Fault diagnosis method and system for device operation and maintenance based on large model knowledge graph

Abstract

This invention provides a fault diagnosis method and system for equipment operation and maintenance based on a large-scale model knowledge graph, belonging to the field of data analysis technology. In this invention, firstly, target hardware operation data of the target device is acquired, and a historical fault feature map constructed based on at least one historical hardware operation data and corresponding fault tags is obtained. Secondly, semantic features are extracted and aggregated for each local hardware operation data in the target hardware operation data to form aggregated hardware operation features. Then, based on the correlation between the historical fault feature map and the aggregated hardware operation features, the aggregated hardware operation features are updated to form updated hardware operation features. Finally, feature restoration is performed based on the updated hardware operation features to form target fault diagnosis data. Based on the above method, the problem of relatively low reliability in fault diagnosis in existing technologies can be improved.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and more specifically, to a fault diagnosis method and system for equipment operation and maintenance based on a large model knowledge graph. Background Technology

[0002] Modern equipment tends to be more complex and diverse in function, thus generally having more components. Timely and effective location of component faults is fundamental to equipment maintenance. However, with the continuous improvement of equipment intelligence and digitalization, the field of equipment operation and maintenance places higher demands on the accuracy and timeliness of fault diagnosis. Traditional equipment fault diagnosis methods mostly rely on manual rules, expert systems, or shallow machine learning models based on feature engineering, such as support vector machines and decision trees. These methods suffer from poor adaptability and weak generalization ability when faced with complex, multi-source, and multimodal equipment operating data. Especially when operating conditions change significantly, the diagnostic accuracy drops markedly, seriously affecting the stable operation of equipment and maintenance decisions. In other words, existing technologies suffer from relatively low reliability in fault diagnosis. Summary of the Invention

[0003] In view of this, the purpose of this invention is to provide a fault diagnosis method and system for equipment operation and maintenance based on a large model knowledge graph, so as to improve the problem of relatively low reliability of fault diagnosis in the prior art.

[0004] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions: A fault diagnosis method for equipment operation and maintenance based on a large-scale knowledge graph model includes: Acquire target hardware operation data of the target device, and acquire a historical fault feature map constructed based on at least one historical hardware operation data and corresponding fault labels. The target hardware operation data includes local hardware operation data of different multiple dimensions, including at least two of vibration, temperature and sound. Each map member in the historical fault feature map is used to reflect the semantic features of the historical hardware operation data or the fault label. Semantic feature extraction and semantic feature aggregation are performed on each local hardware operation data in the target hardware operation data to form hardware operation aggregated features; Based on the correlation between the historical fault feature map and the hardware operation aggregate features, the hardware operation aggregate features are updated to form hardware operation update features. Based on the hardware operation update characteristics, feature restoration is performed to form target fault diagnosis data, wherein the target fault diagnosis data is used to reflect the faulty components in the target device.

[0005] In some preferred embodiments, in the above-described fault diagnosis method for equipment operation and maintenance based on large model knowledge graphs, the step of extracting and aggregating semantic features from each local hardware operation data in the target hardware operation data to form aggregated hardware operation features includes: Initial semantic features are extracted from each local hardware operation data in the target hardware operation data to obtain the corresponding initial hardware operation features. Deep semantic features are extracted from each of the initial hardware operation features to obtain the corresponding hardware operation deep features. Based on the semantic correlation between each of the hardware operation depth features, the semantic features of multiple hardware operation initial features are aggregated to form hardware operation aggregate features.

[0006] In some preferred embodiments, in the above-described fault diagnosis method for equipment operation and maintenance based on a large model knowledge graph, the step of extracting initial semantic features from each local hardware operation data in the target hardware operation data to obtain the corresponding initial hardware operation features includes: The hardware operation local data is subjected to one-dimensional convolution processing to form hardware temporal convolution features, and the temporal information extracted from the hardware operation local data is embedded to form hardware temporal embedding features. The temporal information includes at least one of mean, root mean square, peak value, kurtosis, skewness, peak factor, and impulse factor. The target device includes a wind turbine. Two-dimensional convolution processing is performed on the frequency domain data of the local data of hardware operation to form hardware frequency domain convolution features, and the frequency domain information extracted from the frequency domain data is embedded to form hardware frequency domain embedding features, wherein the frequency domain information includes at least one of the fundamental frequency, 2 times the fundamental frequency, and 3 times the fundamental frequency; Based on the hardware temporal embedding features, the hardware temporal convolutional features are constrained and adjusted to form hardware temporal adjusted features, and based on the hardware frequency domain embedding features, the hardware frequency domain convolutional features are constrained and adjusted to form hardware frequency domain adjusted features. The hardware time-domain adjustment features and the hardware frequency-domain adjustment features are fused to form a corresponding hardware initial operation feature.

[0007] In some preferred embodiments, in the above-described fault diagnosis method for equipment operation and maintenance based on large model knowledge graphs, the step of fusing the hardware time-domain adjustment features and the hardware frequency-domain adjustment features to form a corresponding initial hardware operation feature includes: The hardware temporal adjustment features are subjected to various convolutional processes of different sizes to form multiple corresponding hardware convolutional features. The hardware temporal adjustment features are also subjected to fully connected processing to form hardware temporal fully connected features, wherein the number of feature parameters included in the hardware temporal fully connected features is equal to the number of the multiple hardware convolutional features. Based on the feature parameters in the hardware time-domain fully connected features, a weighted mean is calculated for the multiple hardware convolutional features, and based on the result of the weighted mean calculation and the hardware frequency-domain adjustment features, a corresponding hardware initial operation feature is obtained.

[0008] In some preferred embodiments, in the above-described fault diagnosis method for equipment operation and maintenance based on a large model knowledge graph, the step of extracting deep semantic features from each of the initial hardware operation features to obtain the corresponding deep hardware operation features includes: The sliding window feature sequence of the initial hardware operation features is subjected to depth mining in the first direction to form the first hardware operation feature; The sliding window feature sequence of the initial hardware operation features is subjected to second-direction depth mining to form second hardware operation features, wherein the first-direction depth mining and the second-direction depth mining belong to two opposite-direction depth mining. The first hardware operation feature and the second hardware operation feature are fused together to form a hardware operation depth feature.

[0009] In some preferred embodiments, in the above-described fault diagnosis method for equipment operation and maintenance based on a large model knowledge graph, the step of performing first-direction depth mining on the sliding window feature sequence of the initial hardware operation features to form first hardware operation features includes: Using the a-th first hardware operation feature, the (a+1)-th sliding window feature in the sliding window feature sequence of the initial hardware operation feature is focused on the feature based on the attention mechanism to form the first hardware focus feature based on the attention mechanism, where a is an integer greater than or equal to 0, and the 0th first hardware operation feature is formed during the training process of the corresponding neural network model. Using the a-th first hardware operation feature, the (a+1)-th sliding window feature in the sliding window feature sequence of the initial hardware operation feature is subjected to feature focusing based on a gating mechanism to form the first hardware focusing feature based on the gating mechanism; Using the a-th first hardware operation feature, the (a+1)-th sliding window feature in the sliding window feature sequence of the initial hardware operation feature is subjected to feature focusing based on the convolution mechanism to form the first hardware focusing feature based on the convolution mechanism; The first hardware focusing feature based on the attention mechanism, the first hardware focusing feature based on the gating mechanism, and the first hardware focusing feature based on the convolution mechanism are fused to form the (a+1)th first hardware operation feature. The final output first hardware operating characteristic is determined based at least on the last formed first hardware operating characteristic.

