Special equipment operation fault prediction method and system based on BP neural network

Abstract

The present invention relates to the field of risk management technology, specifically to a special equipment operation fault prediction method and system based on BP neural network, which collects operating status data of sensors at fixed positions of special equipment, combines the identity information of moving parts on the equipment, and forms an operation record of the equipment component; collects key operating data, environmental parameters and equipment structural properties, and outputs dynamic stress parameters; extracts equipment health trend indicators by performing state correlation analysis on the operation records of the equipment components; fuses the health indicators with stress parameters to form a feature vector for operation and maintenance analysis; inputs the feature vector into a preset risk prediction model, and outputs the operation and maintenance indicators of the equipment component; combines the model output results with a preset risk matrix and management rule library to automatically generate maintenance scheduling suggestions and operation intervention instructions for special equipment, which are used to assist the management system in implementing equipment life cycle management and risk control.

Description

Technical Field

[0001] The present invention relates to the technical field of risk management, and in particular to a method and system for predicting operation failures of special equipment based on a BP neural network. Background Art

[0002] In the operation and management of special equipment, how to plan maintenance, optimize spare parts inventory, and reduce losses caused by unexpected downtime are ongoing issues. Traditional equipment maintenance strategies are mostly based on fixed time periods or respond after a failure occurs. This approach is difficult to cope with the complex operating conditions of special equipment.

[0003] In order to improve the predictability of equipment management, data-driven predictive maintenance methods are introduced in existing technologies. However, some problems still exist. For example, when building a network model, the feature selection and processing methods of the input data are relatively simple, making it difficult to mine information related to faults, resulting in limited accuracy and generalization capabilities of the prediction model. In addition, the neural network training process converges slowly and is prone to falling into local optimality, which becomes a bottleneck in operation management scenarios that require rapid response and decision-making. Therefore, the above problems make the effectiveness and reliability of the fault prediction results based on the existing BP neural network technology still need to be improved when guiding actual management activities such as maintenance scheduling, resource allocation and risk avoidance. Summary of the Invention

[0004] The purpose of the present invention is to provide a special equipment operation fault prediction method and system based on BP neural network, obtain the working condition data of sensors at fixed positions of special equipment, and identify the identity of moving components passing through fixed positions to generate component status data; obtain operating data, environmental data and equipment inherent parameters that affect the load of special equipment; analyze the operating data and environmental data, and calculate the dynamic stress data borne when in a fixed position; compare and analyze the component status data to extract component health status characteristics; couple the component health status characteristics with the dynamic stress data to generate a fused feature vector; input the fused feature vector into the equipment prediction model, and output the prediction results, including the failure probability, remaining service life and risk level of the moving component; based on the prediction results, and in accordance with a preset risk matrix and management rule library, generate maintenance and operation intervention instructions for special equipment.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The special equipment operation fault prediction method based on BP neural network includes:

[0007] Acquire operating condition data of sensors at fixed positions of special equipment and identify the identity of moving components passing through the fixed positions; bind the operating condition data with the identity and current time and position information to generate component status data;

[0008] Obtain operational data, environmental data, and inherent equipment parameters that affect the load of special equipment; analyze operational and environmental data based on the operational physics model to calculate the dynamic stress data on the moving component when it is in a fixed position; conduct multi-dimensional comparative analysis of component status data, including longitudinal health trend analysis of the same moving component at different positions and lateral abnormal location analysis of different moving components at the same position, to extract component health status characteristics; couple component health status characteristics with dynamic stress data to generate a fused feature vector;

[0009] The fused feature vector is input into the equipment prediction model to output the prediction results, including the failure probability, remaining service life and risk level of the motion component; based on the prediction results and in accordance with the preset risk matrix and management rule library, maintenance and operation intervention instructions for special equipment are generated.

[0010] Preferably, the operating data includes real-time passenger flow, load distribution, operating speed and car position information; the environmental data includes multi-point wind speed and direction, temperature, humidity and air pressure data; the equipment inherent parameters include line three-dimensional geometric parameters, material mechanical properties and design load parameters;

[0011] The operational physical model includes a data preprocessing layer, a load calculation layer, and a stress analysis layer: the data preprocessing layer preprocesses operational data, environmental data, and inherent equipment parameters to generate a standardized data set; the load calculation layer converts the preprocessed real-time passenger flow into a car load distribution, calculates the dynamic load based on the operational speed and car position information, analyzes the wind speed and direction data, and generates total load distribution data in real time; the stress analysis layer receives the total load distribution data, calculates the wire rope tension distribution, the bracket force, and the stress concentration coefficient of key nodes, and generates dynamic stress data of the moving components at each fixed position.

