Equipment operation effect prediction method and device

Abstract

The invention provides an equipment operation effect prediction method and device, and relates to the technical field of computers. The method specifically comprises the following steps: acquiring collected data which corresponds to equipment to be evaluated and is related to a predicted equipment operation effect, and classifying the collected data to obtain a plurality of target data sets; the multiple target data sets are input into reference prediction models corresponding to the data categories of the target data sets, multiple reference prediction indexes are obtained, and the multiple reference prediction models are the same in network structure; the multiple reference prediction indexes are input into a target prediction model, the target prediction index is obtained, and the target prediction model is a model which is the same as the reference prediction model in network architecture and different from the reference prediction model in network structure; and determining a target prediction result corresponding to the to-be-evaluated equipment according to the collected data and the target prediction index. The problem of low accuracy of an equipment operation effect prediction method in the prior art is solved, and the accuracy of the equipment operation effect prediction method is improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, specifically to a method and apparatus for predicting the operating effect of equipment. Background Technology

[0002] With the continuous development and changes of modern technology, equipment has undergone four generations of development over half a century, and its technical performance has been significantly improved. It has become an important tool. The prediction of equipment operation effect is essentially to identify potential problems in advance, optimize resource allocation, and improve system stability through data-driven methods. In the process of operation simulation, planning, and operation scheme design, the prediction of equipment operation effect has become a key process.

[0003] Existing methods for predicting equipment operating performance mainly include expert review, statistical analysis, and mathematical analysis. However, expert review relies heavily on the experience and reasoning abilities of experts, resulting in a high degree of subjectivity. Statistical analysis requires analyzing large amounts of data for summarization, which becomes difficult to implement when there are many influencing factors and their relationships are extremely complex. All of these methods have shortcomings in handling equipment performance prediction. Therefore, existing technologies have incorporated general neural networks and support vector machines into predictive analysis. However, for predictions involving different combinations of influencing factors, the current prediction methods assign different weights to different factors to differentiate their impact on operating performance. Nevertheless, the accuracy of equipment operating performance predictions obtained using this method still needs improvement.

[0004] There is currently no effective technical solution to the problem of low accuracy in existing methods for predicting equipment operating effects. Summary of the Invention

[0005] The embodiments described herein provide a method and apparatus for predicting equipment operating performance, in order to address the problem of low accuracy in existing equipment operating performance prediction methods.

[0006] According to a first aspect of this disclosure, a method for predicting the operating effect of equipment is provided, comprising: acquiring collected data related to the predicted operating effect of the equipment to be evaluated, and classifying the collected data to obtain multiple target datasets; inputting the multiple target datasets into reference prediction models corresponding to the data categories of the target datasets respectively to obtain multiple reference prediction indices, wherein the multiple reference prediction models have the same network structure; inputting the multiple reference prediction indices into a target prediction model to obtain a target prediction index, wherein the target prediction model is a model with the same network architecture as the reference prediction model but a different network structure; and determining the target prediction result corresponding to the equipment to be evaluated based on the collected data and the target prediction index.

[0007] According to a second aspect of this disclosure, a device for predicting equipment operation performance is provided, comprising: an acquisition unit, configured to acquire collected data related to the predicted equipment operation performance corresponding to the equipment to be evaluated, and classify the collected data to obtain multiple target datasets; a first input unit, configured to input the multiple target datasets into reference prediction models corresponding to the data categories of the target datasets respectively, to obtain multiple reference prediction indices; a second input unit, configured to input the multiple reference prediction indices into a target prediction model, to obtain a target prediction index; and a determination unit, configured to determine the target prediction result corresponding to the equipment to be evaluated based on the collected data and the target prediction index.

[0008] Optionally, the first input unit includes a target dataset receiving unit, used to enable the reference prediction model to receive a target dataset of the corresponding category, wherein the target dataset includes multiple target data; a first feature extraction unit, used to perform first feature extraction on the multiple target data based on the data characteristics of the multiple target data to obtain multiple first extracted data; a second feature extraction unit, used to perform second feature extraction on the multiple first extracted data based on the data characteristics of the multiple first extracted data to obtain multiple second extracted data; and a reference prediction index generation unit, used to pair and fuse the multiple second extracted data to obtain multiple reference prediction indices.

[0009] Optionally, the first feature extraction unit includes a target data decomposition function determination unit, used to determine the decomposition function corresponding to each target data based on the data characteristics of multiple target data; a target data decomposition unit, used to decompose each target data using the decomposition function corresponding to each target data to obtain multiple target high-frequency data and multiple target low-frequency data; and a first extracted data generation unit, used to pair and fuse multiple target high-frequency data and multiple target low-frequency data according to the weight pairs corresponding to the pairing and fusion to obtain multiple first extracted data.

[0010] Optionally, the second feature extraction unit includes a first extraction data decomposition function determination unit, used to determine the decomposition function corresponding to each type of first extraction data based on the data characteristics of multiple types of first extraction data; and a first extraction data decomposition unit, used to decompose each type of first extraction data using the decomposition function corresponding to each type of first extraction data to obtain multiple types of second extraction data including multiple types of first high-frequency data and multiple types of first low-frequency data.

[0011] Optionally, the first input unit further includes a sample data classification unit, used to collect sample data related to the predicted operating effect of the device to be evaluated, and classify the sample data to obtain multiple target sample sets; a current sample set determination unit, used to determine the number of sample dimensions in the multiple target sample sets respectively, and determine the target sample set with the most sample dimensions as the current sample set; an initial prediction model construction unit, used to construct an initial prediction model corresponding to the current sample set based on the current sample set, and determine the decomposition function corresponding to the edge of the initial prediction model based on the data characteristics of the current samples included in the current sample set; a first reference prediction model generation unit, used to iteratively train the initial prediction model based on the data characteristics of the samples included in the current sample set and multiple decomposition functions, so as to adjust the weights in the initial prediction model and obtain a reference prediction model corresponding to the current sample set; and a second reference prediction model generation unit, used to iteratively train the reference prediction model corresponding to the current sample set in combination with the target sample sets other than the current sample set, to obtain the reference prediction models corresponding to each of the target sample sets other than the current sample set.

