Industrial equipment heterogeneous data acquisition and intelligent diagnosis method based on ai fusion algorithm

Abstract

The application is suitable for the technical field of industrial equipment state monitoring and intelligent fault diagnosis, and provides an industrial equipment heterogeneous data acquisition and intelligent diagnosis method based on an AI fusion algorithm.The method comprises the following steps: obtaining initial heterogeneous data of a plurality of acquisition nodes, obtaining initial equipment state features through feature extraction and AI fusion, and obtaining an initial diagnosis probability through an intelligent diagnosis model, and determining a candidate state set to be distinguished when there is diagnosis ambiguity; generating a candidate acquisition action according to a data source type and an adjustable acquisition parameter, calculating an expected state distinguishing benefit based on a reference feature distribution, and determining a target acquisition action in combination with data source reliability and acquisition resource consumption to obtain secondary heterogeneous data; and incrementally fusing secondary features and initial equipment state features and diagnosing again. The method can actively acquire data with high state distinguishing value for diagnosis ambiguity, reduces invalid acquisition and resource consumption, and improves the accuracy and reliability of intelligent diagnosis of industrial equipment.

Description

Technical Field

[0001] This invention belongs to the field of industrial equipment condition monitoring and intelligent fault diagnosis technology, and particularly relates to a method for heterogeneous data acquisition and intelligent diagnosis of industrial equipment based on AI fusion algorithms. Background Technology

[0002] As industrial equipment becomes increasingly intelligent, devices such as motors, pumps, fans, compressors, and machine tools are typically equipped with multiple data acquisition nodes to obtain data from various sources, including vibration, current, temperature, pressure, acoustics, and rotational speed. Current technologies generally perform time alignment, preprocessing, and feature extraction on this multi-source heterogeneous data. Then, AI fusion models are used to form equipment status characteristics, and intelligent diagnostic models are employed to identify the operating status of industrial equipment, thereby improving the accuracy of equipment condition monitoring and fault diagnosis.

[0003] However, different device states may exhibit similar characteristics under the current data collection conditions, causing intelligent diagnostic models to obtain similar diagnostic probabilities for multiple preset device states, thus leading to diagnostic ambiguity. To address this, existing technologies typically involve continuously collecting data according to a fixed data source type and collection parameters, or uniformly increasing the sampling frequency and extending the sampling duration when a device malfunctions. Since different data sources and collection parameters have varying abilities to distinguish between different device states, these methods easily collect a large amount of data that contributes little to the current diagnostic ambiguity. This not only increases the consumption of data collection, transmission, and computing resources but also makes it difficult to guarantee the acquisition of effective information for distinguishing easily confused device states. Furthermore, different collection nodes may be affected by noise, outliers, and communication packet loss, further reducing the diagnostic value of newly added data.

[0004] Therefore, existing heterogeneous data acquisition and intelligent diagnostic technologies for industrial equipment lack effective linkage between the data acquisition process and diagnostic results. In particular, there is a lack of technical means to evaluate the state discrimination capabilities of different candidate acquisition actions when there is diagnostic ambiguity in the initial diagnosis, and to determine the subsequent acquisition method based on the data source reliability and acquisition resource consumption. How to acquire more state-discriminating heterogeneous data in a targeted manner based on the current diagnostic ambiguity, while reducing invalid acquisition and resource consumption, and improving the accuracy and reliability of intelligent diagnostics for industrial equipment, has become an urgent technical problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide a method for heterogeneous data acquisition and intelligent diagnosis of industrial equipment based on AI fusion algorithms, aiming to solve the problems mentioned in the background art.

[0006] This invention is implemented as follows: a method for heterogeneous data acquisition and intelligent diagnosis of industrial equipment based on AI fusion algorithms, characterized in that the method includes:

[0007] Acquire initial heterogeneous data collected from multiple acquisition nodes, extract initial features, and fuse them using an AI fusion model to obtain initial device status features;

[0008] The initial device state features are input into the intelligent diagnostic model to obtain the initial diagnostic probability of each preset device state, and when the initial diagnostic probability characterizes a diagnostic ambiguity, a set of candidate states to be distinguished is determined.

[0009] Based on the data source type and adjustable acquisition parameters of each acquisition node, candidate acquisition actions are generated for each acquisition node. Based on the reference feature distribution of each preset device state under each candidate acquisition action, the expected state differentiation benefit of each candidate acquisition action for the set of candidate states to be differentiated is calculated.

[0010] The target collection action is determined based on the expected state of each candidate collection action, the corresponding data source credibility, and the collection resource consumption. The collection node corresponding to the target collection action is then controlled to execute the target collection action to obtain secondary heterogeneous data.

