CN118066934A

CN118066934A - Intelligent fault positioning system for weapon equipment

Info

Publication number: CN118066934A
Application number: CN202410195918.XA
Authority: CN
Inventors: 金俊; 段莉娟
Original assignee: Avic Power Science & Technology Engineering Co ltd
Current assignee: Avic Power Science & Technology Engineering Co ltd
Priority date: 2024-02-22
Filing date: 2024-02-22
Publication date: 2024-05-24

Abstract

The invention relates to the technical field of weapon testing, in particular to an intelligent fault positioning system for weapon equipment, which improves the accuracy and efficiency of fault positioning by integrating data acquisition, fault association analysis, a neural network model and an intelligent learning module. The data acquisition module collects operation data and history labeling judgment data of the weapon equipment, and then a component fault tree is generated through the fault association module. The fault locating module uses the constructed neural network model to convert the operation data and the historical data into characteristic point values, calculates probability distribution of fault types, and accordingly determines the most probable fault types and positions. The intelligent learning module adjusts the configuration of the weight matrix according to the component fault tree, optimizes the loss function to adapt to the change of the fault mode, and refines fault diagnosis by dynamically adjusting the fault tree coding matrix, thereby enhancing the adaptability of the model to the newly-appearing fault mode and providing an efficient and intelligent solution for maintenance and repair of weaponry.

Description

Intelligent fault positioning system for weapon equipment

Technical Field

The invention relates to the technical field of weapon testing, in particular to an intelligent fault positioning system for weapon equipment.

Background

In the field of maintenance and fault diagnosis of modern weaponry, conventional maintenance and fault diagnosis methods are facing increasing challenges as technology evolves and the complexity of the operational environment increases. Fault Tree Analysis (FTA), while performing well in the reliability analysis of simple or static systems, is a well established method widely used in the industry, and its limitations are emerging when faced with increasingly complex weapon systems today. The conventional FTA method is highly dependent on expert knowledge to build fault trees. This not only means that a lot of time and expertise are required in constructing and updating the fault tree, but also that when new fault modes occur as equipment technology is continually advanced, existing fault trees may be difficult to update in time, thereby affecting the accuracy of fault diagnosis. Traditional FTA is mainly aimed at static system analysis, and is difficult to adapt to dynamic changes of equipment operation environment and failure modes. In practical application, the running state of the system may be affected by various factors, such as environmental condition change, system aging, etc., and these dynamic factors often lead to diversified fault modes, which increases the difficulty of fault diagnosis. Meanwhile, the traditional method has limitations in processing and analyzing large-scale data, lacks effective data processing tools and algorithms, and is difficult to extract useful information from complex data.

Disclosure of Invention

In order to solve the problems, the invention provides an intelligent fault positioning system for weapon equipment.

In order to achieve the above purpose, the invention adopts the following technical scheme:

An intelligent fault location system for a weapon equipment, comprising: the system comprises a data acquisition module, a fault correlation module, a fault positioning module and an intelligent learning module, wherein the data acquisition module, the fault correlation module, the intelligent learning module and the fault positioning module are sequentially connected, and the data acquisition module is connected with the fault positioning module;

The data acquisition module is used for collecting operation data and history labeling judgment data of the weapon equipment;

the fault association module is used for generating a component fault tree according to the history labeling judgment data;

The fault positioning module is used for constructing a neural network model, converting the operation data and the history labeling judgment data of the weapon equipment into characteristic point values in the neural network model, inputting the characteristic point values into the neural network model in a matrix mode, calculating probability distribution of each fault type through the neural network model, determining the fault type and the position according to the probability distribution, and generating a fault positioning result;

the intelligent learning module is used for adjusting the configuration of the weight matrix according to the component fault tree, wherein the calculation formula of the loss function L is as follows:

L＝L_pred+λL_tree(P，F)，

Wherein L _pred is a prediction error term, λL _tree (P, F) is a constraint term based on the logic of the component fault tree, λ is a balance factor, and F is the coding matrix of the component fault tree;

The intelligent learning module is also used for adjusting the component fault tree coding matrix F at regular time according to the fault positioning result and the history marking judgment data.

