A fault root cause positioning method and device for device value attenuation

Abstract

The application provides a device value attenuation-oriented fault root cause positioning method and device, and relates to the technical field of computers. The method comprises the following steps: obtaining a residual value index RVI of a target device and a device operation log, wherein the residual value index RVI is used for quantifying the economic value attenuation degree of the device; obtaining a state feature vector according to the residual value index RVI and obtaining a text feature vector according to the device operation log; taking the state feature vector as a query vector, taking the text feature vector as a key vector and a value vector, and obtaining a context feature vector through an attention mechanism; and obtaining a fault root cause category of the residual value index RVI change of the target device according to the context feature vector. The application can significantly improve the discovery efficiency, root cause positioning accuracy and operation and maintenance intelligent level of the device value attenuation-related fault, and realize automatic reasoning from value anomaly to explainable fault cause.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method and apparatus for locating the root cause of equipment value degradation. Background Technology

[0002] Currently, for root cause localization of equipment value degradation, existing technologies based on inter-component fault correlation analysis can only discover statistical correlations between components, failing to establish a reasoning link between macroeconomic value indicators and micro-log semantics, and thus struggling to answer the business question of what specific fault caused value degradation. Furthermore, while symmetric attention architectures based on multimodal data fusion involve equal participation of features from all modalities, they lack the technical means to proactively guide log analysis with value status, making it difficult to focus on key logs explaining value anomalies. In addition, traditional methods rely on manual review of massive amounts of unstructured logs and maintenance reports, resulting in low efficiency and high dependence on expert experience, making automated and intelligent root cause localization difficult.

[0003] Therefore, there is an urgent need for an intelligent root cause localization method that can automatically output explanatory root causes and related evidence of failures, guided by the equipment residual value index. Summary of the Invention

[0004] This invention provides a method and system for locating the root cause of equipment value degradation. This method can significantly improve the efficiency of detecting equipment value degradation-related faults, the accuracy of root cause location, and the level of intelligent operation and maintenance.

[0005] Firstly, a method for locating the root cause of equipment value degradation is provided, including: Obtain the Residual Value Index (RVI) and equipment operation logs of the target equipment. The Residual Value Index (RVI) is used to quantify the degree of economic value decay of the equipment. Based on the Residual Value Index (RVI), a state feature vector is obtained; based on the equipment operation log, a text feature vector is obtained. By using the state feature vector as the query vector and the text feature vector as the key and value vectors, an attention mechanism is used to obtain the context feature vector. Based on the context feature vector, the root cause category of the failure is obtained for the change in the residual value index (RVI) of the target equipment.

[0006] In this way, by encoding the Residual Value Index (RVI), which reflects the decline in the economic value of equipment, into a state feature vector and using it as the unique query vector for the attention mechanism, and simultaneously encoding the equipment operation log into a text feature vector as a key and value, the asymmetric attention architecture actively guides the semantic analysis of logs and the association between value state. Finally, it outputs a specific fault root cause category that can directly explain the value decline, realizing automated reasoning from "value anomaly" to "explainable fault cause". This overcomes the shortcomings of existing technologies that lack value state guidance and cannot establish a correlation between macro indicators and micro logs, and significantly improves the intelligence level and accuracy of fault root cause localization.

[0007] In some possible implementations, the state feature vector is used as the query vector, and the text feature vector is used as the key and value vectors. An attention mechanism is then used to obtain the context feature vector, including: The attention mechanism is used to calculate the similarity between the state feature vector and each text feature vector, and the similarity is normalized to obtain the attention weight. The context feature vector is obtained by weighting and summing the text feature vectors according to the attention weights.

[0008] In this way, by calculating the similarity between the value state features and each log feature and normalizing it to obtain the attention weight, and then performing a weighted summation of the log features, the log information with high attention weight is significantly enhanced, while the log with low weight is suppressed, thereby generating a context feature vector focused on the explanation of value decay, providing high-quality feature input for subsequent root cause classification.

[0009] In some possible implementations, an attention mechanism is used to calculate the similarity between the state feature vector and each text feature vector, and the similarity is normalized to obtain the attention weights, including: The original similarity score is obtained by calculating the dot product between the status feature vector and the text feature vector of each device operation log. The original similarity score is scaled according to the dimension of the text feature vector to obtain the scaled score; Based on the scaled score, the attention weight of each device operation log is determined.

[0010] In this way, the attention weights are calculated by scaling the dot product. On the one hand, the dot product efficiently measures the similarity between the state vector and the log vector. On the other hand, scaling by dividing by the square root of the dimension avoids the softmax gradient from becoming too large, thus ensuring the rationality of the attention weight distribution and the stability of training.

[0011] In some possible implementations, the dot product between the state feature vector and the text feature vector of each device operation log is calculated to obtain the original similarity score; the original similarity score is scaled according to the dimension of the text feature vector to obtain the scaled score; based on the scaled score, the attention weight of each device operation log is determined, including: The attention weights are calculated using the following formula: Where Q is the state feature vector. For the first i The text feature vector corresponding to each device operation log. For feature dimension, For the first i Attention weights for each device operation log.