[0012] Preferably, the process of acquiring the component health status characteristics includes: extracting time domain characteristics from component status data; performing frequency domain transformation on the time domain characteristics to generate frequency domain characteristic vectors; performing longitudinal trend analysis, forming a time series of historical characteristic data of the same motion component at different fixed positions, calculating the trend slope and degradation rate, and generating a health degradation index; performing horizontal comparative analysis, performing statistical analysis on the characteristic data of different motion components at the same fixed position during the same period, calculating the deviation and abnormality score of each component characteristic value, identifying abnormal components, and outputting a health status characteristic vector containing time domain characteristics, frequency domain characteristic vectors, health degradation indicators and abnormality scores.

[0013] Preferably, the process of obtaining the fused feature vector includes: performing dimensional matching on the health state feature vector and the dynamic stress data to ensure the spatiotemporal consistency of the feature data and the stress data; assigning weight coefficients to the health state feature according to the feature importance, and assigning weight coefficients to the dynamic stress data according to the load influence degree to generate a weighted feature sub-vector; calculating the correlation coefficient between the health state feature and the dynamic stress data to generate an interactive feature item; and performing vector splicing on the weighted feature sub-vector and the interactive feature item to form a fused feature vector.

[0014] Preferably, the device prediction model includes an input layer, a hidden layer and an output layer:

[0015] The input layer receives the fused feature vector and standardizes the input data; the hidden layer adopts a three-layer feedforward neural network structure, the first layer is used to extract nonlinear feature relationships, the second layer performs feature compression, and the third layer performs feature optimization; the output layer sets three output neurons, corresponding to failure probability, remaining service life and risk level respectively. The failure probability outputs the probability value, the remaining service life outputs the time value, and the risk level outputs the level classification result.

[0016] Preferably, a two-dimensional risk matrix is ​​constructed, with the probability of failure as the horizontal axis and three probability intervals of low, medium and high set, and the severity of the failure consequences as the vertical axis and three levels of minor, general and severe set, forming a nine-risk level grid; a management rule library is established, and corresponding decision rules are set for different risk levels, including observation-level rules, warning-level rules and action-level rules;

[0017] When multiple rules are triggered simultaneously, they are prioritized according to risk level, time urgency, and resource availability to ensure that key risks are addressed first. Based on historical prediction accuracy and actual failure occurrences, the probability thresholds and severity assessment criteria of the risk matrix are dynamically adjusted to optimize the triggering conditions and execution parameters of management rules. Based on the failure probability and risk level in the prediction results, the corresponding risk grid is located in the risk matrix to identify the applicable set of management rules.

[0018] Extract matching decision rules from the management rule base and combine them with the current equipment operating status, maintenance resource availability, and operating schedule. Generate equipment monitoring enhancement instructions and preventive maintenance recommendations for observation-level risks, generate operating parameter adjustment instructions for warning-level risks, and generate emergency shutdown instructions and mandatory maintenance work orders for action-level risks. Track the execution status of intervention instructions, collect execution result data, evaluate the effectiveness of intervention measures, and input feedback information into the management rule base for rule optimization to form a closed-loop management decision-making system.

[0019] Special equipment operation fault prediction system based on BP neural network, including:

[0020] The data acquisition unit acquires the working condition data and identity identification of the sensor at the fixed position of the special equipment; binds the working condition data with the identity identification and the current time and location information to generate component status data;

[0021] The fusion analysis unit analyzes operational and environmental data based on the operational physics model, calculates the dynamic stress data on the moving component when it is in a fixed position, performs multi-dimensional comparative analysis on the component status data, extracts component health status characteristics, and couples the component health status characteristics with the dynamic stress data to generate a fusion feature vector.

[0022] The instruction generation unit inputs the fused feature vector into the equipment prediction model and outputs the prediction result; based on the prediction result and in accordance with the preset risk matrix and management rule library, generates maintenance and operation intervention instructions for special equipment.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] 1. This invention implements a fault prediction mechanism by coupling the health status characteristics of motion components with the actual dynamic stress data of the load. This invention extracts time-domain and frequency-domain characteristics, trends, and deviations from health status characteristics, and combines these with stress signature analysis to enhance sensitivity and identification of potential fault indicators. It also outputs multiple indicators, including failure probability, remaining life, and risk level, effectively improving prediction accuracy and generalization capabilities.