[0010] In some preferred embodiments, in the above-described fault diagnosis method for equipment operation and maintenance based on large model knowledge graphs, the step of performing second-direction depth mining on the sliding window feature sequence of the initial hardware operation features to form second hardware operation features includes: Using the b-th second hardware operation feature, the (b-1)-th sliding window feature in the sliding window feature sequence of the initial hardware operation feature is focused on based on an attention mechanism to form a second hardware focus feature based on an attention mechanism, where b is an integer greater than or equal to 2 and less than or equal to B, and the B-th second hardware operation feature is formed during the training process of the corresponding neural network model. Using the b-th second hardware operation feature, the (b-1)-th sliding window feature in the sliding window feature sequence of the initial hardware operation feature is subjected to feature focusing based on a gating mechanism to form a second hardware focusing feature based on a gating mechanism; Using the b-th second hardware operation feature, the (b-1)-th sliding window feature in the sliding window feature sequence of the initial hardware operation feature is subjected to feature focusing based on the convolution mechanism to form the second hardware focusing feature based on the convolution mechanism; The second hardware focusing feature based on the attention mechanism, the second hardware focusing feature based on the gating mechanism, and the second hardware focusing feature based on the convolution mechanism are fused to form the (b-1)th second hardware operation feature. The final output second hardware operating characteristics are determined based at least on the last formed second hardware operating characteristics.

[0011] In some preferred embodiments, in the above-described fault diagnosis method for equipment operation and maintenance based on a large model knowledge graph, the step of updating the hardware operation aggregate features based on the correlation between the historical fault feature graph and the hardware operation aggregate features to form updated hardware operation features includes: For each graph member in the historical fault feature graph that represents the semantic features of historical hardware operation data, based on the feature similarity between the graph member and the hardware operation aggregate feature, it is determined whether the hardware operation aggregate feature is related to the graph member. The hardware operation aggregation features are used as new graph members and connected with related graph members to update the historical fault feature graph and form a new fault feature graph. The new map members are used as targets for feature transfer and updates in the new fault feature map to form hardware operation update features.

[0012] In some preferred embodiments, in the above-described fault diagnosis method for device operation and maintenance based on a large model knowledge graph, the step of using the new graph member as the target for feature transfer and update in the new fault feature graph to form hardware operation update features includes: Using the new graph member as the endpoint, traverse the new fault feature graph to form at least one graph member traversal path; For each of the graph member traversal paths, along the direction from the start point to the end point of the graph member traversal path, each graph member is sequentially associated and focused to form the hardware operation focus feature corresponding to the graph member traversal path. Based on the hardware operation focus features corresponding to each of the graph member traversal paths, the hardware operation update features are determined.

[0013] This invention also provides a fault diagnosis system for equipment operation and maintenance based on a large model knowledge graph, including a processor and a memory. The memory is used to store a computer program, and the processor is used to execute the computer program to implement the above-mentioned fault diagnosis method for equipment operation and maintenance based on a large model knowledge graph.

[0014] The fault diagnosis method and system for equipment operation and maintenance based on a large-scale model knowledge graph provided in this invention first acquires the target hardware operation data of the target device and obtains a historical fault feature map constructed based on at least one historical hardware operation data and corresponding fault tags. Second, semantic features are extracted and aggregated for each local hardware operation data in the target hardware operation data to form aggregated hardware operation features. Then, based on the correlation between the historical fault feature map and the aggregated hardware operation features, the aggregated hardware operation features are updated to form updated hardware operation features. Finally, feature restoration is performed based on the updated hardware operation features to form target fault diagnosis data. Based on the above method, on the one hand, by utilizing the semantic features of at least two dimensions of local hardware operation data, the basis for fault diagnosis can be more sufficient and effective, thereby improving the reliability of fault diagnosis. On the other hand, since the hardware operation aggregate features formed by semantic feature extraction and semantic feature aggregation are updated based on historical fault feature maps, prior knowledge can be fully utilized. Compared with schemes that classify faults directly based on the relationship with prior knowledge, this scheme does not overly rely on prior knowledge while making use of it, but also fully considers the current data. The reliability of the target fault diagnosis data can be further improved. Therefore, it can improve the problem of relatively low reliability of fault diagnosis in the existing technology.

[0015] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0016] Figure 1 The structural block diagram of the fault diagnosis system for equipment operation and maintenance based on a large model knowledge graph provided in the embodiments of the present invention.

[0017] Figure 2 The flowchart illustrates the steps of the fault diagnosis method for equipment operation and maintenance based on a large model knowledge graph provided in this embodiment of the invention.

[0018] Figure 3 This is a schematic diagram of a historical fault feature map provided in an embodiment of the present invention.

[0019] Figure 4 This is a schematic diagram of the initial semantic feature extraction provided in an embodiment of the present invention.

[0020] Figure 5 This is a schematic diagram of feature fusion provided in an embodiment of the present invention.

[0021] Figure 6 This is a schematic diagram of first-direction depth mining provided in an embodiment of the present invention.

[0022] Figure 7 This is a schematic diagram of second-direction depth mining provided in an embodiment of the present invention.

[0023] Figure 8 This is a schematic diagram illustrating the updating of historical fault feature maps provided in an embodiment of the present invention.

[0024] Figure 9 This is a schematic diagram of associated focusing provided in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0026] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0027] like Figure 1 As shown, this embodiment of the invention provides a fault diagnosis system for equipment operation and maintenance based on a large-scale knowledge graph. The fault diagnosis system for equipment operation and maintenance based on a large-scale knowledge graph may include a memory and a processor.

[0028] In detail, the memory and the processor are electrically connected directly or indirectly to enable data transmission or interaction. For example, they can be electrically connected via one or more communication buses or signal lines. The memory may store at least one software functional module (computer program) that exists in the form of software or firmware. The processor can be used to execute the executable computer program stored in the memory, thereby implementing the fault diagnosis method for equipment operation and maintenance based on a large model knowledge graph provided in this embodiment of the invention.

[0029] Alternatively, in an alternative example, the memory may be, but is not limited to, random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), etc.

[0030] Alternatively, in an alternative example, the processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), a system on chip (SoC), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0031] Alternatively, in an alternative example, the fault diagnosis system for equipment operation and maintenance based on a large model knowledge graph can be a server with data processing capabilities.