Further, the operational data of the weapon equipment includes temperature readings, pressure values, vibration data and current and voltage measurements under different operations.

Further, the generating the component fault tree according to the history labeling judgment data includes:

and judging fault cases and corresponding fault information of the components of the weapon equipment in the data by using the history labels, and constructing a fault tree of known fault types and related components thereof through superposition of the fault cases.

Further, the converting the operation data and the history labeling judgment data of the weapon equipment into the characteristic point values for matrix input includes:

and carrying out standardization processing on the collected operation data and the history labeling judgment data, wherein the standardization processing comprises noise removal, missing value filling and data standardization.

Further, the configuration of the weight matrix according to the component fault tree comprises the following steps:

Analyzing a coding matrix F of a component fault tree, and identifying a fault path;

according to the fault path, the weight of the relevant path in the neural network is adjusted through a gradient back propagation algorithm;

And dynamically adjusting constraint terms lambdaL _tree (P, F) based on the logic of the component fault tree in the loss function according to the weight adjustment result.

Further, the step of adjusting the component fault tree coding matrix F according to the fault locating result and the history labeling judgment data includes:

Carrying out weighted average on the fault positioning result and the history labeling judgment data according to a preset time interval to obtain an adjustment score of each fault type;

and adjusting the value interval of each element in the parameter fault tree coding matrix F according to the adjustment score of each fault type.

Further, the fault locating result and the historical labeling judgment data are weighted and averaged according to a preset time interval, wherein the weight of the historical labeling judgment data is greater than that of the fault locating result.

Further, the adjusting the value interval of each element in the parameter fault tree coding matrix F according to the adjustment score of each fault type includes:

When the adjustment score is larger than a preset threshold value, generating a first bias according to the difference value between the adjustment score and the preset threshold value, and adjusting the value interval of each element in the parameter fault tree coding matrix F according to the first bias;

And when the adjustment score is smaller than or equal to the preset threshold value, generating a second bias according to the difference value between the adjustment score and the preset threshold value, and adjusting the value interval of each element in the parameter fault tree coding matrix F according to the second bias.

The invention has the beneficial effects that: according to the invention, the fault tree of the components is generated according to the historical labeling judgment data by collecting the operation data and the historical labeling judgment data of the weapon equipment, the fault positioning process is optimized by utilizing the machine learning technology, and particularly, the accuracy of fault diagnosis is improved by adjusting the configuration of the weight matrix. The key of the process is the design of the loss function L, which directly influences the model training effect and the fault prediction accuracy. The loss function L consists of two parts: a prediction error term and a constraint term based on component fault tree logic. The prediction error term measures the deviation between the model prediction result and the actual fault type, and aims to minimize the error and improve the accuracy of model prediction. The constraint item based on the element fault tree logic is to constrain the weight matrix according to the structure of the fault tree so as to ensure the logic consistency and reliability of the model prediction result. The intelligent learning module automatically adjusts the configuration of the neural network weight matrix by optimizing the loss function L. The process involves gradient descent optimization of the loss function, and weight values are adjusted according to the partial derivatives of the loss function with respect to weights, so as to achieve the purposes of reducing prediction errors and meeting the logical constraints of the fault tree. The intelligent learning also fuses the real-time fault locating result with the history labeling judgment data. And the intelligent learning module can continuously optimize the fault diagnosis model to adapt to the change of the system and the newly-appearing fault mode. The process not only improves the accuracy and efficiency of fault location, but also enhances the adaptability of the system to complex and dynamic weapon equipment environments, and remarkably improves the performance of maintenance and diagnosis work.

By combining automatic fault tree updating, a machine learning model utilizing operation data and self-adaptive optimization of an intelligent learning module, the invention provides an efficient, accurate and self-adaptive solution for fault diagnosis of modern weaponry, and effectively reduces the difficulty of fault diagnosis in the weaponry.

Drawings

FIG. 1 is a schematic diagram of a weapon intelligent test system according to the present invention.