[0011] The secondary features of the secondary heterogeneous data are extracted, and the secondary features are incrementally fused with the initial device state features through the AI ​​fusion model to obtain updated device state features. The updated device state features are then input into the intelligent diagnostic model to obtain updated diagnostic results.

[0012] As a further limitation of the technical solution of this embodiment of the invention, the initial heterogeneous data includes data from at least two different data source types; the extraction of initial features includes: mapping data from different data source types to the same diagnostic time window according to a unified time benchmark, preprocessing and extracting features from the data in each diagnostic time window respectively, and obtaining the initial features corresponding to each data source type.

[0013] As a further limitation of the technical solution of this embodiment of the invention, when there is diagnostic ambiguity in the initial diagnostic probability representation, the set of candidate states to be distinguished is determined, including: calculating the diagnostic entropy and the probability interval between the maximum initial diagnostic probability and the second largest initial diagnostic probability based on each of the initial diagnostic probabilities; when the diagnostic entropy is greater than a preset entropy threshold and / or the probability interval is less than a preset interval threshold, determining that there is diagnostic ambiguity, and selecting at least two preset device states from the preset device states to form the set of candidate states to be distinguished based on each of the initial diagnostic probabilities.

[0014] As a further limitation of the technical solution of the present invention, the adjustable acquisition parameters include at least two of the following: sampling frequency, sampling duration, acquisition frequency band, and triggering conditions; each candidate acquisition action is composed of an acquisition node and a set of parameter combinations of the adjustable acquisition parameters corresponding to the acquisition node.

[0015] As a further limitation of the technical solution of this embodiment of the invention, the reference feature distribution is obtained according to the historical samples corresponding to each preset device state, and according to the data source type and adjustable acquisition parameters corresponding to each candidate acquisition action; for any two candidate device states in the set of candidate states to be distinguished, the predicted separability is determined according to the feature center distance and intra-class dispersion of the corresponding reference feature distribution, and the weight of the predicted separability is determined by the initial diagnosis probability corresponding to the two candidate device states, and the predicted separability between each candidate device state in the set of candidate states to be distinguished is weighted to obtain the expected state distinction benefit.

[0016] As a further limitation of the technical solution of the present invention, the reliability of the data source is determined based on at least two of the current data integrity rate, noise level, outlier ratio and communication packet loss rate of the corresponding collection node. The reliability of the data source increases with the increase of the current data integrity rate and decreases with the increase of any one of the noise level, outlier ratio and communication packet loss rate.

[0017] As a further limitation of the technical solution of this embodiment of the invention, the resource consumption for acquisition is determined based on at least two of the following: the amount of sampled data generated by the corresponding candidate acquisition action, the sampling duration, the amount of transmitted data, and the amount of calculated data; the acquisition action evaluation value of the candidate acquisition action is determined based on the product of the expected state distinction benefit of each candidate acquisition action and the corresponding data source credibility, as well as the resource consumption for acquisition, and the candidate acquisition action with the largest acquisition action evaluation value is determined as the target acquisition action.

[0018] As a further limitation of the technical solution of this embodiment of the invention, the incremental fusion includes: determining the incremental fusion weight of the secondary feature based on the expected state differentiation benefit of the target acquisition action and the corresponding data source credibility, wherein the incremental fusion weight increases as the expected state differentiation benefit and the data source credibility increase; the AI ​​fusion model fuses the secondary feature with the initial device state feature according to the incremental fusion weight to obtain the updated device state feature.

[0019] As a further limitation of the technical solution of this embodiment of the invention, the updated diagnostic result includes the updated diagnostic probability corresponding to each preset device state; the updated diagnostic probability is used to determine whether the updated diagnostic result meets the preset diagnostic confidence condition; if the preset diagnostic confidence condition is not met and the cumulative collection resource consumption has not reached the preset collection resource limit, the updated diagnostic probability replaces the initial diagnostic probability, the set of candidate states to be distinguished and the target collection action are re-determined, and secondary heterogeneous data acquisition, incremental fusion and re-diagnosis are performed in sequence until the preset diagnostic confidence condition is met or the cumulative collection resource consumption reaches the preset collection resource limit.

[0020] Compared with existing technologies, this invention acquires initial heterogeneous data collected by multiple acquisition nodes and forms initial device state features through an AI fusion model. When there is diagnostic ambiguity in the initial diagnostic probability representation obtained by the intelligent diagnostic model, a set of candidate states to be distinguished is determined. Furthermore, based on the reference feature distribution of each preset device state under different candidate acquisition actions, the expected state distinction benefit of each candidate acquisition action is calculated. The target acquisition action is determined by combining the corresponding data source credibility and acquisition resource consumption, so as to control the corresponding acquisition node to acquire secondary heterogeneous data for the current diagnostic ambiguity.