[0032] Combination Figure 2 This invention also provides a fault diagnosis method for equipment operation and maintenance based on a large-scale knowledge graph, which can be applied to the aforementioned fault diagnosis system for equipment operation and maintenance based on a large-scale knowledge graph. The method steps defined in the relevant process of the fault diagnosis method for equipment operation and maintenance based on a large-scale knowledge graph can be implemented by the fault diagnosis system for equipment operation and maintenance based on a large-scale knowledge graph. The following will describe... Figure 2 The specific process shown will be explained in detail.

[0033] Step S110: Obtain the target hardware operation data of the target device, and obtain the historical fault feature map constructed based on at least one historical hardware operation data and the corresponding fault label.

[0034] In this embodiment of the invention, the fault diagnosis system for equipment operation and maintenance based on a large model knowledge graph can acquire target hardware operation data of the target device and acquire a historical fault feature graph constructed based on at least one historical hardware operation data and corresponding fault tags. The target hardware operation data includes local hardware operation data of different dimensions, including at least two of vibration, temperature, and sound (it should be noted that in most modern equipment, when a component malfunctions, there is generally at least one of abnormal vibration, abnormal temperature, and abnormal sound; at least two can achieve mutual constraints between different dimensions, thereby improving the accuracy of potential semantic features). Each graph member in the historical fault feature graph reflects the semantic features of the historical hardware operation data or the fault tags. For example, a neural network model can be used to extract and aggregate semantic features from the historical hardware operation data to form corresponding hardware operation aggregate features, i.e., corresponding semantic features. Furthermore, the fault tags can be embedded (through a word embedding model) to obtain corresponding semantic features. Figure 3 In the historical fault feature map, the map members of historical hardware operation data and the map members of corresponding fault labels are connected to indicate a correlation. Whether there is a correlation between map members of two historical hardware operation data is determined by whether the feature similarity between the corresponding hardware operation aggregate features exceeds a threshold. Furthermore, it should be noted that the target hardware operation data can be formed by collecting information from vibration sensors, temperature sensors, and sound sensors deployed on the target device. Each dimension of hardware operation local data belongs to time-series data, i.e., it belongs to the time domain. For example, the hardware operation local data of the vibration dimension can be the vibration amplitude 1 at time point 1, the vibration amplitude 2 at time point 2, the vibration amplitude 3 at time point 3, the vibration amplitude 4 at time point 4, ..., the vibration amplitude n at time point n.

[0035] Step S120: Extract semantic features and aggregate semantic features for each local hardware operation data in the target hardware operation data to form aggregated hardware operation features.

[0036] In this embodiment of the invention, after obtaining the target hardware operating data, the fault diagnosis system for equipment operation and maintenance based on a large model knowledge graph can perform semantic feature extraction and semantic feature aggregation on each local hardware operating data in the target hardware operating data to form aggregated hardware operating features. That is, semantic features can be extracted from each local hardware operating data separately, and then each semantic feature can be aggregated to form aggregated hardware operating features. This ensures both a small granularity in the semantic feature extraction process, resulting in higher accuracy, and a high richness in the resulting aggregated hardware operating features. Furthermore, it should be noted that semantic feature extraction refers to mining the potential semantic information in the data and representing it in the form of vectors or matrices. That is, in this embodiment of the invention, the representation of each feature can be a vector or a matrix.

[0037] Step S130: Based on the correlation between the historical fault feature map and the hardware operation aggregate features, update the hardware operation aggregate features to form hardware operation update features.

[0038] In this embodiment of the invention, after obtaining the hardware operation aggregated features, the equipment operation and maintenance fault diagnosis system based on a large model knowledge graph can update the hardware operation aggregated features based on the correlation between the historical fault feature graph and the hardware operation aggregated features, forming updated hardware operation features. That is, the hardware operation aggregated features can be constrained based on the historical fault feature graph, thereby achieving effective utilization of prior knowledge (not simply splicing or superimposing, but updating based on correlation, which effectively reduces the introduction of invalid or interfering information) and ensuring the effective utilization of the hardware operation aggregated features. This results in higher semantic representation accuracy of the formed updated hardware operation features.

[0039] Step S140: Based on the hardware operation update features, feature restoration is performed to form target fault diagnosis data.

[0040] In this embodiment of the invention, after obtaining the hardware operation update features, the equipment operation and maintenance fault diagnosis system based on a large model knowledge graph can perform feature restoration based on the hardware operation update features to form target fault diagnosis data. The target fault diagnosis data reflects the faulty components in the target equipment. It should be noted that feature restoration can refer to classification output, that is, outputting the parameter distribution of whether each component of the target equipment has a fault, and then determining the faulty components based on this parameter distribution, thereby obtaining the target fault diagnosis data, such as a faulty fan or a faulty compressor pump.

[0041] Based on the above method, on the one hand, by utilizing the semantic features of at least two dimensions of local hardware operation data, the basis for fault diagnosis can be more sufficient and effective, thereby improving the reliability of fault diagnosis. On the other hand, by updating the hardware operation aggregated features formed by semantic feature extraction and aggregation based on historical fault feature maps, prior knowledge can be fully utilized. Compared with schemes that directly classify faults based on the relationship with prior knowledge, this scheme does not overly rely on prior knowledge while fully considering current data, further improving the reliability of the target fault diagnosis data. Therefore, it can improve the problem of relatively low reliability of fault diagnosis in existing technologies.

[0042] Regarding step S110 in the above content, it should be further explained that the detailed process of obtaining the target hardware operation data of the target device is not limited and can be configured according to actual needs.

[0043] For example, in some feasible implementations, to ensure the timeliness of fault diagnosis, the hardware operation data collected by each sensor can be acquired in real time to obtain real-time target hardware operation data, enabling timely subsequent fault diagnosis and improving the real-time performance of maintenance. In other feasible implementations, the hardware operation data collected by each sensor can be stored first. When the target device malfunctions, if it is necessary to locate the faulty component, the target hardware operation data can be retrieved from the corresponding storage device for fault diagnosis.

[0044] Regarding step S120 in the above content, it should be further explained that the detailed process of semantic feature extraction and semantic feature aggregation for each hardware operation local data in the target hardware operation data is not limited and can be configured accordingly according to actual needs.

[0045] For example, in some feasible implementations, in order to improve the efficiency of latent semantic feature extraction and reduce computational costs, each of the hardware operation local data can be convolved (time series data can be convolved in one dimension), and then the results of the convolution processing can be concatenated or superimposed to obtain the corresponding hardware operation aggregate features.

[0046] For example, in some other feasible implementations, in order to improve the reliability of latent semantic feature extraction and make the resulting hardware operation aggregate features have high semantic representation accuracy, the above step S120 may further include the following sub-steps, the specific contents of which are as follows.

[0047] Step S121: Extract initial semantic features from each local hardware operation data in the target hardware operation data to obtain the corresponding initial hardware operation features.