[0012] Optionally, the second reference prediction model generation unit includes a target sample decomposition function determination unit, used to determine the decomposition function corresponding to the model edge in the reference prediction model based on the data characteristics of the target samples included in the target sample sets other than the current sample set; and a reference prediction model training unit, used to iteratively train the reference prediction model based on the data characteristics of the target samples included in the target sample sets other than the current sample set and the determined multiple decomposition functions, so as to adjust the weights in the reference prediction model and obtain the reference prediction model corresponding to each target sample set.

[0013] Optionally, the above-mentioned sample data classification unit further includes a sample set expansion unit, which is used to expand each sample set according to the sample data characteristics corresponding to each target sample set by using a sample expansion model corresponding to the data characteristics, and to determine the expanded sample set as the target sample set.

[0014] Optionally, the second input unit further includes a training sample acquisition unit for acquiring training samples corresponding to the target prediction model and determining the training sample characteristics; a training sample expansion unit for expanding the training samples based on the sample expansion model corresponding to the training sample characteristics and determining the expanded samples as training samples; a model to be trained construction unit for constructing a model to be trained based on the training samples and determining the decomposition function corresponding to the model edge of the model to be trained based on the training sample characteristics; and a target prediction model generation unit for iteratively training the model to be trained based on the training sample characteristics and the decomposition function to adjust the fusion weights in the model to be trained and obtain the target prediction model.

[0015] According to a third aspect of this disclosure, a computer-readable storage medium is provided, which stores computer instructions for causing a computer to execute the above-described method for predicting the operating effect of a device.

[0016] The above-mentioned method for predicting equipment operation results can improve the accuracy of equipment operation result prediction and solve the problem of low accuracy in existing equipment operation result prediction methods. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the hardware environment for an optional device performance prediction method according to an embodiment of the present invention; Figure 2 This is a flowchart of an optional device operation effect prediction method according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an optional equipment operation effect prediction device according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an optional electronic device according to an embodiment of the present invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0020] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or devices. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] To address the low accuracy of existing equipment operation performance prediction methods, this application provides a method for predicting equipment operation performance. As an optional implementation, the above-mentioned method can be applied, but is not limited to, to methods such as... Figure 1 The device operation effect prediction system shown consists of terminal device 102 and server 104. For example... Figure 1 As shown, terminal device 102 is connected to server 104 via network 110. Network 110 may include, but is not limited to, wired networks and wireless networks. Terminal device 102 is also equipped with a display 106, a processor 108, and a memory 112. The display 106 can be used to display target prediction results, etc. The processor 108 can be used to process data such as requests for device performance, etc. The memory 112 can be used to store relevant information related to this application. Server 104 can be a single server, a server cluster composed of multiple servers, or a cloud server. Server 104 includes a database 114 and a processing engine 116. The database 114 can be used to store target data sets, sample data, training samples, etc., and the processing engine 116 is used to process the stored target data sets, sample data, training samples, etc.

[0022] According to one aspect of the present invention, the above-described device operation effect prediction system may further perform the following steps: First, the terminal device 102 executes S102, sending a device operation effect prediction request to the server 104 via the network 110; then, the server 104 executes as follows: Figure 1 As shown in S104 to S110, the prediction of equipment operation effect is completed.

[0023] To facilitate a detailed explanation of the content of this application, the following section first introduces an important concept involved in this application: the wavelet Kolmogorov-Arnold network (Wav-KAN) architecture. The Wavelet Kolmogorov-Arnold Network (Wav-KAN) architecture is an innovative neural network architecture that combines wavelet transform and Kolmogorov-Arnold network (KAN). Wav-KAN incorporates wavelet functions used for wavelet transform into the Kolmogorov-Arnold network architecture: by replacing the weights of the Kolmogorov-Arnold network architecture with wavelet functions, on the one hand, the network can adapt to the data structure, thereby improving accuracy, speeding up training, and enhancing robustness; on the other hand, the network can effectively capture high-frequency and low-frequency components of the input data, enabling multi-resolution data analysis and high-precision prediction.

[0024] As an optional implementation method, please refer to, for example... Figure 2 The flowchart shown illustrates a method for predicting equipment performance. This method may include the following steps: S202, acquire the collected data related to the predicted operating effect of the device to be evaluated, and classify the collected data to obtain multiple target datasets; S204, input multiple target datasets into reference prediction models corresponding to the data categories of the target datasets respectively to obtain multiple reference prediction indices, wherein the network structures of the multiple reference prediction models are the same; S206, Input multiple reference prediction indices into the target prediction model to obtain the target prediction index, wherein the target prediction model is a model with the same network architecture as the reference prediction model but a different network structure. S208 determines the target prediction result corresponding to the device to be evaluated based on the collected data and the target prediction index.

[0025] It should be noted that the object whose operational effect needs to be predicted is the equipment to be evaluated. The collected data includes both data describing the equipment itself, such as static equipment model and equipment maintenance records, and dynamic data such as temperature and engine speed during equipment operation, as well as data describing the external environment of the equipment, such as ambient temperature, air pressure, and lighting conditions during equipment operation. The amount of data collected is set according to actual needs, including but not limited to tens of thousands to millions of data points.