[0021] Compared to existing methods that use fixed data acquisition methods or uniformly increase the amount of data acquisition during anomalies, this invention can selectively change the subsequent data acquisition method according to the current candidate state to be distinguished, reducing the acquisition of data with low diagnostic contribution and the resulting data transmission and computing resource consumption. At the same time, by fusing the secondary features with the initial equipment state feature increments and then performing a second diagnosis, it can supplement the effective features needed to distinguish easily confused equipment states, improving the accuracy, reliability, and resource utilization efficiency of intelligent diagnosis of industrial equipment. It is applicable to multi-source condition monitoring and intelligent operation and maintenance scenarios such as motors, pumps, fans, compressors, and machine tools, and has good prospects for industrial applications. Attached Figure Description

[0022] Figure 1 A flowchart of the method provided in the embodiments of the present invention;

[0023] Figure 2 This is a flowchart illustrating the process of determining the expected state differentiation benefit for each candidate acquisition action in the method provided in this embodiment of the invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0025] Figure 1A flowchart of the method provided by an embodiment of the present invention is shown.

[0026] Specifically, the method for heterogeneous data acquisition and intelligent diagnosis of industrial equipment based on AI fusion algorithms includes the following steps:

[0027] Step S100: Obtain initial heterogeneous data collected by multiple acquisition nodes, extract initial features, and fuse them through an AI fusion model to obtain initial device status features.

[0028] This invention primarily applies to industrial equipment operation status monitoring and intelligent fault diagnosis, and is suitable for industrial equipment equipped with multiple acquisition nodes capable of acquiring operational data from various data sources. The industrial equipment can be motors, pumps, fans, compressors, gearboxes, machine tool spindles, or other equipment requiring multi-source operation monitoring. Each acquisition node is connected to a corresponding data source located within the industrial equipment body, drive system, control system, or operating environment, and is used to acquire data characterizing the operational status of the industrial equipment according to a preset initial acquisition method. Different acquisition nodes can correspond to different data source types, such as vibration data, current data, temperature data, pressure data, acoustic data, and rotational speed data, thereby forming initial heterogeneous data for performing initial intelligent diagnosis.

[0029] It should be noted that the initial heterogeneous data referred to in this invention refers to data acquired by multiple acquisition nodes, including at least two different data source types, before determining whether subsequent secondary heterogeneous data acquisition is necessary. Data from different data source types differ in their physical meaning, data structure, sampling frequency, sampling time, and numerical range, and cannot be directly used as a unified representation of equipment status. For example, vibration data typically has a high sampling frequency and is used to reflect the periodic vibration, impact, and frequency components of industrial equipment; temperature data changes relatively slowly and is used to reflect the thermal state of equipment; current data can reflect load changes and electromechanical coupling states. Therefore, before fusing data from different data source types, it is necessary to first establish a corresponding temporal relationship between the data from each data source type and extract initial features adapted to their respective data characteristics.

[0030] Existing intelligent diagnostic methods for industrial equipment typically focus on improving the recognition accuracy of a single diagnostic model. This involves continuously collecting data under predetermined data sources, sampling frequencies, and sampling durations, and then inputting the collected data into the diagnostic model to categorize the fault. When multiple equipment states exhibit similar data performance under the existing data collection conditions, current methods usually improve the diagnostic results by increasing model complexity, increasing the number of training samples, or increasing the overall sampling frequency. However, they rarely consider what discriminative information the current collected data lacks to differentiate between multiple candidate equipment states, or whether it is necessary to reverse-engineer subsequent data collection methods based on the diagnostic ambiguities revealed in the initial diagnosis. This can lead to two problems: firstly, continuously collecting data with low diagnostic discriminative value increases the consumption of collection, transmission, and computational resources; secondly, even with increased data volume, it may not be possible to obtain the key features needed to differentiate between easily confused equipment states.

[0031] Based on this, the present invention does not merely complete a one-time diagnosis using initial heterogeneous data, but rather uses the initial equipment state characteristics formed by the initial heterogeneous data as the basis for subsequent diagnostic ambiguity analysis and target acquisition action determination. Specifically, step S100 first utilizes multiple acquisition nodes to obtain initial heterogeneous data of the target industrial equipment under the current acquisition conditions, and through a unified time base, diagnostic time window mapping, data preprocessing, feature extraction, and AI fusion, forms initial equipment state characteristics that can comprehensively characterize the current operating state of the target industrial equipment. Subsequent steps obtain the initial diagnostic probability of each preset equipment state based on the initial equipment state characteristics, and further determine the set of candidate states to be distinguished when diagnostic ambiguity exists, and then evaluate the expected state distinction benefits of different candidate acquisition actions for the set of candidate states to be distinguished. Therefore, the initial equipment state characteristics formed in step S100 are not only used to obtain the initial diagnostic results, but also to establish the data foundation for subsequent "diagnostic results driving data acquisition".