[0048] In this embodiment of the invention, initial semantic features can be extracted from each local hardware operation data in the target hardware operation data to obtain corresponding initial hardware operation features. For example, initial semantic features can be extracted from the local hardware operation data in the vibration dimension to obtain initial hardware operation features in the vibration dimension; similarly, initial semantic features can be extracted from the local hardware operation data in the sound dimension to obtain initial hardware operation features in the sound dimension; and similarly, initial semantic features can be extracted from the local hardware operation data in the temperature dimension to obtain initial hardware operation features in the temperature dimension. Furthermore, initial semantic feature extraction can refer to extracting shallow semantic features from the local hardware operation data.

[0049] Step S122: Perform deep semantic feature extraction on each of the initial hardware operation features to obtain the corresponding hardware operation deep features.

[0050] In this embodiment of the invention, after obtaining the initial hardware operation features, deep semantic feature extraction can be performed on each of the initial hardware operation features to obtain corresponding hardware operation deep features. For example, deep semantic feature extraction can be performed on the initial hardware operation features of the vibration dimension to obtain hardware operation deep features of the vibration dimension; similarly, deep semantic feature extraction can be performed on the initial hardware operation features of the sound dimension to obtain hardware operation deep features of the sound dimension; and deep semantic feature extraction can be performed on the initial hardware operation features of the temperature dimension to obtain hardware operation deep features of the temperature dimension. Furthermore, deep semantic feature extraction can refer to extracting deep semantic features (abstract semantic features) from the initial hardware operation features to capture more detailed potential semantic features.

[0051] Step S123: Based on the semantic correlation between each of the hardware operation depth features, the semantic features of multiple hardware operation initial features are aggregated to form hardware operation aggregate features.

[0052] In this embodiment of the invention, after obtaining the hardware operation depth features, multiple initial hardware operation features can be aggregated semantically based on the semantic correlation between each hardware operation depth feature to form aggregated hardware operation features. For example, each hardware operation depth feature can be concatenated, and then the resulting concatenated feature can be pooled and compressed to reduce its size (effectively reducing the computational cost of subsequent self-attention processing) while filtering out redundant information. Then, self-attention processing is performed on the resulting pooled feature to capture the relevant semantic information between semantic features of different dimensions, thereby achieving effective semantic feature aggregation between different semantic features.

[0053] Regarding step S121 above, it should be further explained that the detailed process of performing initial semantic feature extraction on each hardware operation local data in the target hardware operation data is not limited. For example, in some feasible implementations, in order to improve the reliability of initial semantic feature extraction, that is, to mine more reliable semantic features from the hardware operation local data, step S121 above may further include the following sub-steps.

[0054] Step S121a: Perform one-dimensional convolution processing on the local hardware operation data to form hardware temporal convolution features, and embed the temporal information extracted from the local hardware operation data to form hardware temporal embedding features.

[0055] In this embodiment of the invention, combined with Figure 4 (The waveform and spectrum diagrams in the figure are for illustrative purposes only.) One-dimensional convolution processing can be performed on the local data of the hardware operation (e.g., the size of the convolution kernel can be 1). 3. The convolution stride can be 1 (with edge padding), forming hardware temporal convolutional features. Temporal information extracted from the local hardware operation data is then embedded (using a word embedding model) to form hardware temporal embedding features. The temporal information includes at least one of the following: mean, root mean square (RMS), peak, kurtosis, skewness, crest factor, and impulse factor. These can be used to reflect signal amplitude changes, waveform characteristics, and impact degree. The target equipment includes a fan (when a fan malfunctions, it may produce abnormal vibrations, abnormal sounds, abnormal temperatures, etc.).

[0056] Step S121b involves performing two-dimensional convolution processing on the frequency domain data of the hardware operation local data to form hardware frequency domain convolution features, and then embedding the frequency domain information extracted from the frequency domain data to form hardware frequency domain embedding features.

[0057] In the embodiments of the invention, the frequency domain data of the hardware operation local data (which can be obtained by performing a Fourier transform on the hardware operation local data, and the specific transformation process can refer to relevant prior art) can be subjected to two-dimensional convolution processing (e.g., the size of the convolution kernel can be 3). 3. The convolution stride can be 1 (with edge padding), forming hardware frequency domain convolutional features. Frequency domain information extracted from the frequency domain data is then embedded (using a word embedding model) to form hardware frequency domain embedding features. The frequency domain information includes at least one of the following: fundamental frequency (e.g., related to rotor imbalance, concentricity, etc. in wind turbines), 2nd fundamental frequency, and 3rd fundamental frequency (i.e., low-order harmonics, related to shaft bending, misalignment, etc. in wind turbines).

[0058] Step S121c: Based on the hardware temporal embedding features, constrain and adjust the hardware temporal convolution features to form hardware temporal adjustment features, and based on the hardware frequency domain embedding features, constrain and adjust the hardware frequency domain convolution features to form hardware frequency domain adjustment features.

[0059] In this embodiment of the invention, after obtaining the hardware temporal embedding feature and the hardware temporal convolutional feature, the hardware temporal convolutional feature can be constrained and adjusted based on the hardware temporal embedding feature, such as by using a cross-attention mechanism to form a hardware temporal adjustment feature. For example, based on the cross-attention mechanism, semantic features related to the hardware temporal embedding feature can be mined from the hardware temporal convolutional feature, and then the semantic features can be concatenated with the hardware temporal convolutional feature (e.g., by addition, to mitigate the potential loss of semantic information) to obtain the hardware temporal adjustment feature. Similarly, after obtaining the hardware frequency domain embedding feature and the hardware frequency domain convolutional feature, the hardware frequency domain convolutional feature can be constrained and adjusted based on the hardware frequency domain embedding feature, such as by using a cross-attention mechanism to form a hardware frequency domain adjustment feature. For example, based on the cross-attention mechanism, semantic features related to the hardware frequency domain embedding feature can be mined from the hardware frequency domain convolutional feature, and then the semantic features can be concatenated with the hardware frequency domain convolutional feature (e.g., by addition, to mitigate the potential loss of semantic information) to obtain the hardware frequency domain adjustment feature.

[0060] Step S121d: Perform feature fusion on the hardware time-domain adjustment feature and the hardware frequency-domain adjustment feature to form a corresponding hardware operation initial feature.

[0061] In this embodiment of the invention, after obtaining the hardware time-domain adjustment feature and the hardware frequency-domain adjustment feature, the hardware time-domain adjustment feature and the hardware frequency-domain adjustment feature can be fused to form a corresponding hardware operation initial feature. In this way, the hardware operation initial feature can fully represent the semantic features of the two domains, namely, effectively improving the semantic representation capability of the hardware operation initial feature.

[0062] Regarding step S121d above, it should be further explained that the detailed process of feature fusion of the hardware time-domain adjustment feature and the hardware frequency-domain adjustment feature is not limited. For example, in some feasible implementations, in order to more effectively capture semantic features under more fields of view and avoid the problem of capturing invalid semantic information due to different contributions of different fields of view to the actual data, the above step S121d may further include the following sub-steps.

[0063] Step d1 involves performing various convolutional processes with different sizes on the hardware temporal adjustment features to form multiple corresponding hardware convolutional features, and performing fully connected processing on the hardware temporal adjustment features to form hardware temporal fully connected features.