Detailed Description

Referring to fig. 1, the present invention relates to an intelligent testing system for weaponry, comprising: the system comprises a data acquisition module, a fault correlation module, a fault positioning module and an intelligent learning module, wherein the data acquisition module, the fault correlation module, the intelligent learning module and the fault positioning module are sequentially connected, and the data acquisition module is connected with the fault positioning module;

L＝L_pred+λL_tree(P，F)，

It should be noted that, the data acquisition module monitors the critical performance parameters of the weapon equipment, such as temperature, pressure, vibration, current, voltage, etc., in real time, and collects the historical fault data, including the validated fault instance and the corresponding diagnostic information thereof. These data provide the underlying data support for subsequent failure analysis. The fault association module automatically builds a component fault tree by using historical fault data through a data analysis technology, and the process involves complex algorithms such as cluster analysis and pattern recognition to determine fault transmission paths and logic dependency relations among different components. The dynamic construction and updating mechanism of the fault tree ensures that the model can reflect the latest system state and fault mode. The fault locating module converts the collected operation data and the historical fault data into feature vectors which can be processed by the neural network model. The module calculates probability distribution of each potential fault type by adopting a deep learning technology, so as to accurately identify the most probable fault type and the specific position thereof. This process relies on complex mathematical models and algorithms, such as probabilistic statistical analysis and optimization algorithms, to ensure accuracy of fault prediction. The intelligent learning module adjusts the weight matrix configuration according to the logic of the component fault tree, and adopts a loss function optimization technology to carry out model training. The loss function is combined with a prediction error term and a fault tree logic constraint term, and the neural network weight is adjusted through an iterative optimization algorithm such as a gradient descent method so as to reduce the prediction error and meet the fault logic constraint. In addition, the module updates the coding matrix F of the component fault tree according to the fault positioning result and the historical fault data at regular time, and the fault tree model is dynamically optimized by adopting a self-adaptive algorithm, so that the quick response capability of the system to a new fault mode is ensured.

In particular, temperature readings are key indicators for assessing the thermal stability of the equipment operating environment and its internal components. Abnormal temperature changes may indicate heat dissipation problems, overload, or internal component damage. The system monitors the temperature of the key part in real time through the temperature sensor and records the temperature fluctuation condition so as to analyze whether the equipment is in a normal working range. Monitoring of pressure values is critical to hydraulic systems, sealed tanks, and other equipment components. The data provided by the pressure sensor may be used to evaluate the sealing performance of the system, the operating condition of the hydraulic system, and the structural load-carrying capacity. Abnormal pressure readings may be indicative of leaks, blockages, or compromised structural integrity. The acquisition of vibration data is an important means of identifying anomalies in the operation of mechanical systems. By vibration analysis, problems such as bearing damage, unbalance, dislocation or gear failure can be detected. The system captures vibration characteristics generated when the equipment is operated by utilizing the vibration sensor, and analyzes the frequency, the amplitude and the mode of the vibration sensor for early fault early warning and state monitoring. The monitoring of the current and voltage measurements reflects the operating state of the electrical system of the installation. Real-time data of current and voltage can reveal the operating efficiency of the electrical components, the integrity of the circuit, and possible points of failure under different operating conditions. For example, a sudden increase in current may indicate a short circuit event, while a drop in voltage may be due to overload or poor electrical connection.

Specifically, based on the analysis tool of the graph theory, the dependency relationship among components is identified and mapped. The algorithm takes into account direct and indirect correlations between components, including series and parallel dependencies, as well as more complex logic structures such as "or gate" and gate "relationships. These relationships define how other components are affected when one fails, and how these effects accumulate resulting in system level failure. The system uses information in fault cases, such as indexes of fault occurrence frequency, fault detection difficulty, severity of influence of faults on system performance and the like, to give weight to each node and edge of the fault tree. This step is accomplished by quantitative methods such as using bayesian networks to estimate the probability of fault propagation, or applying fuzzy logic to deal with uncertainties in fault diagnosis. Finally, the fault tree is presented in a visual form, wherein nodes represent specific fault modes or components, edges represent dependency relations among the components, and the weight reflects the characteristics and the importance of the relations. The fault tree not only can intuitively display the fault logic structure of the system, but also can provide decision support for subsequent fault diagnosis and prevention.