[0026] Optionally, after obtaining the above-mentioned collected data, the data can be preprocessed as needed, including but not limited to normalization, missing value handling, and outlier detection.

[0027] In addition, the data collection can be classified from different dimensions. For example, it can be divided into three categories based on three dimensions: static characteristics, dynamic behavior, and external interaction environment: device attribute data, dynamic operation data, and environmental data. It should be noted that, without classification, the number of detailed data categories may range from several to hundreds. The classification in this application is essentially a large-scale classification of a large amount of data (several or even hundreds of categories) related to the prediction of the operating effect of the device.

[0028] The reference prediction model and the target prediction model mentioned above use the same network architecture, including but not limited to the wavelet Kolmogorov-Arnold Networks (Wav-KAN), i.e., the Wav–KAN network architecture.

[0029] The Kolmogorov-Arnold Network (KAN) in the Wav-KAN network architecture represents a new form of neural network design that challenges traditional concepts such as the Multilayer Perceptron (MLP). At the heart of the KAN is an elegant yet somewhat abstract mathematical theorem proposed by Kolmogorov and Arnold: the Kolmogorov-Arnold Representation Theorem.

[0030] Kolmogorov-Arnold Representation Theorem: Any multivariable continuous function can be decomposed into a superposition of a finite number of single-variable functions; that is, for any function defined on a cube... On Any continuous function of variables ,exist Functions and Functions All of these functions are single-variable and continuous, making

[0031] This theorem states that any multivariate continuous function can essentially be decomposed into the sum of univariate functions. These internal functions are univariate, meaning they each accept only one input. Based on this theorem, this application constructs a KAN network and replaces the weights in the KAN network with wavelet functions, i.e.: using... replace ,use replace Thus, the core of the Wav-KAN network is obtained:

[0032] As the above analysis shows, the Wav-KAN network replaces linear weights with learnable univariate functions placed on edges rather than nodes. Within nodes, only some univariate functions from the previous layer are summed. Each function is adjustable, allowing the network to learn activations and transformations of the input. This change improves the flexibility and accuracy of the network architecture. Furthermore, because each function has a specific and easily understood role, it is easier to see what the model is "thinking," thus increasing the interpretability of the Wav-KAN network. Simultaneously, since Wav-KAN can fit complex patterns in data more accurately with fewer parameters, it exhibits good data adaptability. Moreover, during backpropagation, Wav-KAN can calculate the gradients of univariate functions, thereby updating and minimizing the loss function. Therefore, Wav-KAN is more effective in learning complex and high-dimensional functions.

[0033] While the reference prediction model and the target prediction model use the same network architecture (Wav-KAN network architecture), their network structures differ due to differences in the number of nodes (which can be understood as the number of variables in the Wav-KAN network) and the number of model edges (which can be understood as the number of weights in the Wav-KAN network). However, for the various reference prediction models mentioned above, although the weights assigned to each data point differ within each model, the overall number of nodes and model edges is the same; therefore, the network structures of these multiple reference prediction models are identical.

[0034] The aforementioned reference prediction index and target prediction index are both specific numerical values, one or more. For example, one of the reference prediction indices (or target prediction indices) might be calculated as 0.88, with the index range set to 0-1. A higher index indicates a better predicted performance for the device. Specifically, for the reference prediction index, the reference prediction index output by the reference prediction model for each category of target dataset indicates the predicted device performance from the perspective of that category of target dataset. In other words, multiple reference prediction indices represent predictions of the device from different perspectives. For example, a reference prediction index of 0.5 for the device attribute dataset and 0.8 for the environment dataset indicates that, from the perspective of device attributes, the predicted performance of the device is average (e.g., there may be problems such as outdated equipment models leading to low efficiency), while from the perspective of the external environment, the predicted performance is good (e.g., the current temperature allows the device to operate at its optimal state). The target prediction index, on the other hand, is a comprehensive prediction of the device's performance, combining predictions from different perspectives. For example, a target prediction index of 0.88 indicates that, from an overall perspective, the expected performance of the device is excellent.

[0035] Optionally, after obtaining the target prediction result, the various parts of the above-mentioned equipment operation effect prediction process can be optimized and updated using a loss function and an optimizer. The loss function includes, but is not limited to, cross-entropy or mean squared error, and the optimizer includes, but is not limited to, the AdamW optimizer.

[0036] To facilitate understanding, a Boeing aircraft will be used as the device to be evaluated, and the solution in this application will be explained using the example of "predicting the operating effect of an aircraft".

[0037] According to step S202 above: data for the Boeing aircraft is acquired from various sources, including the aircraft's flight management system, engine indication and crew alarm system, atmospheric data system, and aviation meteorological information system. This acquired data may or may not affect the aircraft's operational performance. The acquired data specifically includes: flight altitude data, engine speed data, pilot acceleration or turning data, outside temperature data, atmospheric parameters, etc. The number of data can be set from tens of thousands to millions of records depending on actual needs. This acquired data is divided into three categories based on three dimensions: static characteristics, dynamic behavior, and external interaction environment: equipment attribute data, dynamic operation data, and flight environment data. These three categories of data form three target datasets: equipment attribute dataset, dynamic operation dataset, and flight environment dataset. The equipment attribute dataset includes data such as the model, usage time, and maintenance records of various aircraft equipment; the dynamic operation dataset includes data such as engine speed, fuel flow, and pitch angle; and the flight environment dataset includes data such as air pressure, wind speed, humidity, and ambient temperature. Each of the above three data categories corresponds to a reference prediction model, i.e., a sub-prediction network model.