[0025] 2. This invention constructs a multi-dimensional health status feature vector by integrating health trend analysis and lateral anomaly comparison for fixed-position moving components. Longitudinal trend analysis tracks the long-term health degradation trajectory of a component, deriving trend slopes and degradation rates. Lateral comparative analysis compares different components at the same location to identify abnormally deviated components, achieving spatial comparative localization. This not only improves health status identification but also enables the identification of high-risk individuals, effectively enhancing the targeted and timely nature of device management.

[0026] 3. This invention establishes a two-dimensional risk matrix composed of failure probabilities and a supporting management rule base, enabling a closed-loop intelligent decision-making process from risk identification to operational intervention. This invention integrates prediction results with equipment management strategies to form a three-level intervention mechanism: observation, warning, and action. Appropriate intervention instructions are automatically generated based on different risk levels, and dynamic sorting and priority execution are implemented based on actual resource availability, equipment status, and operational plans, improving the timeliness, rationality, and flexibility of decision execution. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A schematic diagram of the process flow of the special equipment operation fault prediction method based on BP neural network provided by the present invention;

[0028] Figure 2 This is a schematic diagram of the structure of the special equipment operation fault prediction system based on BP neural network provided by the present invention;

[0029] Figure 3 A schematic diagram of the special equipment operation fault prediction structure provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0031] Example 1:

[0032] See also Figures 1 to 3 The present invention provides a special equipment operation fault prediction method based on BP neural network, which is applied to the special equipment operation fault prediction system based on BP neural network, including a data acquisition unit, a fusion analysis unit and an instruction generation unit. The specific method flow is provided in Figure 1 Refer to the system structure diagram Figure 2 The technical solutions are as follows:

[0033] Acquire operating condition data of sensors at fixed positions of special equipment and identify the identity of moving components passing through the fixed positions; bind the operating condition data with the identity and current time and position information to generate component status data;

[0034] Obtain operational data, environmental data, and inherent equipment parameters that affect the load of special equipment; analyze operational and environmental data based on operational physical models to calculate the dynamic stress data borne by moving components in fixed positions;

[0035] Operational data includes real-time passenger flow, load distribution, operating speed, and car position information; environmental data includes multi-point wind speed and direction, temperature, humidity, and air pressure data; and inherent equipment parameters include line 3D geometry parameters, material mechanical properties, and design load parameters.

[0036] The operational physical model includes a data preprocessing layer, a load calculation layer, and a stress analysis layer. The data preprocessing layer preprocesses operational data, environmental data, and inherent equipment parameters to generate a standardized data set. Specifically, the preprocessing layer normalizes real-time passenger flow data, performs spatial interpolation calculations on multi-point wind speed and direction data, and performs coordinate transformation on the three-dimensional geometric parameters of the line to generate a standardized input data set.

[0037] The load calculation layer converts pre-processed real-time passenger flow data into car load distribution, calculates dynamic loads based on operating speed and car position information, analyzes wind speed and direction data, and generates total load distribution data in real time. Specifically, based on the principles of structural mechanics, the pre-processed passenger flow data is converted into car load distribution, dynamic loads are calculated based on operating speed and car position information, wind speed and direction data are converted into wind-induced loads through an aerodynamic model, and real-time total load distribution data is output.

[0038] The stress analysis layer receives the total load distribution data, calculates the wire rope tension distribution, the bracket force and the stress concentration coefficient of the key nodes, and generates the dynamic stress data of the moving components at each fixed position; specifically, it receives the total load distribution data, combines the material mechanical properties and design load parameters, and uses the finite element analysis method to calculate the wire rope tension distribution, the bracket force and the stress concentration coefficient of the key nodes, and generates the dynamic stress data of the moving components at each fixed position, providing basic load data for subsequent fusion analysis.

[0039] In this embodiment, sensor operating condition data at fixed locations on special equipment is combined with the identity, time, and location information of moving components to generate component status data, improving the accuracy and pertinence of equipment monitoring. By combining real-time operating data, environmental factors, and inherent equipment parameters, an operational physics model is constructed. Layered data processing and analysis are performed to calculate dynamic stresses at key locations. This improves the real-time perception of special equipment operating status and the accuracy of stress response calculations, providing data support for equipment safety assessments and operation and maintenance decisions, and enhancing the system's intelligence and risk warning capabilities.