[0064] In this embodiment of the invention, combined with Figure 5 The hardware temporal adjustment features can be subjected to various convolutional processes with different sizes (e.g., the kernel size can be 1...). 3.1 5. The convolution stride can all be equal to 1, and edge padding is performed on all of them. This forms multiple corresponding hardware convolutional features (one convolutional processing of one size corresponds to one hardware convolutional feature, to capture potential semantic features under different views). Furthermore, the hardware temporal adjustment features are fully connected to form hardware temporal fully connected features. The number of feature parameters included in the hardware temporal fully connected features is equal to the number of the multiple hardware convolutional features, i.e., they have a one-to-one correspondence.

[0065] Step d2: Based on the feature parameters in the hardware time-domain fully connected features, perform weighted mean calculation on the multiple hardware convolutional features, and based on the result of the weighted mean calculation and the hardware frequency-domain adjustment features, obtain a corresponding hardware initial operation feature.

[0066] In this embodiment of the invention, after obtaining the plurality of hardware convolutional features and the hardware temporal fully connected features, a weighted mean can be calculated on the plurality of hardware convolutional features based on the feature parameters in the hardware temporal fully connected features. Furthermore, based on the result of the weighted mean calculation and the hardware frequency domain adjustment features, a corresponding initial hardware running feature is obtained. For example, the result of the weighted mean calculation and the hardware frequency domain adjustment features can be added together to obtain the corresponding initial hardware running feature. It should be noted that since frequency domain data generally includes more detailed information, semantic feature capture of different perspectives is not necessary for the hardware frequency domain adjustment features to improve computational efficiency. However, the corresponding weighted mean calculation result can also be obtained by following the processing method for the hardware temporal adjustment features. Then, the results of the two weighted mean calculations in the temporal and frequency domains can be summed to obtain the initial hardware running feature. In other words, while capturing multiple hardware convolutional features under different views, a linear mapping can be achieved through fully connected processing. This allows the semantic features of the hardware temporal adjustment features to determine the degree of attention for different views, thus obtaining hardware temporal fully connected features. Specifically, the hardware temporal fully connected features can be normalized so that each feature parameter can be used to represent the weights, and the sum of each weight is equal to 1. Then, a weighted average can be calculated for the multiple hardware convolutional features based on each weight. In this way, adaptive attention to the semantic features of different views can be achieved.

[0067] Regarding step S122 in the above content, it should be further explained that the detailed process of performing deep semantic feature extraction on each of the initial hardware operation features is not limited. For example, in some feasible implementations, in order to capture more detailed information during the deep semantic feature extraction process, step S122 may further include the following sub-steps.

[0068] Step S122a: Perform depth mining in the first direction on the sliding window feature sequence of the initial hardware operation features to form the first hardware operation features.

[0069] In this embodiment of the invention, the sliding window feature sequence of the initial hardware operation features can be subjected to depth mining in a first direction to form a first hardware operation feature. That is, the initial hardware operation features can be segmented using a sliding window to form a more granular sliding window feature, thereby improving the accuracy of depth mining.

[0070] Step S122b: Perform second-direction depth mining on the sliding window feature sequence of the initial hardware operation features to form the second hardware operation features.

[0071] In this embodiment of the invention, the sliding window feature sequence of the initial hardware operation features can be subjected to second-direction depth mining to form second hardware operation features. The first-direction depth mining and the second-direction depth mining are two opposite-direction depth mining operations. For example, the first-direction depth mining can refer to mining from the first sliding window feature to the last sliding window feature in the sliding window feature sequence; the second-direction depth mining can refer to mining from the last sliding window feature to the first sliding window feature in the sliding window feature sequence. Thus, by using two different directions, the semantic features of the association between sliding window features can be fully captured.

[0072] Step S122c: The first hardware operation feature and the second hardware operation feature are fused to form a hardware operation depth feature.

[0073] In this embodiment of the invention, after obtaining the first hardware operation feature and the second hardware operation feature, the first hardware operation feature and the second hardware operation feature can be fused to form a hardware operation depth feature. For example, the first hardware operation feature and the second hardware operation feature can be added together to obtain the hardware operation depth feature.

[0074] Regarding step S122a in the above content, it should be further explained that the detailed process of performing first-direction depth mining on the sliding window feature sequence of the initial hardware operation features is not limited. For example, in some feasible implementations, in order to effectively capture the associated semantic features between different sliding window features, step S122a may further include the following sub-steps.

[0075] Step a1: Using the a-th first hardware operation feature, the (a+1)-th sliding window feature in the sliding window feature sequence of the initial hardware operation feature is focused on based on an attention mechanism to form a first hardware focus feature based on an attention mechanism.

[0076] In this embodiment of the invention, combined with Figure 6The attention-based first hardware focusing feature can be formed by using the a-th first hardware operation feature to focus the (a+1)-th sliding window feature in the sliding window feature sequence of the initial hardware operation features. Here, a is an integer greater than or equal to 0, and the 0th first hardware operation feature is formed during the training process of the corresponding neural network model (during training, the initial 0th first hardware operation feature can be a random vector or a vector of all "0"s, etc.). That is, based on the 0th first hardware operation feature (i.e., belonging to prior knowledge), cross-attention processing can be performed on the 1st sliding window feature, and the result of the cross-attention processing can be added to the 1st sliding window feature to obtain the attention-based first hardware focusing feature. Similarly, based on the 1st first hardware operation feature, cross-attention processing can be performed on the 2nd sliding window feature, and the result of the cross-attention processing can be added to the 2nd sliding window feature to obtain the attention-based first hardware focusing feature.

[0077] Step a2: Using the a-th first hardware operation feature, the (a+1)-th sliding window feature in the sliding window feature sequence of the initial hardware operation feature is focused on based on a gating mechanism to form a first hardware focusing feature based on a gating mechanism.

[0078] In this embodiment of the invention, the a-th first hardware operation feature can also be used to perform feature focusing based on a gating mechanism on the (a+1)-th sliding window feature in the sliding window feature sequence of the initial hardware operation feature, forming a first hardware focusing feature based on a gating mechanism. That is, based on the 0th first hardware operation feature, the 1st sliding window feature can be subjected to gating adjustment processing, and the result of the gating adjustment processing can be added to the 1st sliding window feature to obtain the first hardware focusing feature based on a gating mechanism. The gating adjustment process can be as follows: linearly mapping the a-th first hardware operation feature (which can be implemented through a fully connected network layer), then performing nonlinear activation on the obtained linearly mapped feature (e.g., using functions such as sigmoid) to obtain a gating parameter matrix, and then using each parameter in the gating parameter matrix as the weight of the corresponding position in the (a+1)-th sliding window feature for multiplication to achieve weighted summation.

[0079] Step a3: Using the a-th first hardware operation feature, the (a+1)-th sliding window feature in the sliding window feature sequence of the initial hardware operation feature is focused on based on a convolution mechanism to form the first hardware focusing feature based on the convolution mechanism.