[0038] According to step S204 above: input the dynamic operation class dataset into the corresponding reference prediction model to obtain the corresponding reference prediction index 0.88 (the index range is set to 0-1, the higher the index, the better the dynamic operation status of the equipment, and thus the better the dynamic operation class data indicates the aircraft's operation effect). Input the equipment attribute class dataset into the corresponding reference prediction model to obtain the corresponding reference prediction index 0.92. And so on, input each target dataset into different reference prediction models to obtain multiple reference prediction indices. Each reference prediction index indicates the aircraft's operation effect from a different perspective.

[0039] According to step S206 above: the reference prediction indexes such as the dynamic operation reference prediction index of 0.88 and the equipment attribute reference prediction index of 0.92 are input into the target prediction model. The target prediction model summarizes these reference prediction indices and comprehensively analyzes the evaluation results of the aircraft operation effect from different perspectives to obtain a target prediction index of 0.90 (close to 1, indicating that the overall prediction effect of aircraft operation is relatively good).

[0040] According to step S208 above: the initial collected data is combined with the target prediction index and interpreted by the system. For example, if the collected data shows that the engine temperature is slightly higher when the aircraft is flying in the troposphere, this situation will often lead to an increase of 1% in the aircraft's fuel consumption rate. Combined with the target prediction index of 0.9, the target prediction result can be obtained: the overall operation of the aircraft is predicted to be good, but because the engine is not in optimal condition, it will lead to an increase of 1% in the fuel consumption rate.

[0041] The above-mentioned equipment operation effect prediction method can achieve more accurate prediction of equipment operation effect, solving the problem of low accuracy in existing equipment operation effect prediction methods.

[0042] As an optional implementation, multiple target datasets are input into reference prediction models corresponding to the data categories of the target datasets to obtain multiple reference prediction indices, including: S121: The reference prediction model receives the target dataset of the corresponding category, which includes multiple target data. S122: Based on the data characteristics of multiple target data, perform first feature extraction on multiple target data to obtain multiple first extracted data; S123: Based on the data characteristics of multiple first extracted data, perform second feature extraction on multiple first extracted data to obtain multiple second extracted data; S124: Pair and fuse multiple second-extracted data to obtain multiple reference prediction indices.

[0043] The “data characteristics” in the above-mentioned “data characteristics of target data” and “data characteristics of first extracted data” refer to the inherent features of the data, which can be used to guide the subsequent data processing. Data characteristics include, but are not limited to, statistical characteristics, time domain characteristics, frequency domain characteristics, etc.

[0044] It should be noted that S122 above: Based on the data characteristics of multiple target data, first feature extraction is performed on multiple target data to obtain multiple first extracted data, including: determining the decomposition function corresponding to each target data based on the data characteristics of multiple target data; using the decomposition function corresponding to each target data to decompose each target data to obtain multiple target high-frequency data and multiple target low-frequency data; pairing and fusing the multiple target high-frequency data and multiple target low-frequency data according to the weight pairs corresponding to the pairing and fusion to obtain multiple first extracted data, wherein one target high-frequency data is paired and fused with one target low-frequency data, and different pairing and fusion processes correspond to different weight pairs.

[0045] In addition, S123 above: Based on the data characteristics of multiple first extracted data, perform second feature extraction on multiple first extracted data to obtain multiple second extracted data, including: based on the data characteristics of multiple first extracted data, determine the decomposition function corresponding to each first extracted data; use the decomposition function corresponding to each first extracted data to perform data decomposition on each first extracted data to obtain multiple second extracted data including multiple first high-frequency data and multiple first low-frequency data.

[0046] For the above data (target data and first extracted data) decomposition process, the tools used include, but are not limited to, the Discrete Wavelet Transform (DWT) in wavelet functions. The wavelet functions (wavelet basis functions) used in the Discrete Wavelet Transform (DWT) are basis functions used for time-frequency analysis, including but not limited to: Haar wavelet, Daubechies wavelet, Biorthogonal wavelet, etc. Different wavelet basis functions have different characteristics, such as symmetry, support length (the range of non-zero values ​​of the wavelet function in the time domain), etc. Therefore, different wavelet basis functions can be selected for data with different data characteristics. For example, for data with many abrupt changes, the Haar wavelet can be selected to capture the abrupt changes, and for relatively smooth data, the Daubechies wavelet can be selected to reflect the trend. It should be noted that since wavelet decomposition functions are particularly effective in capturing the characteristics of data, when performing data decomposition, selecting an appropriate wavelet decomposition function according to the data characteristics has an important impact on the prediction results of equipment performance.

[0047] After dynamically selecting wavelet basis functions based on the different characteristics of the data, the discrete wavelet transform can efficiently decompose the data into high-frequency (local details) and low-frequency (global trends) components, thereby enabling multi-resolution analysis and providing a more comprehensive interpretation of the data's structure and characteristics. Furthermore, when using discrete wavelet transform for wavelet analysis, once the initial set of coefficients is calculated, it is not necessary to recalculate them when calculating additional coefficients at higher resolutions. This efficiency stems from the fact that the wavelet transform uses previously calculated coefficients to generate more details, ensuring that early calculation results remain valid and reusable, simplifying the process and significantly reducing computational overhead.

[0048] Regarding the Discrete Wavelet Transform (DWT) described above: Unlike the Continuous Wavelet Transform (CWT), which analyzes signals at every possible scale and displacement, DWT uses discrete sampling intervals, making it suitable for digital applications. When data points are close together (high sampling rate), the wavelet basis can be amplified to capture high-frequency details (fine details) in that local region. When data points are scattered (low sampling rate), the wavelet basis can be scaled down to capture low-frequency trends (overall shape) using global information. This invention examines some examples of uniformly and irregularly sampled data. In fact, by using wavelets, this invention combines local details in areas of dense data points with broader trends in areas of sparse data points.