[0032] Specifically, multiple acquisition nodes collect data from their respective data source types and configure corresponding time information for the collected data. This time information can be a sampling timestamp generated by the acquisition nodes, or a time stamp calibrated by an industrial equipment controller, edge acquisition terminal, or unified clock source. Since different acquisition nodes may have different data sampling frequencies and sampling start times, the data from different data source types are first correlated according to a unified time reference. This unified time reference is used to convert data from different acquisition nodes to a comparable time reference, enabling data from different data source types to be matched according to the actual operating period of their corresponding industrial equipment.

[0033] After completing the unified time base processing, data from different data source types are mapped to the same diagnostic time window. The diagnostic time window is a continuous time interval determined for primary equipment status analysis. For example, the industrial equipment operation process can be divided into sliding intervals according to a preset time length, and data whose timestamps fall within the same time interval under the unified time base can be included in the same diagnostic time window. For different data source types with different sampling frequencies, it is not required that each data source have the same number of data points within the same diagnostic time window; it is only necessary to ensure that the data from each data source type corresponds to the same actual operating time interval of the industrial equipment. If necessary, one or more methods, such as resampling, interpolation, data truncation, or timestamp correction, can be used based on the data characteristics of each data source type to ensure that data from different data source types correspond to the same diagnostic time window in a temporal sense.

[0034] For example, in one embodiment, a vibration acquisition node acquires vibration data at a first sampling frequency, a current acquisition node acquires current data at a second sampling frequency, and a temperature acquisition node acquires temperature data at a third sampling frequency, wherein the first, second, and third sampling frequencies may be different from each other. (Based on time intervals) As a diagnostic time window, vibration data, current data, and temperature data within the specified time interval are acquired, and these three types of data are collectively determined as the initial heterogeneous data corresponding to this diagnostic time window. This avoids the situation where data from different data sources actually correspond to different equipment operating times due to mechanical correspondence based solely on data sequence numbers.

[0035] For data from different data source types mapped to each diagnostic time window, preprocessing is performed separately. This preprocessing aims to reduce the impact of acquisition noise, outlier values, different units of measurement, and missing data on subsequent feature extraction. Depending on the data source type, the preprocessing method may vary, including but not limited to one or more of the following: filtering and denoising, outlier removal or correction, missing value handling, detrending, normalization, and standardization. For example, vibration data may undergo filtering and detrending, temperature data may undergo outlier handling and normalization, and current data may undergo denoising and amplitude standardization. Different data source types should employ preprocessing methods adapted to their data characteristics; it is not required that all data use the same preprocessing algorithm.

[0036] Feature extraction is performed on the preprocessed data from each data source type to obtain initial features corresponding to each data source type. These initial features characterize the equipment operating status reflected by the corresponding data source type within the current diagnostic time window. For time-series data, the initial features may include one or more of the following: time-domain features, frequency-domain features, time-frequency features, or deep features obtained by a feature extraction network. For example, initial features for vibration data may include root mean square value, peak value, kurtosis, spectral energy, or specific frequency band energy; initial features for current data may include current amplitude variation, harmonic characteristics, or frequency band energy; and initial features for temperature data may include temperature value, temperature change rate, and temperature statistical characteristics within the window. The specific features described above are only used to illustrate that different data source types can extract corresponding initial features according to their respective data characteristics and do not constitute a limitation on the types of initial features.

[0037] Furthermore, feature extraction methods corresponding to different data source types can be set separately, or a trained feature extraction network can be used to convert data from different data source types into feature vectors of a preset dimension. Through the above processing, at least two initial features corresponding to different data source types are formed within each diagnostic time window. Although each initial feature reflects the operating status of different physical dimensions of industrial equipment, after processing with a unified time base, they all correspond to the same diagnostic time window, and therefore can be used as input for cross-data source fusion in the AI ​​fusion model.

[0038] Initial features corresponding to each data source type are input into the AI ​​fusion model, which then fuses these initial features to obtain initial device state features. The AI ​​fusion model establishes the correlation between initial features from different data source types and converts initial features representing different physical states into a unified device state feature representation. The AI ​​fusion model can employ a neural network model capable of multi-source feature fusion, such as an attention-based fusion network, a gating-based fusion network, a Transformer-based fusion network, or other models capable of forming a comprehensive feature representation based on initial features corresponding to multiple data source types.