It should be noted that, during the collection process, the operation data may be subjected to various interferences, such as electromagnetic interference, signal changes caused by equipment aging, and the like, so that noise occurs in the data. The system adopts signal processing technology, such as a low-pass filter, a median filter and the like, to remove or reduce the influence of noise and improve the signal-to-noise ratio of data. For example, if the temperature sensor readings suddenly change dramatically, which is not physically, likely due to noise, the system will correct for such data anomalies by filtering techniques. In practice, there may be missing values in the collected data set due to sensor failures, data transmission errors or other technical problems. The system needs to handle these missing values in order not to affect the integrity and accuracy of the subsequent analysis. Common methods include estimating the missing values using the mean, median, or by time series analysis, interpolation, etc. of the historical data. For example, if pressure data for a certain period is missing, the system may estimate the missing value by analyzing the data for neighboring time points. Different operating parameters may have different dimensional and numerical ranges, and direct use of these raw data as model inputs may result in inefficiency or bias in the results of model training. Therefore, the system needs to perform normalization processing on the data, and convert the data to a uniform scale. Common normalization methods include min-max normalization (normalization) and Z-score normalization (standard deviation normalization). By this processing, for example, converting all sensor data into a numerical value between 0 and 1, or into a distribution having zero mean and unit variance, the training efficiency and prediction accuracy of the model can be effectively improved. After the preprocessing is completed, the system converts the processed data into characteristic point values and organizes the characteristic point values into a matrix form suitable for being input into a neural network model. The feature matrix comprises key information extracted from the operation data and the historical labeling judgment data, and provides a basis for fault diagnosis and prediction. Through the series of professional data processing steps, the system ensures the quality and consistency of data, and provides a solid data basis for accurately diagnosing the running state and potential faults of the weapon equipment.

It should be noted that the process of parsing the coding matrix F involves performing deep analysis on the matrix to extract accurate information of the fault propagation path. In particular, this step requires the use of advanced algorithms, such as graph theory analysis or network flow algorithms, to traverse the structure of the fault tree, identifying all possible paths from the root node (i.e., the final fault phenomenon) to the leaf node (i.e., the primary fault cause). Each path represents a logical chain of components that may fail causing a failure of the upper level system. For example, assuming that a weapon has failed a, possible causes include failure of components X and Y according to the fault tree. If the failure of component X is again likely to be caused by the failure of components W and V, then the information of these components and their associated paths will be contained in the encoding matrix F. The analysis coding matrix F identifies fault paths such as 'A is caused by X and Y, X is caused by W and V', and the paths are clearly identified, so that basis is provided for subsequent weight adjustment. By accurately analyzing the coding matrix F of the component fault tree, the system can determine the contribution degree of each component fault to the final fault phenomenon, which is the basis for adjusting the weight of the neural network and dynamically adjusting the constraint term of the loss function through a gradient back propagation algorithm. The pertinence and the effectiveness of the weight adjustment process are ensured, and the accuracy and the reliability of the fault diagnosis model are further improved. The gradient back-propagation algorithm is an optimization algorithm for updating the weights and biases of the neural network according to the gradient of the loss function in order to minimize the loss function, i.e. to reduce the difference between the model predicted output and the actual result. In this step, the algorithm first calculates the gradient of the loss function (e.g., mean square error or cross entropy loss) with respect to the network weights, which involves forward propagating the neural network to calculate the output error, and then back propagating the error to the network to calculate the partial derivatives for each weight. These partial derivatives indicate the rate of change of the loss function with respect to each weight, thereby guiding how the weights are adjusted to reduce the overall error. In particular, for the fault paths resolved from the coding matrix F, the gradient back-propagation algorithm makes precise adjustments to those neural network weights that directly participate in the particular fault delivery path. This means that if a component failure is identified as part of a path that contributes highly to a particular failure event, then the neural network weights associated with that path will be preferentially adjusted to ensure that the model is more sensitive to the characteristics of such failure modes. For example, if the system identifies that the fault path is "fault a to component X to component W", the weights of the connecting components X and W in the corresponding neural network layer on the path will be adjusted according to the gradient back-propagation algorithm. By accurately calculating the effect of these weights on the loss function and updating it in the gradient descent direction, the algorithm can ensure that the neural network can produce a more accurate prediction when similar failure modes occur.