[0080] In this embodiment of the invention, the a-th first hardware operation feature can also be used to perform convolution-based feature focusing on the (a+1)-th sliding window feature in the sliding window feature sequence of the initial hardware operation feature, forming a first hardware focusing feature based on the convolution mechanism. That is, based on the 0th first hardware operation feature, convolution focusing processing can be performed on the 1st sliding window feature, and the result of the convolution focusing processing can be added to the 1st sliding window feature to obtain the first hardware focusing feature based on the convolution mechanism. The convolution focusing processing can be performed by adding the a-th first hardware operation feature and the (a+1)-th sliding window feature, and then performing convolution processing on the result of the addition (the feature size remains unchanged after convolution processing).

[0081] Step a4: The first hardware focusing feature based on the attention mechanism, the first hardware focusing feature based on the gating mechanism, and the first hardware focusing feature based on the convolution mechanism are fused to form the (a+1)th first hardware operation feature.

[0082] In this embodiment of the invention, after obtaining the first hardware focusing feature based on the attention mechanism, the first hardware focusing feature based on the gating mechanism, and the first hardware focusing feature based on the convolution mechanism, the first hardware focusing feature based on the attention mechanism, the first hardware focusing feature based on the gating mechanism, and the first hardware focusing feature based on the convolution mechanism can be fused to form the (a+1)th first hardware operation feature. For example, the first hardware focusing features obtained based on the three mechanisms can be averaged to obtain the (a+1)th first hardware operation feature.

[0083] Step a5: Determine the final output first hardware operating feature based at least on the last formed first hardware operating feature.

[0084] In this embodiment of the invention, after obtaining each first hardware operating feature, the final output first hardware operating feature can be determined based at least on the last formed first hardware operating feature. For example, the last formed first hardware operating feature can be determined as the final output first hardware operating feature, or the average of each formed first hardware operating feature can be calculated to obtain the final output first hardware operating feature.

[0085] Regarding step S122b in the above content, it should be further explained that the detailed process of performing second-direction depth mining on the sliding window feature sequence of the initial hardware operation features is not limited. For example, in some feasible implementations, in order to effectively capture the associated semantic features between different sliding window features, step S122b may further include the following sub-steps.

[0086] Step b1: Using the bth second hardware operation feature, the (b-1)th sliding window feature in the sliding window feature sequence of the initial hardware operation feature is focused on based on an attention mechanism to form a second hardware focusing feature based on an attention mechanism.

[0087] In this embodiment of the invention, combined with Figure 7 The b-th second hardware operation feature can be used to focus the (b-1)-th sliding window feature in the sliding window feature sequence of the initial hardware operation features based on an attention mechanism, forming a second hardware focusing feature based on an attention mechanism, as described previously. Here, b is an integer greater than or equal to 2 and less than or equal to B. The B-th second hardware operation feature is formed during the training process of the corresponding neural network model (at the initial training moment, the B-th second hardware operation feature can be a random vector or a vector of all "0"s). In other words, the B-th second hardware operation feature can be used to focus the (B-1)-th sliding window feature (the last sliding window feature) based on an attention mechanism, forming a second hardware focusing feature based on an attention mechanism.

[0088] Step b2: Using the bth second hardware operation feature, the (b-1)th sliding window feature in the sliding window feature sequence of the initial hardware operation feature is subjected to feature focusing based on a gating mechanism to form a second hardware focusing feature based on a gating mechanism.

[0089] In this embodiment of the invention, the b-th second hardware operation feature can also be used to perform feature focusing based on a gating mechanism on the (b-1)-th sliding window feature in the sliding window feature sequence of the initial hardware operation feature, forming a second hardware focusing feature based on a gating mechanism, as described above. That is, the B-th second hardware operation feature can be used to perform feature focusing based on a gating mechanism on the (B-1)-th sliding window feature, forming a second hardware focusing feature based on a gating mechanism.

[0090] Step b3: Using the b-th second hardware operation feature, the (b-1)-th sliding window feature in the sliding window feature sequence of the initial hardware operation feature is focused on using a convolution mechanism to form a second hardware focusing feature based on a convolution mechanism.

[0091] In this embodiment of the invention, the b-th second hardware operation feature can also be used to perform feature focusing based on a convolution mechanism on the (b-1)-th sliding window feature in the sliding window feature sequence of the initial hardware operation feature, forming a second hardware focusing feature based on a convolution mechanism, as described above. That is, the B-th second hardware operation feature can be used to perform feature focusing based on a convolution mechanism on the (b-1)-th sliding window feature, forming a second hardware focusing feature based on a convolution mechanism.

[0092] Step b4: The second hardware focusing feature based on the attention mechanism, the second hardware focusing feature based on the gating mechanism, and the second hardware focusing feature based on the convolution mechanism are fused to form the (b-1)th second hardware operation feature.

[0093] In this embodiment of the invention, after obtaining the second hardware focusing feature based on the attention mechanism, the second hardware focusing feature based on the gating mechanism, and the second hardware focusing feature based on the convolution mechanism, the second hardware focusing feature based on the attention mechanism, the second hardware focusing feature based on the gating mechanism, and the second hardware focusing feature based on the convolution mechanism can be fused to form the (b-1)th second hardware operation feature, as described above.

[0094] Step b5: Determine the final output second hardware operating characteristics based at least on the last formed second hardware operating characteristics.

[0095] In this embodiment of the invention, after obtaining each formed second hardware operating feature, the final output second hardware operating feature can be determined based at least on the last formed second hardware operating feature, as described above.

[0096] Regarding step S130 in the above content, it should be further explained that the detailed process of updating the hardware operation aggregate features based on the correlation between the historical fault feature map and the hardware operation aggregate features is not limited and can be configured according to actual needs.

[0097] For example, in some feasible implementations, in order to ensure the reliability of updating the hardware operation aggregate features, that is, to effectively utilize the correlation between the historical fault feature map and the hardware operation aggregate features, so that the semantic representation accuracy of the formed hardware operation update features is higher, the above step S130 may further include the following sub-steps.

[0098] Step S131: For each graph member in the historical fault feature graph that represents the semantic features of historical hardware operation data, determine whether the hardware operation aggregate feature is related to the graph member based on the feature similarity between the graph member and the hardware operation aggregate feature.

[0099] In this embodiment of the invention, for each graph member (i.e., semantic feature) in the historical fault feature graph used to characterize historical hardware operation data, the correlation between the hardware operation aggregate feature and the graph member is determined based on the feature similarity (such as cosine similarity) between the graph member and the hardware operation aggregate feature. For example, if the feature similarity is greater than a similarity threshold, it is determined to be correlated; if the feature similarity is not greater than the similarity threshold, it is determined to be uncorrelated. This similarity threshold can be configured according to actual needs, such as values ​​like 0.8 or 0.9. Alternatively, one or more graph members with the highest feature similarity can be determined to be correlated.

[0100] Step S132: The hardware operation aggregation feature is used as a new graph member and connected with the relevant graph members to update the historical fault feature graph and form a new fault feature graph.