[0040] Perform multi-dimensional comparative analysis of component status data, including longitudinal health trend analysis of the same motion component at different locations and lateral abnormal location analysis of different motion components at the same location, to extract component health status characteristics;

[0041] The process of obtaining the component health status characteristics includes:

[0042] Extract time-domain features from component status data, including the peak value, root mean square value, and kurtosis coefficient of vibration signals, the mean and rate of change of temperature data, and the energy distribution of acoustic signals;

[0043] Performing frequency domain transformation on the time domain features, extracting the main frequency components and harmonic characteristics of the vibration signal using fast Fourier transform, extracting the time-frequency characteristics of the signal through wavelet transform, and generating a frequency domain feature vector;

[0044] Perform longitudinal trend analysis, construct a time series from the historical characteristic data of the same motion component at different fixed positions, and use the moving average method and exponential smoothing method to calculate the trend slope and degradation rate to generate a health degradation index;

[0045] Perform horizontal comparative analysis, conduct statistical analysis on the feature data of different motion components at the same fixed position during the same period, calculate the deviation and anomaly score of the feature values ​​of each component, identify abnormal components, and output a comprehensive health status feature vector containing time domain features, frequency domain feature vectors, health degradation indicators and anomaly scores.

[0046] In this embodiment, component status data is analyzed vertically and horizontally to extract time-domain, frequency-domain, and time-frequency features, enabling a more accurate assessment of the operating status of moving components. Combining time series analysis with comparative statistical methods, this approach can not only identify health degradation trends of individual components but also locate abnormal individuals within a group, enabling a comprehensive assessment of the cableway's overall health. This provides a reliable basis for subsequent predictive maintenance and fault warnings, improving the intelligence of equipment monitoring and operational efficiency.

[0047] Couple the component health status characteristics with the dynamic stress data to generate a fusion feature vector;

[0048] The process of obtaining the fused feature vector includes:

[0049] Dimensionally match the component health status feature vector and dynamic stress data, ensuring the spatiotemporal consistency of the feature data and stress data through timestamps and location information;

[0050] A weighted fusion method is used to assign weight coefficients to health state features according to feature importance and to assign weight coefficients to dynamic stress data according to load influence, generating weighted feature sub-vectors.

[0051] Through feature interaction analysis, the correlation coefficient between health status features and dynamic stress data is calculated to generate interactive feature terms, including the product of vibration amplitude and stress, and the product of temperature change rate and load change rate;

[0052] The weighted feature sub-vectors and interactive feature items are vector-concatenated to form a high-dimensional fusion feature vector containing the original features, stress information, and feature interaction relationships. The vector dimension is the sum of the health feature dimension, stress feature dimension, and interactive feature dimension.

[0053] In this embodiment, by coupling component health status features with dynamic stress data to construct a fused feature vector, the integrity and relevance of the features are effectively improved. Timestamps and location information ensure the temporal and spatial consistency of the data. Weighted fusion enhances the representation of key features, and interactive features reveal the inherent connection between health status and load changes. The high-dimensional fused feature vector integrates information on health, stress, and their interactions, providing a data foundation for subsequent anomaly detection, risk assessment, and predictive maintenance, and improving the robustness of model analysis.

[0054] Inputting the fused feature vector into an equipment prediction model and outputting prediction results, including the failure probability, remaining service life, and risk level of the motion component;

[0055] The device prediction model includes an input layer, a hidden layer, and an output layer:

[0056] The input layer receives the fused feature vector, normalizes the input data, maps the feature value to the interval [0,1], and sets the number of input neurons to be equal to the dimension of the fused feature vector;

[0057] The hidden layer adopts a three-layer feedforward neural network structure. The number of neurons in the first hidden layer is 1.5 times that of the input layer. The ReLU activation function is used to process the input data and extract nonlinear feature relationships. The number of neurons in the second hidden layer is 0.8 times that of the first hidden layer. The Sigmoid activation function is used for feature compression. The number of neurons in the third hidden layer is 0.6 times that of the second hidden layer. The Tanh activation function is used for feature optimization.

[0058] The output layer sets three output neurons, corresponding to the failure probability, remaining service life and risk level respectively. The failure probability output neuron uses the Sigmoid activation function to output a probability value in the range of 0-1, the remaining service life output neuron uses the linear activation function to output the time value, and the risk level output neuron uses the Softmax activation function to output the discrete level classification result, providing quantitative prediction data for subsequent decision analysis.

[0059] In this example, by constructing a multi-layered equipment prediction model, efficient processing of fused feature vectors and deep learning modeling were achieved. The overall model structure was optimized and rationalized, providing a reliable prediction basis for equipment operation risk assessment and health management.

[0060] Based on the prediction results and in accordance with a preset risk matrix and management rule base, maintenance and operation intervention instructions for special equipment are generated.

[0061] The process of constructing the risk matrix and management rule base includes:

[0062] Construct a two-dimensional risk matrix with three probability intervals (low, medium, and high) on the horizontal axis for the probability of failure, and three levels (minor, general, and severe) on the vertical axis for the severity of the consequences of failure. This results in nine risk level grids, each corresponding to a specific management response strategy.