The adjustment of the loss function L is dynamically optimized according to the weight adjustment result, and involves the fine adjustment of a prediction error term, a constraint term based on component fault tree logic and a balance factor lambda in the loss function. The loss function can be expressed as:

L＝L_pred+λL_tree(P，F)，

The prediction error term L _pred plays a critical role in the loss function. The method directly quantifies the difference between the output of the neural network model and actual fault data, and reflects the performance of the model in terms of fault positioning accuracy. The design of this term is intended to ensure that the model is able to accurately predict the type and location of the fault of the weapon equipment with minimal error. The prediction error term L _pred is implemented using Mean Square Error (MSE), cross entropy loss, or other loss function that is appropriate for the particular task. For a weapon intelligent fault location system, which loss function to select depends on the nature of the fault data and the type of model output. If the fault localization problem is modeled as a regression problem, i.e., predicting a particular parameter value of the fault (e.g., temperature, pressure bias, etc.), a Mean Square Error (MSE) is used. The MSE calculates the average of the squares of the differences between the model predicted and actual values, given by:

Where y _i is the value of the ith actual failure data point, Is the model corresponding predictive value, n is the total number of data points.

In constraint terms λL _tree (P, F) of the component fault tree logic, P represents the fault probability distribution of model prediction, which is the prediction result given by the neural network for each possible fault type. F is the coding matrix of the component failure tree, in which the failure dependency relationship and failure propagation path between components are recorded in detail. The fault tree structure described by F is used to evaluate the logical dependencies between fault types in the model prediction. If the high probability prediction of a particular fault type is inconsistent with the dependency in F (i.e., if the model predicts that a component fails with a high probability, but that component does not directly lead to an advanced fault in the fault tree), the cost of the prediction is increased. For each failure mode, a different weight is assigned to adjust the value of λL _tree (P, F) depending on its location and path in the failure tree. This approach ensures that the model considers not only the probability of failure occurrence, but also the logical relationship between failures in the prediction. The estimated logical consistency cost is integrated into the main loss function L, and forms a final optimization target together with the prediction error term L _pred. Thus, the model may attempt to not only reduce the prediction error during the training process, but also to maintain the logical consistency of the prediction results. The innovative constraint design provides a powerful guide for the model, enabling it to understand and learn more deeply the complex failure modes and failure propagation mechanisms.

In some embodiments, the system first collects and aggregates up-to-date fault location results and corresponding historical labeling decision data according to preset time intervals (e.g., monthly, quarterly). These data represent the most recent fault diagnosis performance of the system and the previously accumulated fault handling experience, respectively. Next, the system performs a weighted average process on each type of fault type data, where the history label determination data may be weighted higher than the most recent fault location results to ensure that new adjustments can reflect the most current fault pattern and diagnostic hole findings while retaining the history effective experience. For example, if the most recent fault location results indicate a significant difference in frequency and condition of occurrence of a particular type of fault from historical data, the system may calculate an adjustment score for that type of fault that reflects possible changes in the fault pattern or improvement in fault diagnosis accuracy. After the adjustment scores of each fault type are obtained, the system further adjusts the value interval of the corresponding element in the fault tree coding matrix F according to the scores. This step is based on an assumption that: the dependency between failure modes and components may change over time, technological advances, or changes in equipment usage conditions. Therefore, by adjusting the coding matrix F, the system can dynamically reflect these changes, thereby improving the accuracy and effectiveness of fault diagnosis. Specifically, if the adjustment score for a fault type indicates an increase in the strength of the dependencies between components associated with the fault, the system will increase the element values representing those dependencies in the encoding matrix accordingly; conversely, if the adjustment score indicates that the dependency weakens, the system reduces the element values. Such adjustments are not limited to components on the direct fault propagation path, but also include indirect effects and potential new fault propagation paths. By the method, the fault tree coding matrix F is updated regularly, so that the current and expected fault modes can be reflected more accurately, and a more reliable logic basis is provided for fault prediction and fault positioning. The dynamic adjustment mechanism ensures the adaptability and long-term effectiveness of the intelligent fault positioning system of the weapon equipment, and improves the performance of the system in practical application.