[0039] In one implementation, the initial features corresponding to each data source type are denoted as follows: , ... ,in The number of data source types participating in the current diagnostic time window fusion, and ;Will , ... The AI ​​fusion model is input, and different initial features are mapped and fused through the AI ​​fusion model to output the initial device state features. The initial device state characteristics It also includes device status information from multiple data source types, used to characterize the overall operating status of the target industrial equipment within the current diagnostic time window.

[0040] Through the above processing, initial heterogeneous data with different sampling frequencies, data structures, and physical meanings are first unified into the same diagnostic time window, and then initial features corresponding to each data source type are formed. These features are then used by an AI fusion model to form initial device state features, thereby avoiding mismatch of state information caused by directly fusing heterogeneous data that has not undergone time-correspondence processing. Simultaneously, these initial device state features serve as input to the subsequent intelligent diagnostic model, providing a unified data foundation for obtaining the initial diagnostic probability of each preset device state, identifying diagnostic ambiguities, and further determining the target acquisition action for the set of candidate states to be distinguished.

[0041] Furthermore, the method also includes the following steps:

[0042] Step S200: Input the initial device state features into the intelligent diagnostic model to obtain the initial diagnostic probability of each preset device state, and determine the set of candidate states to be distinguished when the initial diagnostic probability characterization has diagnostic ambiguity.

[0043] In this embodiment of the invention, the intelligent diagnostic model can be pre-trained based on historical samples with equipment status labels, and is used to identify the current equipment status of the industrial equipment according to the initial equipment status features. The preset equipment status may include a normal status and different pre-defined fault states. After inputting the initial equipment status features into the intelligent diagnostic model, the initial diagnostic probability corresponding to each preset equipment status is obtained. Each initial diagnostic probability is used to characterize the probability that the industrial equipment is currently in the corresponding preset equipment status.

[0044] Further, the diagnostic entropy and the probability interval between the maximum and second-largest initial diagnostic probabilities are calculated based on each initial diagnostic probability. The diagnostic entropy characterizes the overall uncertainty of the initial diagnostic probability corresponding to each preset device state. Let the initial diagnostic probability corresponding to the i-th preset device state be... Then the diagnostic entropy It can be represented as:

[0045]

[0046] The higher the diagnostic entropy, the more dispersed the initial diagnostic probability, and the higher the uncertainty of the current diagnostic result.

[0047] At the same time, the initial diagnostic probabilities are sorted according to their numerical values, and the highest initial diagnostic probability is denoted as... Let the second largest initial diagnosis probability be denoted as Then the probability interval The smaller the probability interval, the more difficult it is for the intelligent diagnostic model to clearly distinguish between the two preset device states with the highest probabilities.

[0048] When the diagnostic entropy is greater than a preset entropy threshold and / or the probability interval is less than a preset interval threshold, it is determined that the initial diagnostic probability representation has diagnostic ambiguity. Subsequently, at least two preset device states are selected from the preset device states to form a set of candidate states to be distinguished based on each of the initial diagnostic probabilities. Specifically, a preset number of preset device states can be selected in descending order of initial diagnostic probabilities, or at least two preset device states with initial diagnostic probabilities greater than a preset probability threshold can be selected.

[0049] For example, the initial diagnostic probabilities corresponding to the three preset device states are 0.46, 0.42, and 0.12, respectively, with a probability interval of 0.04 between the maximum and the second-highest initial diagnostic probabilities. When this probability interval is less than a preset interval threshold, a diagnostic ambiguity is determined, and the two preset device states with initial diagnostic probabilities of 0.46 and 0.42 are identified as device states in the candidate state set to be distinguished, providing a basis for subsequently determining candidate acquisition actions for the candidate state set to be distinguished.

[0050] Furthermore, the method also includes the following steps:

[0051] Step S300: Generate candidate acquisition actions for each acquisition node according to the data source type and adjustable acquisition parameters of each acquisition node. Based on the reference feature distribution of each preset device state under each candidate acquisition action, calculate the expected state differentiation benefit of each candidate acquisition action for the set of candidate states to be differentiated.

[0052] In this embodiment of the invention, for each acquisition node, adjustable acquisition parameters are determined according to the corresponding data source type. A parameter combination is generated based on the optional parameter values ​​of the adjustable acquisition parameters, and an acquisition node and its corresponding set of adjustable acquisition parameter combinations are identified as a candidate acquisition action. The adjustable acquisition parameters include at least two of the following: sampling frequency, sampling duration, acquisition bandwidth, and triggering condition. The adjustable acquisition parameters and parameter ranges may differ for different data source types.

[0053] For example, for a data acquisition node corresponding to vibration data, different sampling frequencies, sampling durations, and acquisition bandwidths can be combined to form multiple candidate acquisition actions; similarly, for a data acquisition node corresponding to temperature data, different sampling frequencies and sampling durations can be combined to form corresponding candidate acquisition actions. Thus, each candidate acquisition action represents a method of data acquisition using a specific set of adjustable acquisition parameters at the corresponding acquisition node.

[0054] like Figure 2 The flowchart shown illustrates the process of determining the expected state differentiation benefit for each candidate acquisition action. The reference feature distribution is obtained based on historical samples corresponding to each preset device state, according to the data source type and adjustable acquisition parameters corresponding to each candidate acquisition action. For any two candidate device states in the set of candidate states to be differentiated, the predicted separability is determined based on the feature center distance and intra-class dispersion of the corresponding reference feature distribution. The weight of the predicted separability is determined using the initial diagnostic probabilities corresponding to the two candidate device states. The predicted separability between each candidate device state in the set of candidate states to be differentiated is then weighted to obtain the expected state differentiation benefit.

[0055] Furthermore, the method also includes the following steps:

[0056] Step S400: Based on the expected state of each candidate collection action, the revenue, the corresponding data source credibility, and the collection resource consumption, determine the target collection action, and control the collection node corresponding to the target collection action to execute the target collection action to obtain secondary heterogeneous data.

[0057] In this embodiment of the invention, after obtaining the expected state differentiation benefit corresponding to each candidate collection action, the candidate collection action is also evaluated in combination with the corresponding data source credibility and collection resource consumption, so as to avoid selecting the collection method with poor current data quality or high resource consumption based solely on the state differentiation capability, so that the final determined target collection action takes into account the state differentiation capability, data reliability and resource utilization efficiency.

[0058] Specifically, the data source reliability is used to characterize the reliability of the data currently collected by the corresponding acquisition node, and it is determined based on at least two of the following: current data integrity rate, noise level, outlier ratio, and communication packet loss rate. The current data integrity rate characterizes the completeness of the actual acquired data relative to the expected acquired data; the noise level characterizes the degree of noise interference affecting the acquired data; the outlier ratio characterizes the proportion of outlier data in the acquired data; and the communication packet loss rate characterizes the degree of data loss during data transmission. The data source reliability increases with the increase of the current data integrity rate and decreases with the increase of any one of the noise level, outlier ratio, and communication packet loss rate. For example, the above parameters can be normalized and weighted to obtain a data source reliability with a value ranging from 0 to 1.

[0059] The resource consumption for data acquisition is used to characterize the acquisition, transmission, and computation resources occupied in executing the corresponding candidate acquisition action. It is determined based on at least two of the following: sampled data volume, sampling duration, transmitted data volume, and computed data volume. Different parameters can be normalized and weighted to obtain the resource consumption for the corresponding candidate acquisition action; wherein, the larger the sampled data volume, sampling duration, transmitted data volume, or computed data volume, the greater the corresponding resource consumption.

[0060] Furthermore, the evaluation value of each candidate acquisition action is determined by multiplying the expected state of the acquisition action by the corresponding data source credibility and the corresponding acquisition resource consumption. In one implementation, the evaluation value of the i-th candidate acquisition action is... It can be represented as:

[0061]

[0062] in, Differentiate the revenue based on the corresponding expected state. Assess the credibility of the corresponding data source. For the corresponding collection resource consumption, This is a preset constant greater than 0. Therefore, the greater the expected benefit from state differentiation, the higher the reliability of the data source, and the lower the resource consumption for data collection, the higher the corresponding evaluation value of the data collection action. It should be noted that the above expression is only an example, and other calculation methods that can reflect the above relationship can also be used.

[0063] After obtaining the evaluation value of each candidate acquisition action, the candidate acquisition action with the highest evaluation value is determined as the target acquisition action, and the acquisition node corresponding to the target acquisition action is determined. Subsequently, the adjustable acquisition parameters included in the target acquisition action are sent to the corresponding acquisition node, controlling the acquisition node to perform data acquisition according to the target acquisition action to obtain secondary heterogeneous data.

[0064] For example, if a high-frequency vibration candidate acquisition action has a high expected state differentiation benefit, but the corresponding acquisition node currently has a high noise level and consumes a lot of acquisition resources, while another current candidate acquisition action has a slightly lower expected state differentiation benefit, but the data source has higher reliability and consumes less acquisition resources, then the latter may have a higher acquisition action evaluation value and be identified as the target acquisition action. Through this method, secondary heterogeneous data acquisition can be performed on the current candidate state to be differentiated, reducing invalid acquisitions with low reliability or high resource consumption, and providing a data foundation for subsequent secondary feature extraction, incremental fusion, and updating diagnostic results.

[0065] Furthermore, the method also includes the following steps:

[0066] Step S500: Extract the secondary features of the secondary heterogeneous data, incrementally fuse the secondary features with the initial device state features through the AI ​​fusion model to obtain updated device state features, and input the updated device state features into the intelligent diagnostic model to obtain updated diagnostic results.

[0067] In this embodiment of the invention, the secondary heterogeneous data is processed according to a feature extraction method adapted to the initial features to obtain corresponding secondary features. The secondary heterogeneous data is collected by the acquisition node corresponding to the target acquisition action according to corresponding adjustable acquisition parameters. Therefore, the secondary features are used to supplement the distinguishing state information between the current candidate states to be distinguished.

[0068] Further, the incremental fusion weights of the secondary features are determined based on the expected state differentiation benefit of the target acquisition action and the corresponding data source credibility. Specifically, the expected state differentiation benefit and the data source credibility can be normalized respectively, and the incremental fusion weights are determined based on their product. For example, the normalized expected state differentiation benefit can be denoted as... The credibility of the data source is denoted as Then the incremental fusion weight It can be represented as ,in, This is a function that maps the calculation results to a preset weight range. Therefore, the greater the expected state differentiation gain and the higher the corresponding data source credibility, the greater the incremental fusion weight of the secondary features.

[0069] The AI ​​fusion model fuses the secondary features with the initial device state features according to the incremental fusion weights to obtain updated device state features. This method ensures that secondary features actively collected for current diagnostic ambiguity and possessing high state-discrimination value contribute appropriately to the updated device state features, thereby reducing the adverse impact of low-reliability secondary features on re-diagnosis.

[0070] The updated device status features are input into the intelligent diagnostic model to obtain the update diagnostic probability corresponding to each preset device status, and based on this, it is determined whether the update diagnostic result meets the preset diagnostic confidence condition. The preset diagnostic confidence condition may be that the maximum update diagnostic probability reaches a preset probability threshold, and / or the probability interval between the maximum update diagnostic probability and the second largest update diagnostic probability reaches a preset interval threshold. When the preset diagnostic confidence condition is met, the current device status of the industrial equipment is determined based on the update diagnostic result.

[0071] When the updated diagnostic result does not meet the preset diagnostic confidence condition, and the cumulative collection resource consumption has not reached the preset collection resource limit, the initial diagnostic probability is replaced with the updated diagnostic probability. Based on the replaced diagnostic probability, the set of candidate states to be distinguished and the target collection action are re-determined, and secondary heterogeneous data is acquired again for incremental fusion and re-diagnosis. The cumulative collection resource consumption is the cumulative value of the collection resource consumption corresponding to each round of target collection action executed, until the updated diagnostic result meets the preset diagnostic confidence condition or the cumulative collection resource consumption reaches the preset collection resource limit.

[0072] For example, if the diagnostic probabilities of two candidate device states are relatively close in the initial diagnosis, after acquiring secondary heterogeneous data through the target acquisition action and completing incremental fusion, the updated diagnostic probability corresponding to one of the preset device states is significantly improved and meets the preset diagnostic confidence condition. Then, subsequent acquisition is stopped and the corresponding updated diagnostic result is output. If the preset diagnostic confidence condition is still not met, the next round of targeted acquisition is continued.

[0073] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0074] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0075] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0076] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

[0077] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for heterogeneous data acquisition and intelligent diagnosis of industrial equipment based on AI fusion algorithms, characterized in that, The method includes: Acquire initial heterogeneous data collected from multiple acquisition nodes, extract initial features, and fuse them using an AI fusion model to obtain initial device status features; The initial device state features are input into the intelligent diagnostic model to obtain the initial diagnostic probability of each preset device state, and when the initial diagnostic probability characterizes a diagnostic ambiguity, a set of candidate states to be distinguished is determined. Based on the data source type and adjustable acquisition parameters of each acquisition node, candidate acquisition actions are generated for each acquisition node. Based on the reference feature distribution of each preset device state under each candidate acquisition action, the expected state differentiation benefit of each candidate acquisition action for the set of candidate states to be differentiated is calculated. The target collection action is determined based on the expected state of each candidate collection action, the corresponding data source credibility, and the collection resource consumption. The collection node corresponding to the target collection action is then controlled to execute the target collection action to obtain secondary heterogeneous data. The secondary features of the secondary heterogeneous data are extracted, and the secondary features are incrementally fused with the initial device state features through the AI ​​fusion model to obtain updated device state features. The updated device state features are then input into the intelligent diagnostic model to obtain updated diagnostic results.

2. The method for heterogeneous data acquisition and intelligent diagnosis of industrial equipment based on AI fusion algorithm according to claim 1, characterized in that, The initial heterogeneous data includes data from at least two different data source types; the extraction of initial features includes: mapping data from different data source types to the same diagnostic time window according to a unified time benchmark, preprocessing and extracting features from the data within each diagnostic time window to obtain the initial features corresponding to each data source type.

3. The method for heterogeneous data acquisition and intelligent diagnosis of industrial equipment based on AI fusion algorithm according to claim 1, characterized in that, Determining a set of candidate states to be distinguished when there is diagnostic ambiguity in the initial diagnostic probability representation includes: calculating diagnostic entropy and probability interval between the largest and second largest initial diagnostic probabilities based on each initial diagnostic probability; determining that there is diagnostic ambiguity when the diagnostic entropy is greater than a preset entropy threshold and / or the probability interval is less than a preset interval threshold, and selecting at least two preset device states from the preset device states to form the set of candidate states to be distinguished based on each initial diagnostic probability.

4. The method for heterogeneous data acquisition and intelligent diagnosis of industrial equipment based on AI fusion algorithm according to claim 1, characterized in that, The adjustable acquisition parameters include at least two of the following: sampling frequency, sampling duration, acquisition frequency band, and triggering conditions; each candidate acquisition action consists of an acquisition node and a set of parameters of the adjustable acquisition parameters corresponding to that acquisition node.

5. The method for heterogeneous data acquisition and intelligent diagnosis of industrial equipment based on AI fusion algorithm according to claim 1, characterized in that, The reference feature distribution is obtained based on historical samples corresponding to each preset device state, according to the data source type and adjustable acquisition parameters corresponding to each candidate acquisition action. For any two candidate device states in the set of candidate states to be distinguished, the predicted separability is determined based on the feature center distance and intra-class dispersion of the corresponding reference feature distribution, and the weight of the predicted separability is determined by the initial diagnosis probability corresponding to the two candidate device states. The predicted separability between each candidate device state in the set of candidate states to be distinguished is weighted to obtain the expected state distinction benefit.

6. The method for heterogeneous data acquisition and intelligent diagnosis of industrial equipment based on AI fusion algorithm according to claim 1, characterized in that, The reliability of the data source is determined based on at least two of the following: the current data integrity rate, noise level, outlier ratio, and communication packet loss rate of the corresponding collection node. The reliability of the data source increases with the increase of the current data integrity rate and decreases with the increase of any one of the noise level, outlier ratio, and communication packet loss rate.

7. The method for heterogeneous data acquisition and intelligent diagnosis of industrial equipment based on AI fusion algorithm according to claim 1, characterized in that, The resource consumption for data acquisition is determined based on at least two of the following: the amount of sampled data generated by the corresponding candidate acquisition action, the sampling duration, the amount of transmitted data, and the amount of calculated data. The evaluation value of each candidate acquisition action is determined by the product of the expected state of the acquisition action and the corresponding data source credibility, as well as the acquisition resource consumption. The candidate acquisition action with the highest evaluation value is then determined as the target acquisition action.

8. The method for heterogeneous data acquisition and intelligent diagnosis of industrial equipment based on AI fusion algorithm according to claim 1, characterized in that, The incremental fusion includes: determining the incremental fusion weight of the secondary feature based on the expected state differentiation benefit of the target acquisition action and the corresponding data source credibility, wherein the incremental fusion weight increases as the expected state differentiation benefit and the data source credibility increase; the AI ​​fusion model fuses the secondary feature with the initial device state feature according to the incremental fusion weight to obtain the updated device state feature.

9. The method for heterogeneous data acquisition and intelligent diagnosis of industrial equipment based on AI fusion algorithm according to claim 1, characterized in that, The updated diagnostic results include the updated diagnostic probabilities corresponding to each preset device state; the updated diagnostic probabilities are used to determine whether the updated diagnostic results meet preset diagnostic confidence conditions; if the preset diagnostic confidence conditions are not met and the cumulative collection resource consumption has not reached the preset collection resource limit, the updated diagnostic probabilities replace the initial diagnostic probabilities, the set of candidate states to be distinguished and the target collection action are re-determined, and secondary heterogeneous data acquisition, incremental fusion and re-diagnosis are performed sequentially until the preset diagnostic confidence conditions are met or the cumulative collection resource consumption reaches the preset collection resource limit.

Images