[0049] The principle of DWT is to obtain wavelet basis functions (a set of functions) by scaling and translating a single mother wavelet. These wavelet basis functions are orthogonal and normalized. Through wavelet basis functions, signals can be represented in a compact and non-redundant manner. Furthermore, by iteratively applying high-pass and low-pass filters to the signal, DWT can generate approximation coefficients and detail coefficients at different resolution levels, thereby realizing multi-resolution analysis.

[0050] set up For the mother wavelet, This is a scaling function. The signal. In the Layer decomposition into approximate coefficients and detail coefficient As shown below:

[0051]

[0052] in: It is a discrete signal.

[0053] The scale is Location is The scaling function over time.

[0054] The scale is Location is The wavelet function at time.

[0055] This represents the approximation coefficient.

[0056] This represents the detail factor.

[0057] The original signal can be reconstructed from its wavelet coefficients using the inverse discrete wavelet transform (DWT), ensuring no information loss during the transformation process. This reconstruction is given by the following formula:

[0058] in It refers to the number of decomposition layers.

[0059] As can be seen from the above analysis, by using DWT, approximations and details can be created at different resolutions. This process provides a hierarchical framework in which each level captures different frequency components of the signal, making it easier to analyze local features and transient phenomena. This enables the present application to achieve multi-resolution (MRA) analysis, which can adapt to the irregular sampling and missing data that often exist in device performance prediction, as well as complex factor interactions and non-stationary signal characteristics. More detailed information is used where there are more data points, and less detailed information is used where there are fewer data points.

[0060] The pairing and fusion of the target high-frequency data and target low-frequency data, as well as the pairing and fusion of the first high-frequency data and the first low-frequency data, are achieved through the Wav-KAN network architecture. In the Wav-KAN architecture, the shape and parameters of the wavelet function (such as the translation / scaling factor of the Daubechies wavelet) are optimized through training to obtain learnable wavelet functions. These learnable wavelet functions are then used as activation functions (such as Daubechies and Haar bases) to dynamically adapt to the data characteristics. The outputs of the input wavelet functions are summed through nodes, and the high-frequency and low-frequency components generated by the wavelet transform are dynamically integrated to form a high-order feature combination, thereby achieving multi-resolution fusion. This multi-resolution dynamic fusion mechanism can adaptively adjust the weights of high-frequency / low-frequency components according to the data density. In data-dense regions, it enhances the capture of high-frequency details (such as instantaneous equipment fault signals), while in data-sparse regions, it relies on low-frequency trend prediction (such as long-term performance degradation).

[0061] In summary, the Wav-KAN network architecture replaces traditional weights with learnable wavelet functions for its edges. Each edge corresponds to a univariate wavelet activation function, and these wavelet functions are summed through nodes to form an interpretable combination of nonlinear high-order features. Furthermore, through multi-resolution dynamic fusion, the weights of high-frequency / low-frequency components can be adaptively adjusted according to data density. For example, in data-dense regions, high-frequency detail capture is enhanced (such as instantaneous equipment fault signals), while in data-sparse regions, low-frequency trend prediction is relied upon (such as long-term performance degradation), thereby achieving high-precision prediction. By combining the advantages of wavelet transform and the Kolmogorov-Arnold representation theorem, the adaptability, accuracy, and interpretability of the prediction model are improved.

[0062] Taking the example of "predicting the operational performance of an aircraft" above, and using the reference prediction model for dynamic operational data as an example, the dynamic operational target dataset received by the reference prediction model includes multiple target data points: during the aircraft's climb, the corresponding engine speed N1, engine exhaust temperature EGT, fuel flow FF, etc., at different altitudes and when the pilot performs acceleration, turns, etc. The above data is decomposed using a wavelet function to obtain the high-frequency part (target high-frequency data) and low-frequency part (target low-frequency data) of each data point, such as the high-frequency and low-frequency parts of the aircraft's flight altitude, the high-frequency and low-frequency parts of the engine speed, etc. Different weights are assigned to the target high-frequency data and target low-frequency data obtained by wavelet decomposition for fusion. For example, fusing the target low-frequency data corresponding to the flight altitude with the target low-frequency data corresponding to the engine speed can yield the first extracted data. This first extracted data can indicate that when the aircraft climbs to the 8km position, the engine speed increases rapidly. After obtaining multiple first-extracted data through the above pairing and fusion process, a corresponding wavelet decomposition function is determined based on the characteristics of each first-extracted data. Then, using the wavelet decomposition function corresponding to each first-extracted data, the multiple first-extracted data are further decomposed according to different frequency components, resulting in multiple second-extracted data, including multiple first-high-frequency data and multiple first-low-frequency data. For example, the first-high-frequency data indicates a short-term phenomenon where the engine speed rapidly increases while experiencing millisecond-level violent vibrations at the moment the aircraft climbs to 8km. The first-low-frequency data indicates a long-term trend where the engine speed slowly decreases and eventually stabilizes during the transition from the climb phase to the cruise phase. Further pairing and fusion of these multiple second-extracted data yields a reference prediction index of 0.88 for the dynamic operation target data. Through the above two data decompositions (decomposition of target data and first-extracted data) and two data pairing and fusions (pairing and fusion of target high-frequency data and target low-frequency data, and pairing and fusion of first-high-frequency data and first-low-frequency data), more subtle and deeper patterns and regularities within complex data can be uncovered, improving the accuracy of equipment operation effect prediction results.