It should be noted that if the adjustment score of a fault type exceeds a preset threshold, this indicates that the fault type occurs more frequently than expected in the most recent fault location practice, or that the fault type is more characterized than the historical data. At this point, the system calculates the difference between the adjustment score and the preset threshold, and generates a first bias based thereon. The bias reflects the degree to which the description of the relationships between the corresponding components in the fault tree needs to be enhanced. By increasing the interval of values of the relevant elements in the fault tree coding matrix F, the first bias enhances the sensitivity of the model to these frequent or significant fault types to ensure that the system can more accurately identify and respond to these changes. The adjustment caused by the first bias gives the system a higher alertness in the face of such faults, helping to quickly identify and handle such faults that may affect the stability and safety of the system. The second bias reduces the sensitivity of the model to these types of faults that are no longer so frequent or significant by reducing the interval of values of the relevant elements in the fault tree coding matrix F, avoiding excessive reactions of the model to such faults. The adjustment of the second bias helps to prevent false alarms of the model on certain fault types, ensure the accuracy of fault location and the effective allocation of resources, and avoid unnecessary maintenance cost and operation interference. In general, the first bias and the second bias together form a dynamic adjustment mechanism, so that the intelligent fault positioning system of the weapon equipment can flexibly adjust the representation of the fault tree according to the change of actual fault data, thereby maintaining high-efficiency and accurate fault diagnosis performance. This mechanism ensures that the system is able to adapt to the evolution of the failure mode, respond to new failure trends in time, while avoiding excessive sensitivity to less critical failures.

The above embodiments are merely illustrative of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention, and various modifications and improvements made by those skilled in the art to the technical solution of the present invention should fall within the scope of protection defined by the claims of the present invention without departing from the spirit of the design of the present invention.

Claims

1. An intelligent fault location system for a weapon, comprising: the system comprises a data acquisition module, a fault correlation module, a fault positioning module and an intelligent learning module, wherein the data acquisition module, the fault correlation module, the intelligent learning module and the fault positioning module are sequentially connected, and the data acquisition module is connected with the fault positioning module;

L＝L_pred+λL_tree(P，F)，

2. The intelligent fault location system of a piece of equipment as claimed in claim 1, wherein the operational data of the piece of equipment includes temperature readings, pressure values, vibration data and current and voltage measurements under different operations.

3. The intelligent fault location system for a piece of weapon equipment according to claim 1, wherein the generating a component fault tree according to the history labeling judgment data comprises:

4. The intelligent fault location system of a weapon equipment of claim 1, wherein converting the operation data and the history marking judgment data of the weapon equipment into the characteristic point values for matrix input comprises:

5. The intelligent fault location system for a weapon equipment according to claim 1, wherein the configuration of the weight matrix according to the component fault tree comprises the steps of:

6. The intelligent fault location system for a weapon equipment according to claim 1, wherein the device fault tree coding matrix F is adjusted in timing according to the fault location result and the history marking judgment data, and comprises:

7. The intelligent fault location system of a weapon equipment of claim 6, wherein the fault location result is weighted average with historical labeling decision data according to a preset time interval, wherein the historical labeling decision data has a weight greater than the fault location result weight.

8. The intelligent fault location system for a weapon equipment according to claim 6, wherein the adjusting the value interval of each element in the parameter fault tree coding matrix F according to the adjustment score of each fault type comprises: