is used to obtain dynamic weight vectors with different statistical features, and the weight coefficients are:

[0015] Where Wa and ba are model parameters obtained by training with historical data; The loss function is:

[0016] in and The health index and reference health index calculated by the model. It is the attention weight of the i-th parameter at time t. It is the balance coefficient, and n is the total number of parameters.

[0017] In a preferred embodiment of the present invention, the dynamic weight in step four is:

[0018] Where i represents the i-th feature, and k represents the k-th historical sampling point. The weight represents the learning weight, and z represents the dynamic weight, which consists of three indicators.

[0019] In a preferred embodiment of the present invention, in step four, the static weights and dynamic weights are weighted and fused to calculate the device health score, and the final weights are:

[0020] in The coefficient is a fusion coefficient between 0 and 1, where i represents the i-th feature; The method for calculating the device health status is as follows: .

[0021] In a preferred embodiment of the present invention, in step five, the calculated health index is compared with the warning threshold to complete the fault warning. H greater than 0.8 indicates health, H less than 0.5 indicates fault warning, and the device is in a sub-healthy state if it is between the two.

[0022] In a preferred embodiment of the present invention, in step six, a trigger mechanism is set to iteratively update the weights, and steps three and four are repeated when an iteration period is set or a fault alarm occurs, so as to recalculate the dynamic weights and the comprehensive weights and realize online self-optimization of the weights.

[0023] As a preferred embodiment of the present invention, the trainable attention model is trained based on historical time series data, and the model structure adopts a feedforward neural network, a recurrent neural network, or a combination thereof.

[0024] Compared with existing technologies, the proposed scheme of this invention utilizes massive amounts of operation and maintenance data to predict equipment failures, thereby reducing maintenance costs, minimizing losses caused by equipment failures, and lowering the operating costs of rail transit. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating a fault prediction method for rail transit equipment based on a dynamic evolution dual-weight mechanism in this embodiment. Detailed Implementation

[0026] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention, but do not limit the scope of the present invention.

[0027] like Figure 1 As shown in the figure, a fault prediction method for rail transit equipment based on a dynamic evolution dual-weight mechanism in this embodiment includes the following steps: Step 1: Collect multi-dimensional operating parameters of the equipment to determine the set of monitoring data that affect the health status of the equipment, including dynamic parameters and static parameters; Step 2: Determine the static weights based on the contribution of each parameter to the equipment function and prior knowledge; Step 3: Based on the equipment's historical operating data and maintenance records, calculate multidimensional statistical features for each monitoring parameter and construct a statistical feature vector; then, through a trainable attention model, obtain the dynamic weights of different statistical features. Step 4: Calculate the device health score by weighting and fusing the static and dynamic weights. Step 5: Compare the calculated health index with the warning threshold to complete the fault warning; Step 6: Set up a trigger mechanism to iteratively update the weights.

[0028] In a preferred embodiment of the present invention, the set of monitoring data X affecting the health status of the equipment in step one is determined as follows: , Where xi is the i-th evaluation parameter and n is the total number of parameters.

[0029] In a preferred embodiment of the present invention, in step two, static weights are determined based on the Delphi method according to the contribution of each parameter to the device function and prior knowledge. for: , in, S represents static.

[0030] In a preferred embodiment of the present invention, in step three, the statistical feature vector consists of three indicators: historical volatility, threshold deviation, and fault correlation, specifically formulated as follows: ; in This indicates the fluctuation characteristics of the normalized parameters. This represents the frequency characteristics of abnormal parameters after normalization. This represents the correlation characteristics between the normalized parameters and historical failure events; The i-th feature is the historical fluctuation feature. for: ; Where i represents the i-th feature, k represents the k-th historical sampling point, and m represents the number of samples within the sampling window. This represents the mean of the i-th feature within the historical window; After normalization:

[0031] The frequency characteristics of threshold deviation are:

[0032] in, This represents the normal interval for the i-th feature; After normalization:

[0033] The correlation coefficient of the fault correlation is:

[0034] in, Represents the time series of fault events. , This represents the mean within the k-th historical window; after normalization: .