[0101] In this embodiment of the invention, combined with Figure 8 After making relevant judgments, the hardware operation aggregated features can be used as new graph members and connected with relevant graph members to update the historical fault feature graph and form a new fault feature graph. In other words, the new fault feature graph includes the historical fault feature graph and the hardware operation aggregated features.

[0102] Step S133: The new map member is used as the target for feature transfer and update in the new fault feature map to form hardware operation update features.

[0103] In this embodiment of the invention, after obtaining the new fault feature map, the new map members can be used as targets for feature transfer and updates, forming hardware operation update features. That is, based on the correlations in the new fault feature map, relevant semantic features are transferred to the new map members to obtain the hardware operation update features.

[0104] Regarding step S132 in the above content, it should be further explained that the detailed process of feature transfer and updating in the new fault feature map is not limited. For example, in some feasible implementations, in order to take into account both the semantic richness and semantic accuracy of the formed hardware operation update characteristics, the above step S132 may further include the following sub-steps.

[0105] Step S132a: Using the new graph member as the endpoint, traverse the new fault feature graph to form at least one graph member traversal path.

[0106] In this embodiment of the invention, combined with Figure 9 The new graph member can be used as the endpoint to traverse the new fault feature graph, forming at least one graph member traversal path. For example, each other graph member in the new fault feature graph can be used as the starting point, and then traversal and movement can be performed until the new graph member is reached, thus forming a corresponding graph member traversal path. After obtaining multiple graph member traversal paths corresponding to multiple other graph members, filtering can be performed. For example, if two graph member traversal paths have an inclusion relationship, the graph member traversal path that includes (i.e., includes fewer graph members) will be removed. In this way, at least one graph member traversal path can be formed, such as... Figure 8 The graph members include Graph Member 2, Graph Member 1, Graph Member 3, and new graph members.

[0107] Step S132b: For each of the graph member traversal paths, along the direction from the start point to the end point of the graph member traversal path, the graph members are sequentially associated and focused to form the hardware operation focus feature corresponding to the graph member traversal path.

[0108] In this embodiment of the invention, after obtaining the at least one graph member traversal path, for each graph member traversal path, each graph member can be sequentially associated and focused along the direction from the start point to the end point of the graph member traversal path to form the hardware operation focus feature corresponding to the graph member traversal path. For example, for the first graph member (starting point), self-attention processing can be performed on the graph member, and the result of the self-attention processing can be added to the graph member, and then normalization processing can be performed to obtain the first associated focus feature. Then, for each graph member after the first one, cross-attention processing can be performed on the graph member based on the previous associated focus feature, and the result of the cross-attention processing can be added to the graph member, and then normalization processing can be performed to obtain the corresponding associated focus feature, until the last associated focus feature is obtained, and the last associated focus feature is determined as the hardware operation focus feature.

[0109] Step S132c: Based on the hardware operation focus features corresponding to each of the graph member traversal paths, determine the hardware operation update features.

[0110] In this embodiment of the invention, after obtaining the hardware operation focus feature corresponding to each graph member traversal path, the hardware operation update feature can be determined based on the hardware operation focus feature corresponding to each graph member traversal path. For example, the hardware operation update feature can be obtained by calculating the mean of the hardware operation focus feature corresponding to each graph member traversal path.

[0111] Regarding step S140 in the above content, it should be further explained that the detailed process of feature restoration based on the hardware operation update features is not limited and can be configured according to actual needs.

[0112] For example, in some feasible implementations, the hardware operation update features can be fully connected to obtain corresponding fully connected features, wherein the size of the fully connected features can be 1. C, where C equals the number of components in the target device, or the number of components requiring fault diagnosis, or the number of components whose vibration, temperature, and sound change under abnormal conditions, exhibiting a one-to-one correspondence. Then, activation functions such as SigmaOb can be used to activate this fully connected feature, resulting in a corresponding activation parameter matrix. Finally, from this activation parameter matrix, each activation parameter greater than a participation threshold (e.g., 0.5, 0.6, etc.) can be identified, and the component corresponding to that activation parameter is determined as the faulty component. In this way, efficient and reliable location of faulty components in the target device can be achieved.

[0113] In summary, the fault diagnosis method and system for equipment operation and maintenance based on a large-scale model knowledge graph provided by this invention first acquires the target hardware operation data of the target device and obtains a historical fault feature map constructed based on at least one historical hardware operation data and corresponding fault tags. Second, semantic features are extracted and aggregated for each local hardware operation data in the target hardware operation data to form aggregated hardware operation features. Then, based on the correlation between the historical fault feature map and the aggregated hardware operation features, the aggregated hardware operation features are updated to form updated hardware operation features. Finally, feature restoration is performed based on the updated hardware operation features to form target fault diagnosis data. Based on this method, on the one hand, by utilizing the semantic features of at least two dimensions of local hardware operation data, the basis for fault diagnosis can be more sufficient and effective, thereby improving the reliability of fault diagnosis. On the other hand, since the hardware operation aggregate features formed by semantic feature extraction and semantic feature aggregation are updated based on historical fault feature maps, prior knowledge can be fully utilized. Compared with schemes that classify faults directly based on the relationship with prior knowledge, this scheme does not overly rely on prior knowledge while making use of it, but also fully considers the current data. The reliability of the target fault diagnosis data can be further improved. Therefore, it can improve the problem of relatively low reliability of fault diagnosis in the existing technology.

[0114] In the several embodiments provided in this invention, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus and method embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0115] In addition, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0116] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, electronic device, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks. It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. In the absence of further restrictions, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0117] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A fault diagnosis method for device operation and maintenance based on a large model knowledge graph, characterized by, include: Acquire target hardware operation data of the target device, and acquire a historical fault feature map constructed based on at least one historical hardware operation data and corresponding fault labels. The target hardware operation data includes local hardware operation data of different multiple dimensions, including at least two of vibration, temperature and sound. Each map member in the historical fault feature map is used to reflect the semantic features of the historical hardware operation data or the fault label. Semantic feature extraction and semantic feature aggregation are performed on each local hardware operation data in the target hardware operation data to form hardware operation aggregated features; Based on the correlation between the historical fault feature map and the hardware operation aggregate features, the hardware operation aggregate features are updated to form hardware operation update features. Based on the hardware operation update characteristics, feature restoration is performed to form target fault diagnosis data, wherein the target fault diagnosis data is used to reflect the faulty components in the target device.

2. The large model knowledge graph-based device operation and maintenance fault diagnosis method of claim 1, wherein, The step of extracting and aggregating semantic features for each local hardware operation data in the target hardware operation data to form aggregated hardware operation features includes: Initial semantic features are extracted from each local hardware operation data in the target hardware operation data to obtain the corresponding initial hardware operation features. Deep semantic features are extracted from each of the initial hardware operation features to obtain the corresponding hardware operation deep features. Based on the semantic correlation between each of the hardware operation depth features, the semantic features of multiple hardware operation initial features are aggregated to form hardware operation aggregate features.