[0063] As an optional implementation, before inputting multiple target datasets into reference prediction models corresponding to the data categories of the target datasets to obtain multiple reference prediction indices in S206, the method further includes: S131, collecting sample data related to the predictive effect of the device to be evaluated, and classifying the sample data to obtain multiple target sample sets; S132, determining the number of sample dimensions in each of the multiple target sample sets, and determining the target sample set with the most sample dimensions as the current sample set, wherein the number of sample dimensions is the number of sample dimensions included in each target sample in the target sample set; S133, constructing an initial prediction model corresponding to the current sample set based on the current sample set, and determining the decomposition function corresponding to the edge of the initial prediction model based on the data characteristics of the current samples included in the current sample set; S134, iteratively training the initial prediction model based on the data characteristics of the samples included in the current sample set and the multiple decomposition functions to adjust the weights in the initial prediction model, thereby obtaining a reference prediction model corresponding to the current sample set; S135, iteratively training the reference prediction model corresponding to the current sample set in combination with target sample sets other than the current sample set, thereby obtaining reference prediction models corresponding to each of the target sample sets other than the current sample set.

[0064] In step S131 above, the classification method for sample data can refer to the classification in S202 above. Starting from different dimensions, for example, based on three dimensions of static features, dynamic behavior and external interaction environment, the target sample sets are divided into three categories: device attribute target sample set, dynamic operation target sample set, and environment target sample set.

[0065] The "sample dimension" in step S132 above refers to the types of data included in the sample data. For example, when predicting the aircraft's operating status, a sample data point in the dynamic operation target sample set might be: at a certain moment during the climb phase, the aircraft is turning at an altitude of 8km, with engine speed N1=90%, engine exhaust temperature EGT=900℃, and fuel flow rate FF=2300kg / h. This sample data includes five types of data: flight phase, flight altitude, engine speed, engine exhaust temperature, and fuel flow rate—that is, five sample dimensions. The sum of the sample dimensions of all sample data in the target sample set is the "number of sample dimensions." The target sample set with the highest number of sample dimensions is chosen as the current sample set because it contains the most types of data and the most comprehensive information.

[0066] In step S133 above, the decomposition function determined based on the data characteristics of the current samples included in the current sample set is the wavelet decomposition function mentioned in steps S122 and S123 above. The properties and determination method of the wavelet decomposition function will not be elaborated here.

[0067] The implementation of step S135 above can be achieved by: determining the decomposition function corresponding to the model edge in the reference prediction model based on the data characteristics of the target samples included in the target sample sets other than the current sample set; iteratively training the reference prediction model based on the data characteristics of the target samples included in the target sample sets other than the current sample set and the determined multiple decomposition functions to adjust the weights in the reference prediction model, thereby obtaining the reference prediction model corresponding to each target sample set. The network structure of each reference prediction model is the same. For example, after the dynamic operation target sample set in the device attribute class target sample set, dynamic operation class target sample set, and environment class target sample set is determined as the current sample set and the dynamic operation class reference prediction model is obtained through iterative training, the current dynamic operation class reference prediction model is iteratively trained using the other two target sample sets to obtain the device attribute class reference prediction model and the environment class reference prediction model. The generation process of the other two reference prediction models can be implemented in a multi-threaded parallel manner. Since the other two reference prediction models are obtained by iterative training based on the dynamic operation class reference prediction model, the number of nodes and model edges in the three reference prediction models are the same, but the weights used by each model are different.

[0068] Taking the example of "predicting the operational performance of an aircraft" above, we first collect sample data related to the aircraft's operational performance, such as sample data collected during the aircraft's flight over the past month. This sample data is then categorized into three target sample sets: equipment attribute target sample set, dynamic operation target sample set, and flight environment target sample set. Taking the dynamic operation target sample set as an example, one sample in this set is: at a certain moment during the climb phase, the aircraft is turning at an altitude of 8km. At this time, the engine speed N1=90%, the engine exhaust temperature EGT=900℃, and the fuel flow rate FF=2300kg / h. In other words, this sample includes flight phase, flight altitude, engine speed, and engine... The sample set includes five dimensions: engine exhaust temperature and fuel flow. The total number of sample dimensions in the dynamic operation target sample set is obtained by summing the number of sample dimensions in each sample. Similarly, the number of sample dimensions in other target sample sets is calculated. The target sample set with the most sample dimensions is then designated as the current sample set. Assuming that the dynamic operation target sample set is the current sample set, an initial prediction model is constructed based on it. The wavelet decomposition function in the initial prediction model is determined using the dynamic operation target sample data. The initial prediction model is then iteratively trained to adjust the weights, resulting in a reference prediction model corresponding to the dynamic operation target sample set.

[0069] Based on the reference prediction model corresponding to the dynamic operation target sample set, the wavelet function in the reference prediction model is determined according to the data characteristics of the sample data in other target sample sets: equipment attribute target sample set and flight environment target sample set. The reference prediction model is iteratively trained according to the sample data of each type of target to obtain the other two reference prediction models. Since the generation of the other two reference prediction models is based on the reference prediction model corresponding to the dynamic operation target sample set, the number of nodes (such as engine speed node and air pressure node) and the number of edges in the above three reference prediction models are the same, but the weight of each data in the network structure of each reference prediction model is different.

[0070] It should be noted that after collecting sample data related to the predicted operating effect of the device corresponding to the device to be evaluated and classifying the sample data to obtain multiple target sample sets, the process further includes: according to the characteristics of the sample data corresponding to each target sample set, using a sample augmentation model corresponding to the data characteristics to augment each sample set, and determining the augmented sample set as the target sample set. The sample augmentation model mentioned above includes, but is not limited to, the Conditional Generative Adversarial Network (CGAN) model. The CGAN model includes a generator and a discriminator. The generator is used to generate data that is as close as possible to the real samples, and the discriminator is used to distinguish whether the input sample is real data or a fake sample from the generator, and outputs a probability value. During training, the generator improves the sample authenticity by deceiving the discriminator, and the discriminator improves its classification ability by distinguishing between real and fake samples, forming adversarial training, thereby achieving sample augmentation. Furthermore, the inputs of the generator and the discriminator include condition variables. The use of condition variables enables the CGAN model to generate samples of a specified type according to the condition variables. For example, in the above example of "predicting the operational performance of an aircraft", the flight altitude of 9km and the flight state of straight-line constant speed cruise are used as conditional variables. The conditional variables and the existing dynamic operation target sample data are then input into the conditional generative adversarial network. After adversarial training, the expanded dynamic operation target sample set can be obtained and used as the target sample set.