[0063] Establish a management rule library and set corresponding decision rules for different risk levels, including observation-level rules, warning-level rules, and action-level rules. Each rule includes three elements: trigger conditions, execution actions, and time requirements;

[0064] Define a rule priority mechanism. When multiple rules are triggered simultaneously, prioritize them according to risk level, time urgency, and resource availability to ensure that key risks are addressed first.

[0065] Establish a rule update mechanism, and use machine learning methods to dynamically adjust the probability threshold and severity assessment criteria of the risk matrix based on historical prediction accuracy and actual failure occurrence, optimize the triggering conditions and execution parameters of management rules, and achieve continuous improvement of management decisions.

[0066] The specific process of generating maintenance and operation intervention instructions includes:

[0067] Based on the failure probability and risk level in the prediction results, locate the corresponding risk grid in the risk matrix and identify the applicable management rule set;

[0068] Extract matching decision rules from the management rule base and calculate the feasibility score of each intervention measure based on the current equipment operating status, maintenance resource availability, and operation schedule;

[0069] Generates graded intervention instructions. For observation-level risks, it generates equipment monitoring enhancement instructions and preventive maintenance recommendations, including specific monitoring parameters and maintenance time windows. For warning-level risks, it generates operating parameter adjustment instructions, including speed limits, load control thresholds, and operating time limits. For action-level risks, it generates emergency shutdown instructions and mandatory maintenance work orders, including downtime, maintenance content, and responsible personnel assignment.

[0070] Establish an instruction execution feedback mechanism, track the execution status of intervention instructions, collect execution result data, evaluate the effectiveness of intervention measures, input feedback information into the management rule base for rule optimization, and form a closed-loop management decision-making system.

[0071] In this embodiment, by introducing a risk matrix and a management rule base, and combining prediction results to generate hierarchical maintenance and operational intervention instructions, we achieve intelligent and closed-loop risk management for special equipment. The risk matrix describes the combination of failure probability and consequence severity, while the management rule base provides specific response strategies and implementation specifications, ensuring that each type of risk has a clear response measure. Through a rule-based prioritization mechanism and a machine learning-driven dynamic optimization mechanism, we achieve scientific sequencing and continuous improvement of intervention strategies, improving the rationality, timeliness, and resource utilization efficiency of operation and maintenance decisions, and enhancing the safety and reliability of equipment operation.

[0072] The present invention binds the sensor working condition data with the specific motion component identity, and combines it with the running physical model to accurately calculate its dynamic stress under changes in real-time passenger flow, wind load, etc., to build an analytical foundation for deep coupling of "state-stress", which significantly improves the accuracy of state perception. On this basis, a two-dimensional analysis method combining longitudinal trends and horizontal comparisons is adopted, which can not only track the long-term performance degradation of individual components, but also instantly discover abnormal individuals in the group, and achieve comprehensive capture of fault precursors. These multi-dimensional features containing health status and dynamic stress are deeply integrated and input into the optimized neural network model to output the failure probability, remaining life and risk level. Finally, the prediction results are seamlessly connected to an intelligent decision-making engine composed of a risk matrix and a management rule base. The engine can automatically generate hierarchical and specific intervention instructions, and perform self-optimization based on feedback on the execution effect of the instructions, forming a set of intelligent and adaptive management systems from precise prediction to closed-loop control, which fundamentally improves the operational safety, reliability and operation and maintenance efficiency of special equipment. For specific reference, Figure 3 .

[0073] Example 2:

[0074] This embodiment takes a mountain ropeway system as an application scenario and elaborates on the specific implementation process of a special equipment operation fault prediction method based on a BP neural network.

[0075] A sensor array is deployed at key nodes of the cableway system. Each grip is given a unique digital identity through an RFID tag, in the form of a code such as "CLG-077", where CLG represents the grip type and 077 is the serial number.

[0076] When the cable grip CLG-077 passes the fifth support, the system instantly triggers sensor acquisition, capturing vibration waveform data, acoustic spectrum information, temperature readings, and displacement changes. It also records information such as the current timestamp, support coordinate position, and cableway speed. This operating condition data is strongly associated with the CLG-077 identifier and its temporal and spatial information to generate a complete component status record. The record format is: time - component ID - location coordinates - operating parameters - sensor data array.

[0077] Through continuous data collection, the system creates a running trajectory file for each gripper along the entire line, realizing the transition from traditional "measurement point monitoring" to "component file".

[0078] Operational data includes real-time passenger flow provided by the ticketing system, load distribution information from the dispatching system, operating speed, and car position data; environmental data covers multi-point wind field information collected by wind speed and direction meters along the line, temperature and humidity sensor data, and air pressure change records; equipment inherent parameters include the three-dimensional geometric model of the cableway line, the mechanical properties of the wire rope material, and the design load parameters of each bracket.

[0079] The data preprocessing layer standardizes the raw data, maps the real-time passenger flow data to the standard interval using a normalization algorithm, implements Kriging spatial interpolation calculations on multi-point wind speed and direction data, and performs coordinate system transformation on the three-dimensional geometric parameters of the line to ensure the uniformity of the data format.

[0080] The load calculation layer, based on the principles of structural mechanics, converts standardized passenger flow data into actual load distribution for each car. It then calculates dynamic load effects based on the current operating speed and car position. Furthermore, it uses an aerodynamic model to convert wind field data into wind-induced loads acting on the cables, ultimately outputting real-time total load distribution data.

[0081] The stress analysis layer receives the total load data and, combining the wire rope's material properties such as elastic modulus and Poisson's ratio with the designed load-bearing capacity of each support, uses finite element analysis to calculate in real time the tension distribution of each wire rope segment, the stress conditions on the support, and the stress concentration factors at key connection nodes. This layer generates dynamic stress data for the CLG-077 cable grip at support No. 5, including tangential, normal, and torsional stress components.

[0082] First, time-domain features are extracted from the vibration signal. These include the peak value, which reflects the impact intensity; the root mean square value, which represents the overall vibration level; and the kurtosis coefficient, which indicates the signal distribution. The mean and rate of change are extracted from the temperature data, and the energy distribution features are extracted from the acoustic signal.

[0083] Frequency domain transformation is performed on time domain features. Fast Fourier transform is used to analyze the main frequency components and harmonic structure of the vibration signal to identify the inherent frequency characteristics of the equipment operation. Wavelet transform is used to extract the time-frequency localization characteristics of the signal to capture transient abnormal events. A frequency domain feature vector containing frequency amplitude, phase information, and time-frequency distribution is generated.

[0084] A longitudinal trend analysis was performed, extracting historical characteristic data for the CLG-077 cable grip at the fifth support bracket and constructing a time series. A moving average method was used to eliminate short-term fluctuations, while exponential smoothing was used to highlight recent trends. The degradation slope and deterioration rate of the health status were calculated to quantify the long-term health evolution trajectory of the component.

[0085] Conduct a horizontal comparative analysis, collecting characteristic data from all grips at Support 5 during the same period for statistical comparison. Calculate the deviation and anomaly score of CLG-077 relative to the group mean to identify any significant anomalies. If only CLG-077 data is abnormal, the fault is attributed to that grip. If all grips are abnormal at this location, the fault is attributed to Support 5.

[0086] The final output is a comprehensive health status feature vector, which contains information in four dimensions: time domain features, frequency domain feature vectors, health degradation indicators, and group abnormality scores.

[0087] Fusion processing is performed on the health state feature vector and dynamic stress data; first, dimension matching is performed to ensure the spatiotemporal consistency of the two types of data through timestamps and spatial coordinates, ensuring accurate correspondence of the analysis objects.

[0088] An adaptive weighted fusion strategy is employed to assign differentiated weight coefficients to health features such as vibration, temperature, and acoustics based on the importance of features determined through historical data analysis. Dynamic stress components such as tangential stress, normal stress, and torsional stress are also weighted based on the load's impact on equipment life. Weighted feature subvectors are generated to highlight the contribution of key information.

[0089] Perform feature interaction analysis to calculate the correlation coefficient between health status features and dynamic stress data, identifying the inherent relationships between the two. Generate interactive feature terms, such as the product of vibration amplitude and real-time stress, and the product of temperature change rate and load change rate. These interactive terms can reflect the response characteristics of the equipment under specific stress conditions.

[0090] The weighted feature sub-vectors are concatenated with the interactive feature items to form a fused feature vector. This vector contains the original health feature information, the real-time stress state, and the interaction between the two, comprehensively describing the equipment's operating status and load environment.

[0091] The device prediction model uses a three-layer feedforward neural network architecture. The input layer receives the fused feature vector, normalizes the data, and maps the feature values ​​to a uniform range of zero to one to facilitate network training convergence. The number of neurons in the input layer is set to the number of dimensions of the fused feature vector.