[0035] In a preferred embodiment of the present invention, in step three, a trainable attention model is used to obtain dynamic weight vectors with different statistical features, and the weight coefficients are:

[0036] Where Wa and ba are model parameters obtained by training with historical data; The loss function is:

[0037] in and The health index and reference health index calculated by the model. It is the attention weight of the i-th parameter at time t. It is the balance coefficient, and n is the total number of parameters.

[0038] In a preferred embodiment of the present invention, the dynamic weight in step four is:

[0039] Where i represents the i-th feature, and k represents the k-th historical sampling point. The weight represents the learning weight, and z represents the dynamic weight, which consists of three indicators.

[0040] In a preferred embodiment of the present invention, in step four, the static weights and dynamic weights are weighted and fused to calculate the device health score, and the final weights are:

[0041] in The coefficient is a fusion coefficient between 0 and 1, where i represents the i-th feature; The method for calculating the device health status is as follows: .

[0042] In a preferred embodiment of the present invention, in step five, the calculated health index is compared with the warning threshold to complete the fault warning. H greater than 0.8 indicates health, H less than 0.5 indicates fault warning, and the device is in a sub-healthy state if it is between the two.

[0043] In a preferred embodiment of the present invention, in step six, a trigger mechanism is set to iteratively update the weights, and steps three and four are repeated when an iteration period is set or a fault alarm occurs, so as to recalculate the dynamic weights and the comprehensive weights and realize online self-optimization of the weights.

[0044] As a preferred embodiment of the present invention, the trainable attention model is trained based on historical time series data, and the model structure adopts a feedforward neural network, a recurrent neural network, or a combination thereof. Example 1

[0045] Step 1: Collect multi-dimensional operating parameters of the equipment to determine the set of monitoring data affecting the health status of the equipment. This set of operating monitoring data includes dynamic and static parameters. Determine the set of evaluation parameters for the drainage pump. ,in: (Pump start-up interval, min) (Average operating current of a single pump, A) (Average drainage flow rate per pump, m³ / h) (Average drainage time per pump, min) (Water level change rate, m / min).

[0046] Step Two: Based on the Delphi method, domain experts determine the static weights according to the contribution of each parameter to the equipment function and prior knowledge. Multiple water supply and drainage experts, maintenance personnel, and manufacturer technicians were invited to score the parameters using the Delphi method. The static weights were then determined after a consistency check. .

[0047] Step 3: Based on historical equipment operation data and maintenance records, calculate parameter volatility, threshold deviation, and fault correlation characteristics for each monitoring parameter, and construct a statistical feature vector. Then, obtain attention weight vectors for different statistical features using a trainable attention model. Collect one year's worth of operating data from the drainage pump, calculate the volatility of each parameter, and normalize it. Set normal thresholds for each parameter, and calculate the normalized threshold deviation. Statistical analysis of fault events, calculation of fault correlation degree, taking the absolute value and then normalizing. The dynamic permission attention weights are obtained by training the parameters through an LSTM network, with α=[0.238,0.214,0.176,0.192,0.180] and the dynamic weights are... =[0.303,0.240,0.161,0.207,0.089].

[0048] Step 4: Combine static and dynamic weights using a weighted fusion method. , =0.3. The final weights W=[0.2571,0.243,0.2027,0.1749,0.1223] and health H=0.624 were calculated.

[0049] Step 5: Compare the calculated health index with the warning threshold to complete the fault warning. H greater than 0.8 indicates a healthy state, H less than 0.5 triggers a fault warning, and the device is in a sub-healthy state if it falls between the two values. In this embodiment, the device is in a sub-healthy state.

[0050] Step Six: Set up a trigger mechanism to iteratively update the weights. Set the iteration cycle to 30 days or repeat steps three and four when a fault alarm occurs to recalculate the dynamic weights and the overall weights, thereby achieving online self-optimization of the weights.

[0051] The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. 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.