[0071] By expanding the target sample set using Conditional Generative Adversarial Networks (CGANs), problems such as overfitting caused by insufficient samples can be avoided, thereby improving the stability and reliability of the reference prediction model.

[0072] As an optional implementation, before inputting multiple reference prediction indices into the target prediction model to obtain the target prediction index in S208, the method further includes: S141, obtaining training samples corresponding to the target prediction model and determining the training sample characteristics; S142, expanding the training samples based on the sample expansion model corresponding to the training sample characteristics, and determining the expanded samples as training samples; S143, constructing a model to be trained based on the training samples, and determining the decomposition function corresponding to the model edge of the model to be trained based on the training sample characteristics; S144, iteratively training the model to be trained based on the training sample characteristics and the decomposition function to adjust the fusion weights in the model to be trained, thereby obtaining the target prediction model.

[0073] In other words, training samples corresponding to the target prediction model are first obtained, and the training samples are expanded using a sample expansion model corresponding to their characteristics. The sample expansion model includes, but is not limited to, the aforementioned Conditional Generative Adversarial Network (CGAN) model. The expanded samples are used as training samples to construct the model to be trained. The wavelet function corresponding to the model edge of the model to be trained is determined by the characteristics of the training samples (frequency composition, etc.). After determining the wavelet function, the model to be trained is iteratively trained using the characteristics of the training samples and the wavelet function, thereby adjusting the fusion weights of different nodes and different network layers in the model to be trained, and thus obtaining the target prediction model.

[0074] By expanding the training samples using Conditional Generative Adversarial Networks (CGANs), problems such as overfitting caused by insufficient samples can be avoided, thereby improving the stability and reliability of the target prediction model.

[0075] For example, in the above example of "predicting the operational performance of an aircraft", multiple reference prediction indices are input into the target prediction model. Before obtaining the target prediction index, some or all of the data related to the prediction of the aircraft's operational performance obtained during the past month's flights are used as training samples. The training samples are expanded and replaced using a conditional generative adversarial network (CGAN) model (the expanded training samples are determined as the training samples). The wavelet function of the model to be trained is determined based on the frequency characteristics of the training samples. The model to be trained is then iteratively trained using the frequency characteristics of the training samples and the determined wavelet function. After the iterative training is completed, the target prediction model can be obtained.

[0076] The above-mentioned equipment operation effect prediction method can achieve more accurate prediction of equipment operation effect, solving the problem of low accuracy in existing equipment operation effect prediction methods.

[0077] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0078] According to another aspect of the present invention, an apparatus for predicting equipment operating performance for implementing the above-described equipment operating performance prediction method is also provided, such as... Figure 3 As shown, the device includes: an acquisition unit 302, used to acquire collected data related to the predicted operating effect of the device to be evaluated, and classify the collected data to obtain multiple target datasets; a first input unit 304, used to input the multiple target datasets into reference prediction models corresponding to the data categories of the target datasets respectively, to obtain multiple reference prediction indices, wherein the multiple reference prediction models have the same network structure; a second input unit 306, used to input the multiple reference prediction indices into a target prediction model to obtain a target prediction index, wherein the target prediction model is a model with the same network architecture as the reference prediction model but a different network structure; and a determination unit 308, used to determine the target prediction result corresponding to the device to be evaluated based on the collected data and the target prediction index.

[0079] The specific methods of execution of each unit in the above device embodiments have been described in detail in the embodiments related to the method, and will not be elaborated further here.

[0080] According to another aspect of the present invention, an electronic device for implementing the above-described device operation effect prediction method is also provided. This electronic device may be as follows: Figure 4 The terminal device or server shown. This embodiment uses this electronic device as an example for illustration. Figure 4 As shown, the electronic device includes: at least one processor 404; and a memory 402 communicatively connected to the at least one processor 404; wherein the memory 402 stores a computer program executable by the at least one processor 404, the computer program being executed by the at least one processor 404 to cause the at least one processor 404 to perform the steps in any of the above method embodiments. The electronic device may be located in at least one of a plurality of network devices in a computer network. The processor may be configured to execute the above-described device operation effect prediction method via a computer program.

[0081] Alternatively, as those skilled in the art will understand, Figure 4The structure shown is for illustrative purposes only. Electronic devices can also be smartphones (such as Android phones, iOS phones, etc.), tablets, PDAs, mobile internet devices (MIDs), PADs, and other terminal devices. Figure 4 This does not limit the structure of the aforementioned electronic devices. For example, electronic devices may also include components that are more... Figure 4 The more or fewer components shown (such as network interfaces, etc.), or having the same Figure 4 The different configurations shown.

[0082] The memory 402 can be used to store software programs and modules, such as the program instructions / modules corresponding to the device operation effect prediction method and device in the embodiments of the present invention. The processor 404 executes various functional applications and data processing by running the software programs and modules stored in the memory 402, thereby realizing the above-mentioned device operation effect prediction method.

[0083] Those skilled in the art will understand that all or part of the processes in the above method embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes as described in the above method embodiments.