[0092] The hidden layer is designed as a three-layer structure to achieve hierarchical feature extraction. The first hidden layer has 1.5 times the number of neurons in the input layer and uses the ReLU activation function to process the input information and extract nonlinear feature relationships in the data. The second hidden layer has 0.8 times the number of neurons in the first layer and uses the Sigmoid activation function to achieve feature compression and reduce data redundancy. The third hidden layer has 0.6 times the number of neurons in the second layer and uses the Tanh activation function for feature optimization to enhance the model's expressiveness.

[0093] The output layer sets three neurons to output failure probability, remaining service life and risk level respectively; the failure probability neuron uses the Sigmoid activation function to output a probability value ranging from zero to one; the remaining service life neuron uses a linear activation function to directly output the time prediction value; the risk level neuron uses the Softmax activation function to output the classification probability of three discrete levels: low, medium and high.

[0094] After network calculation, the system outputs the prediction results of the rope gripper CLG-077: the failure probability is medium, the estimated remaining service life is six months, and the current risk level is medium.

[0095] A two-dimensional risk matrix is ​​constructed, with the horizontal axis representing the probability of failure divided into three intervals: low, medium, and high, and the vertical axis representing the severity of the consequences of failure divided into three levels: minor, general, and severe, forming nine risk grids; each grid corresponds to a specific management response strategy and intervention intensity.

[0096] Establish a three-level management rule base. Observation-level rules apply to low-risk situations and primarily include recommendations for enhanced monitoring and preventive maintenance. Warning-level rules address medium-risk situations and involve adjustments to operating parameters and restrictive measures. Action-level rules address high-risk situations, including emergency shutdowns and mandatory maintenance requirements. Each rule clearly defines the triggering conditions, execution actions, and completion deadlines.

[0097] Based on the prediction results of the cable gripper CLG-077, the system locates the grid corresponding to medium failure probability and general severity in the risk matrix and identifies the applicable early warning level management rules.

[0098] The system extracts matching decision rules from the rule base and calculates the feasibility of various intervention measures based on the current cableway operating status, maintenance team availability, and holiday operation plans. For the medium-risk condition of CLG-077, the system generates an early warning-level intervention instruction: it recommends reducing the cableway operating speed from six meters per second to four meters per second when passing through the wind outlet section support within the next week to reduce the impact of wind-induced loads on the cable grip; and arranges for a key inspection of CLG-077 during the next routine maintenance, including testing of bearing lubrication and spring elasticity.

[0099] Establish a command execution tracking mechanism to monitor the effectiveness of intervention measures. Collect status data from the CLG-077 during reduced speed operation to assess whether vibration levels have improved. Feedback execution results to the management rule base to optimize rule trigger thresholds and intervention intensity, achieving continuous improvement in management decisions. See Table 1 for details.

[0100] Table 1 Technical effect comparison and verification table

[0101]

[0102] In terms of fault prediction accuracy, the system of the present invention can identify subtle changes in the health status of equipment in advance, providing a sufficient time window for maintenance plan formulation; in terms of operational safety, the active intervention mechanism effectively prevents the occurrence of potential safety accidents, ensuring passenger safety and equipment integrity; in terms of economic benefits, the preventive maintenance strategy significantly reduces downtime losses and emergency repair costs caused by equipment failures.

[0103] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A special equipment operation fault prediction method based on BP neural network, characterized in that: include: Acquire the working condition data of sensors at fixed positions of special equipment and identify the identity of moving components passing through fixed positions; Binding the operating condition data with the identity identifier and current time and location information to generate component status data; Obtain operational data, environmental data and equipment-specific parameters that affect special equipment loads; Based on the operational physics model, operational and environmental data are analyzed to calculate the dynamic stress data on the motion components when they are in a fixed position. Multi-dimensional comparative analysis of component status data is performed, including longitudinal health trend analysis of the same motion component at different positions and lateral abnormal location analysis of different motion components at the same position, to extract component health status characteristics. Couple component health status characteristics with dynamic stress data to generate a fusion feature vector; The fused feature vector is input into the equipment prediction model to output the prediction results, including the failure probability, remaining service life and risk level of the motion component; based on the prediction results and in accordance with the preset risk matrix and management rule library, maintenance and operation intervention instructions for special equipment are generated.

2. The method for predicting operation failure of special equipment based on BP neural network according to claim 1, characterized in that: Operational data includes real-time passenger flow, load distribution, operating speed, and car position information; environmental data includes multi-point wind speed and direction, temperature, humidity, and air pressure data; and inherent equipment parameters include line 3D geometry parameters, material mechanical properties, and design load parameters. The operational physical model includes a data preprocessing layer, a load calculation layer, and a stress analysis layer: the data preprocessing layer preprocesses operational data, environmental data, and inherent equipment parameters to generate a standardized data set; The load calculation layer converts the pre-processed real-time passenger flow into the car load distribution, calculates the dynamic load based on the running speed and car position information, analyzes the wind speed and direction data, and generates the total load distribution data in real time; the stress analysis layer receives the total load distribution data, calculates the wire rope tension distribution, the bracket force and the stress concentration coefficient of the key nodes, and generates the dynamic stress data of the moving components at each fixed position.