3. The device operation and maintenance fault diagnosis method based on a large model knowledge graph according to claim 2, wherein, The step of extracting initial semantic features from each local hardware operation data in the target hardware operation data to obtain the corresponding initial hardware operation features includes: The hardware operation local data is subjected to one-dimensional convolution processing to form hardware temporal convolution features, and the temporal information extracted from the hardware operation local data is embedded to form hardware temporal embedding features. The temporal information includes at least one of mean, root mean square, peak value, kurtosis, skewness, peak factor, and impulse factor. The target device includes a wind turbine. Two-dimensional convolution processing is performed on the frequency domain data of the local data of hardware operation to form hardware frequency domain convolution features, and the frequency domain information extracted from the frequency domain data is embedded to form hardware frequency domain embedding features, wherein the frequency domain information includes at least one of the fundamental frequency, 2 times the fundamental frequency, and 3 times the fundamental frequency; Based on the hardware temporal embedding features, the hardware temporal convolutional features are constrained and adjusted to form hardware temporal adjusted features, and based on the hardware frequency domain embedding features, the hardware frequency domain convolutional features are constrained and adjusted to form hardware frequency domain adjusted features. The hardware time-domain adjustment features and the hardware frequency-domain adjustment features are fused to form a corresponding hardware initial operation feature.

4. The device operation and maintenance fault diagnosis method based on a large model knowledge graph according to claim 3, wherein, The step of fusing the hardware time-domain adjustment features and the hardware frequency-domain adjustment features to form a corresponding hardware initial operation feature includes: The hardware temporal adjustment features are subjected to various convolutional processes of different sizes to form multiple corresponding hardware convolutional features. The hardware temporal adjustment features are also subjected to fully connected processing to form hardware temporal fully connected features, wherein the number of feature parameters included in the hardware temporal fully connected features is equal to the number of the multiple hardware convolutional features. Based on the feature parameters in the hardware time-domain fully connected features, a weighted mean is calculated for the multiple hardware convolutional features, and based on the result of the weighted mean calculation and the hardware frequency-domain adjustment features, a corresponding hardware initial operation feature is obtained.

5. The device operation and maintenance fault diagnosis method based on a large model knowledge graph according to claim 2, wherein, The step of extracting deep semantic features for each of the initial hardware operation features to obtain the corresponding deep hardware operation features includes: The sliding window feature sequence of the initial hardware operation features is subjected to depth mining in the first direction to form the first hardware operation feature; The sliding window feature sequence of the initial hardware operation features is subjected to second-direction depth mining to form second hardware operation features, wherein the first-direction depth mining and the second-direction depth mining belong to two opposite-direction depth mining. The first hardware operation feature and the second hardware operation feature are fused together to form a hardware operation depth feature.

6. The device operation and maintenance fault diagnosis method based on a large model knowledge graph according to claim 5, wherein, The step of performing first-direction depth mining on the sliding window feature sequence of the initial hardware operation features to form the first hardware operation features includes: Using the a-th first hardware operation feature, the (a+1)-th sliding window feature in the sliding window feature sequence of the initial hardware operation feature is focused on the feature based on the attention mechanism to form the first hardware focus feature based on the attention mechanism, where a is an integer greater than or equal to 0, and the 0th first hardware operation feature is formed during the training process of the corresponding neural network model. Using the a-th first hardware operation feature, the (a+1)-th sliding window feature in the sliding window feature sequence of the initial hardware operation feature is subjected to feature focusing based on a gating mechanism to form the first hardware focusing feature based on the gating mechanism; Using the a-th first hardware operation feature, the (a+1)-th sliding window feature in the sliding window feature sequence of the initial hardware operation feature is subjected to feature focusing based on the convolution mechanism to form the first hardware focusing feature based on the convolution mechanism; The first hardware focusing feature based on the attention mechanism, the first hardware focusing feature based on the gating mechanism, and the first hardware focusing feature based on the convolution mechanism are fused to form the (a+1)th first hardware operation feature. The final output first hardware operating characteristic is determined based at least on the last formed first hardware operating characteristic.

7. The device operation and maintenance fault diagnosis method based on a large model knowledge graph according to claim 5, wherein, The step of performing second-direction depth mining on the sliding window feature sequence of the initial hardware operation features to form second hardware operation features includes: Using the b-th second hardware operation feature, the (b-1)-th sliding window feature in the sliding window feature sequence of the initial hardware operation feature is focused on based on an attention mechanism to form a second hardware focus feature based on an attention mechanism, where b is an integer greater than or equal to 2 and less than or equal to B, and the B-th second hardware operation feature is formed during the training process of the corresponding neural network model. Using the b-th second hardware operation feature, the (b-1)-th sliding window feature in the sliding window feature sequence of the initial hardware operation feature is subjected to feature focusing based on a gating mechanism to form a second hardware focusing feature based on a gating mechanism; Using the b-th second hardware operation feature, the (b-1)-th sliding window feature in the sliding window feature sequence of the initial hardware operation feature is subjected to feature focusing based on the convolution mechanism to form the second hardware focusing feature based on the convolution mechanism; The second hardware focusing feature based on the attention mechanism, the second hardware focusing feature based on the gating mechanism, and the second hardware focusing feature based on the convolution mechanism are fused to form the (b-1)th second hardware operation feature. The final output second hardware operating characteristics are determined based at least on the last formed second hardware operating characteristics.

8. The device operation and maintenance fault diagnosis method based on a large model knowledge graph according to any one of claims 1-7, characterized in that, The step of updating the hardware operation aggregate features based on the correlation between the historical fault feature map and the hardware operation aggregate features to form updated hardware operation features includes: For each graph member in the historical fault feature graph that represents the semantic features of historical hardware operation data, based on the feature similarity between the graph member and the hardware operation aggregate feature, it is determined whether the hardware operation aggregate feature is related to the graph member. The hardware operation aggregation features are used as new graph members and connected with related graph members to update the historical fault feature graph and form a new fault feature graph. The new map members are used as targets for feature transfer and updates in the new fault feature map to form hardware operation update features.

9. The device operation and maintenance fault diagnosis method based on a large model knowledge graph according to claim 8, wherein, The step of using the new map members as the target for feature transfer and update in the new fault feature map to form hardware operation update features includes: Using the new graph member as the endpoint, traverse the new fault feature graph to form at least one graph member traversal path; For each of the graph member traversal paths, along the direction from the start point to the end point of the graph member traversal path, each graph member is sequentially associated and focused to form the hardware operation focus feature corresponding to the graph member traversal path. Based on the hardware operation focus features corresponding to each of the graph member traversal paths, the hardware operation update features are determined.

10. A fault diagnosis system for equipment operation and maintenance based on a large-scale knowledge graph model, characterized in that, It includes a processor and a memory, the memory being used to store a computer program, and the processor being used to execute the computer program to implement the fault diagnosis method for equipment operation and maintenance based on a large model knowledge graph as described in any one of claims 1-9.

Images