[0084] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting equipment operating performance, characterized in that, include: Acquire the collected data related to the predicted operating effect of the device to be evaluated, and classify the collected data to obtain multiple target datasets; Multiple target datasets are input into reference prediction models corresponding to the data categories of the target datasets to obtain multiple reference prediction indices, wherein the network structures of the multiple reference prediction models are the same; Multiple reference prediction indices are input into the target prediction model to obtain the target prediction index, wherein the target prediction model is a model with the same network architecture as the reference prediction model but a different network structure. The target prediction result corresponding to the device to be evaluated is determined based on the collected data and the target prediction index.

2. The method according to claim 1, characterized in that, Multiple target datasets are input into reference prediction models corresponding to the data categories of the target datasets to obtain multiple reference prediction indices, including: The reference prediction model receives a target dataset of the corresponding category, and the target dataset includes multiple target data. Based on the data characteristics of the various target data, a first feature is extracted from the various target data to obtain various first extracted data; Based on the data characteristics of various first extracted data, second feature extraction is performed on various first extracted data to obtain various second extracted data; Multiple types of the second extracted data are paired and fused to obtain multiple reference prediction indices.

3. The method according to claim 2, characterized in that, Based on the data characteristics of the various target data, a first feature is extracted from the various target data to obtain various first extracted data, including: Based on the data characteristics of the various target data, determine the decomposition function corresponding to each type of target data; Each type of target data is decomposed using a decomposition function corresponding to each type of target data to obtain multiple types of high-frequency target data and multiple types of low-frequency target data. Multiple target high-frequency data and multiple target low-frequency data are paired and fused according to the weight pairs corresponding to the pairing and fusion to obtain multiple first extracted data. Among them, one target high-frequency data is paired and fused with one target low-frequency data, and different pairing and fusion processes correspond to different weight pairs.

4. The method according to claim 2, characterized in that, Based on the data characteristics of various first extracted data, second feature extraction is performed on various first extracted data to obtain various second extracted data, including: Based on the various data characteristics of the first extracted data, a decomposition function corresponding to each type of first extracted data is determined; The first extracted data is decomposed using a decomposition function corresponding to each type of the first extracted data to obtain multiple types of second extracted data, including multiple types of first high-frequency data and multiple types of first low-frequency data.

5. The method according to claim 1, characterized in that, Before inputting the multiple target datasets into reference prediction models corresponding to the data categories of the target datasets to obtain multiple reference prediction indices, the method further includes: Collect sample data related to the predicted operating effect of the equipment corresponding to the equipment to be evaluated, and classify the sample data to obtain multiple target sample sets; The number of sample dimensions in each of the multiple target sample sets is determined, and the target sample set with the largest number of sample dimensions is determined as the current sample set, wherein the number of sample dimensions is the number of sample dimensions included in each target sample in the target sample set; Construct an initial prediction model corresponding to the current sample set based on the current sample set, and determine the decomposition function corresponding to the edge of the initial prediction model based on the data characteristics of the current samples included in the current sample set; The initial prediction model is iteratively trained based on the data characteristics of the samples included in the current sample set and multiple decomposition functions to adjust the weights in the initial prediction model and obtain the reference prediction model corresponding to the current sample set. The reference prediction model corresponding to the current sample set is iteratively trained by combining the target sample sets other than the current sample set to obtain the reference prediction models corresponding to each of the target sample sets other than the current sample set.

6. The method according to claim 5, characterized in that, Iteratively train the reference prediction model corresponding to the current sample set by combining it with target sample sets other than the current sample set, to obtain reference prediction models corresponding to each target sample set other than the current sample set, including: Determine the decomposition function corresponding to the model edge in the reference prediction model based on the data characteristics of the target samples included in the target sample set other than the current sample set; The reference prediction model is iteratively trained based on the data characteristics of the target samples included in the target sample sets other than the current sample set, and the determined multiple decomposition functions, so as to adjust the weights in the reference prediction model and obtain the reference prediction model corresponding to each target sample set.

7. The method according to claim 5, characterized in that, After collecting sample data related to the predicted operating effect of the equipment to be evaluated and classifying the sample data to obtain multiple target sample sets, the process further includes: Based on the characteristics of the sample data corresponding to each target sample set, a sample augmentation model corresponding to the data characteristics is used to augment each sample set, and the augmented sample set is determined as the target sample set.

8. The method according to claim 1, characterized in that, Before inputting the multiple reference prediction indices into the target prediction model to obtain the target prediction index, the process further includes: Obtain training samples corresponding to the target prediction model, and determine the training sample characteristics of the training samples; The training samples are augmented using a sample augmentation model corresponding to the characteristics of the training samples, and the augmented samples are then used as the training samples. A model to be trained is constructed based on the training samples, and the decomposition function corresponding to the model edge of the model to be trained is determined based on the characteristics of the training samples. The model to be trained is iteratively trained based on the characteristics of the training samples and the decomposition function to adjust the fusion weights in the model to be trained, thereby obtaining the target prediction model.

9. A device for predicting equipment operating performance, characterized in that, include: The acquisition unit is used to acquire the collected data related to the predicted operating effect of the device to be evaluated, and to classify the collected data to obtain multiple target datasets; The first input unit is used to input multiple target datasets into reference prediction models corresponding to the data categories of the target datasets respectively, to obtain multiple reference prediction indices, wherein the network structures of the multiple reference prediction models are the same; The second input unit is used to input multiple reference prediction indices into the target prediction model to obtain the target prediction index, wherein the target prediction model is a model with the same network architecture as the reference prediction model but a different network structure. The determining unit is used to determine the target prediction result corresponding to the device to be evaluated based on the collected data and the target prediction index.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the device operation effect prediction method according to any one of claims 1-8.

Images