3. The method for predicting operation failure of special equipment based on BP neural network according to claim 1, characterized in that: The process of acquiring the component health status characteristics includes: extracting time domain characteristics from component status data; performing frequency domain transformation on the time domain characteristics to generate frequency domain characteristic vectors; performing longitudinal trend analysis, forming a time series of historical characteristic data of the same moving component at different fixed positions, calculating the trend slope and degradation rate, and generating a health degradation index; performing horizontal comparative analysis, performing statistical analysis on the characteristic data of different moving components at the same fixed position during the same period, calculating the deviation and anomaly score of each component characteristic value, identifying abnormal components, and outputting a health status characteristic vector containing time domain characteristics, frequency domain characteristic vectors, health degradation indicators, and anomaly scores.

4. The method for predicting operation failure of special equipment based on BP neural network according to claim 1, characterized in that: The acquisition process of the fused feature vector includes: performing dimension matching on the health state feature vector and the dynamic stress data to ensure the spatiotemporal consistency of the feature data and the stress data; assigning weight coefficients to the health state feature according to feature importance, and assigning weight coefficients to the dynamic stress data according to the load influence degree to generate a weighted feature sub-vector; calculating the correlation coefficient between the health state feature and the dynamic stress data to generate an interactive feature item; and performing vector concatenation on the weighted feature sub-vector and the interactive feature item to form a fused feature vector.

5. The method for predicting operation failure of special equipment based on BP neural network according to claim 1 is characterized in that: The device prediction model includes an input layer, a hidden layer, and an output layer: The input layer receives the fused feature vector and standardizes the input data; the hidden layer adopts a three-layer feedforward neural network structure, the first layer is used to extract nonlinear feature relationships, the second layer performs feature compression, and the third layer performs feature optimization; the output layer sets three output neurons, corresponding to failure probability, remaining service life and risk level respectively. The failure probability outputs the probability value, the remaining service life outputs the time value, and the risk level outputs the level classification result.

6. The method for predicting operation failure of special equipment based on BP neural network according to claim 1, characterized in that: Construct a two-dimensional risk matrix, with the probability of failure as the horizontal axis, setting three probability intervals of low, medium, and high, and the severity of the failure consequences as the vertical axis, setting three levels of minor, general, and severe, forming a nine-risk level grid; Establish a management rule library and set corresponding decision-making rules for different risk levels, including observation-level rules, warning-level rules, and action-level rules; When multiple rules are triggered simultaneously, they are prioritized according to risk level, time urgency, and resource availability to ensure that key risks are addressed first. Based on historical prediction accuracy and actual failure occurrences, the probability thresholds and severity assessment criteria of the risk matrix are dynamically adjusted to optimize the triggering conditions and execution parameters of management rules. Based on the failure probability and risk level in the prediction results, the corresponding risk grid is located in the risk matrix to identify the applicable set of management rules. Extract matching decision rules from the management rule library and combine them with the current equipment operating status, maintenance resource availability, and operating schedule. Generate equipment monitoring enhancement instructions and preventive maintenance recommendations for observation-level risks, operating parameter adjustment instructions for warning-level risks, and emergency shutdown instructions and mandatory maintenance work orders for action-level risks. Track the execution status of intervention instructions, collect execution result data, evaluate the effectiveness of intervention measures, input feedback information into the management rule base for rule optimization, and form a closed-loop management decision-making system.

7. The special equipment operation fault prediction system based on BP neural network is characterized by: The method for predicting operation failures of special equipment based on a BP neural network according to claim 1 comprises: The data acquisition unit acquires the working condition data and identity identification of the sensor at the fixed position of the special equipment; binds the working condition data with the identity identification and the current time and location information to generate component status data; The fusion analysis unit analyzes operational and environmental data based on the operational physics model, calculates the dynamic stress data on the moving component when it is in a fixed position, performs multi-dimensional comparative analysis on the component status data, extracts component health status characteristics, and couples the component health status characteristics with the dynamic stress data to generate a fusion feature vector. The instruction generation unit inputs the fused feature vector into the equipment prediction model and outputs the prediction result; based on the prediction result and in accordance with the preset risk matrix and management rule library, generates maintenance and operation intervention instructions for special equipment.

Images