Claims

1. A fault prediction method for rail transit equipment based on a dynamic evolution dual-weight mechanism, characterized in that, Includes the following steps: Step 1: Collect multi-dimensional operating parameters of the equipment to determine the set of monitoring data that affect the health status of the equipment, including dynamic parameters and static parameters; Step 2: Determine the static weights based on the contribution of each parameter to the equipment function and prior knowledge; Step 3: Based on the equipment's historical operating data and maintenance records, calculate multidimensional statistical features for each monitoring parameter and construct a statistical feature vector; And by using a trainable attention model, dynamic weights for different statistical features are obtained; Step 4: Calculate the device health score by weighting and fusing the static and dynamic weights. Step 5: Compare the calculated health index with the warning threshold to complete the fault warning; Step 6: Set up a trigger mechanism to iteratively update the weights.

2. The method for predicting faults in rail transit equipment based on a dynamic evolution dual-weight mechanism according to claim 1, characterized in that, In step one, the set of monitoring data X that affects the health status of the equipment is determined as follows: ， Where, x i Let be the i-th evaluation parameter, and n be the total number of parameters.

3. The method for predicting rail transit equipment faults based on a dynamic evolution dual-weight mechanism according to claim 1, characterized in that, In step two, static weights are determined based on the Delphi method, taking into account the contribution of each parameter to the device function and prior knowledge. for: ， in, S represents static.

4. The fault prediction method for rail transit equipment based on a dynamic evolution dual-weight mechanism according to claim 1, characterized in that, In step three, the statistical feature vector consists of three indicators: historical volatility, threshold deviation, and fault correlation, with the specific formula as follows: ； in This indicates the fluctuation characteristics of the normalized parameters. This represents the frequency characteristics of abnormal parameters after normalization. This represents the correlation characteristics between the normalized parameters and historical failure events; The i-th feature is the historical fluctuation feature. for: ； Where i represents the i-th feature, k represents the k-th historical sampling point, and m represents the number of samples within the sampling window. This represents the mean of the i-th feature within the historical window; After normalization: ； The frequency characteristics of threshold deviation are: ； in, This represents the normal interval for the i-th feature; After normalization: ； The correlation coefficient of the fault correlation is: ； in, Represents the time series of fault events. , This represents the mean value within the k-th historical window; After normalization: 。 5. The method for predicting rail transit equipment faults based on a dynamic evolution dual-weight mechanism according to claim 1, characterized in that, In step three, a trainable attention model is used to obtain dynamic weight vectors with different statistical features, and the weight coefficients are: ； Among them W a and b a These are the model parameters obtained through training with historical data; The loss function is: ； in and The health index and reference health index calculated by the model. It is the attention weight of the i-th parameter at time t. It is the balance coefficient, and n is the total number of parameters.

6. The method for predicting faults in rail transit equipment based on a dynamic evolution dual-weight mechanism according to claim 1, characterized in that, In step four, the dynamic weights are: ； Where i represents the i-th feature, and k represents the k-th historical sampling point. The weight represents the learning weight, and z represents the dynamic weight, which consists of three indicators.

7. The fault prediction method for rail transit equipment based on a dynamic evolution dual-weight mechanism according to claim 1, characterized in that: In step four, the static weights and dynamic weights are weighted and fused to calculate the device health score. The final weights are: ； in The coefficient is a fusion coefficient between 0 and 1, where i represents the i-th feature; The method for calculating the device health status is as follows: 。 8. The method for predicting faults in rail transit equipment based on a dynamic evolution dual-weight mechanism according to claim 1, characterized in that, In step five, the calculated health index is compared with the warning threshold to complete the fault warning. H greater than 0.8 indicates health, H less than 0.5 indicates fault warning, and the device is in a sub-healthy state if it is between the two.

9. The method for predicting faults in rail transit equipment based on a dynamic evolution dual-weight mechanism according to claim 1, characterized in that, In step six, a trigger mechanism is set to iteratively update the weights. When the iteration period is set or a fault alarm occurs, steps three and four are repeated to recalculate the dynamic weights and the comprehensive weights, thereby achieving online self-optimization of the weights.

10. The method according to claim 5, characterized in that, The trainable attention model is trained based on historical time series data, and the model structure adopts a feedforward neural network, a recurrent neural network, or a combination thereof.

Images