[0012] In this way, the calculation process of attention weight is defined by a clear mathematical formula, which makes the contribution of each log to the explanation of value decay quantifiable and traceable, providing a foundation for the extraction of supporting evidence and the interpretability of root cause localization.

[0013] In some possible implementations, the root cause category of the change in the residual value index (RVI) of the target device is obtained based on the context feature vector, including: The context feature vector is input into the classifier, which outputs the root cause category of the fault and its confidence level. The output category of the classifier includes at least one of the following: heat dissipation system failure, power system failure, network connection problem, storage medium damage, software configuration error, hardware aging, and external environmental interference.

[0014] In this way, by pre-setting root cause categories that cover common equipment failures, the classifier automatically identifies the most likely failure type and confidence level from the context feature vector that integrates value state guidance information, realizing end-to-end mapping from multi-source heterogeneous data to specific failure root causes, avoiding the tedious process of manually checking logs and making experience-based judgments.

[0015] In some possible implementations, the state feature vector is obtained based on the residual value index (RVI), including: The residual value index RVI is mapped to a state feature vector of a preset dimension through a multilayer perceptron, wherein the multilayer perceptron includes at least one fully connected layer and a nonlinear activation function.

[0016] In this way, by mapping the single-value residual value index (RVI) to a high-dimensional state feature vector through a multilayer perceptron, the feature representation capability of the value state is enhanced, enabling it to perform effective dot product similarity calculation with the text feature vector. At the same time, the introduction of a non-linear activation function improves the model's ability to fit the non-linear changes in value decay.

[0017] In some possible implementations, text feature vectors are obtained based on device operation logs, including: The device operation logs are input into a pre-trained language model to extract the semantic feature vector of each device operation log. The pre-trained language model is BERT.

[0018] In this way, by leveraging the powerful semantic understanding capabilities of pre-trained language models such as BERT, the original natural language logs are transformed into semantic feature vectors rich in contextual information, providing high-quality text representations for subsequent interactions with value state features, and significantly improving the model's ability to extract hidden fault information from the logs.

[0019] Secondly, a fault root cause localization device for equipment value degradation is provided, comprising: The first processing module is used to obtain the Residual Value Index (RVI) of the target equipment and the equipment operation log. The Residual Value Index (RVI) is used to quantify the degree of economic value decay of the equipment. The second processing module is used to obtain the state feature vector based on the Residual Value Index (RVI) and the text feature vector based on the equipment operation log. The third processing module is used to take the state feature vector as the query vector and the text feature vector as the key vector and value vector, and obtain the context feature vector through an attention mechanism. The fourth processing module is used to obtain the root cause category of the change in the residual value index (RVI) of the target device based on the context feature vector.

[0020] In some possible implementations, the first processing module is specifically used to: obtain the residual value index (RVI) from the equipment management system, where RVI is a continuous value between [0,1], which comprehensively reflects the residual value of the equipment, predicted maintenance costs, and technological gaps; and obtain the equipment operation log text from the equipment log library, where the log text includes at least one of operation status records, alarm information, and error reports.

[0021] In some possible implementations, the second processing module is specifically used for: The residual value index (RVI) is input into a multilayer perceptron, which includes at least one fully connected layer and a nonlinear activation function, and mapped to a state feature vector of a preset dimension. The device operation logs are input into a pre-trained language model, which can be BERT, RoBERTa, or DistilBERT, and the semantic feature vectors of each log are extracted to form a set of text feature vectors.

[0022] In some possible implementations, the third processing module is specifically used for: The state feature vector is used as the unique query vector in the attention mechanism, and the set of text feature vectors is used as the key and value in the attention mechanism. The dot product between the state feature vector and each text feature vector is calculated to obtain the original similarity score. The original similarity score is scaled according to the dimension of the text feature vector to obtain the scaled score. The scaled score is normalized by the softmax function to obtain the attention weight of each log. The text feature vectors are weighted and summed according to the attention weight to generate the context feature vector.

[0023] In some possible implementations, the third processing module is specifically used to calculate the attention weights using the following formula: Where Q is the state feature vector. For the first i The text feature vector corresponding to each device operation log. For feature dimension, For the first i Attention weights for each device operation log; Context feature vector ,in Let the text feature vector corresponding to the i-th device operation log be used as the value vector.

[0024] In some possible implementations, the fourth processing module is specifically used to: input the context feature vector into a fully connected classifier, the output layer of the classifier adopts the softmax activation function, and outputs the root cause category of the fault and its confidence level; the output category of the classifier includes at least one of the following: heat dissipation system failure, power system failure, network connection problem, storage medium damage, software configuration error, hardware aging, external environmental interference, and others / unknown.

[0025] In some possible implementations, the device also includes: a report generation module, used to extract log entries with attention weights greater than a preset threshold as supporting evidence for the root cause category of the fault, and to generate an analysis report containing the root cause category of the fault, confidence level, supporting evidence and attention weight visualization.

[0026] Thirdly, an electronic device is provided, comprising: one or more processors; one or more memories; and one or more programs, wherein the one or more programs are stored in the one or more memories, and the one or more programs include instructions that, when executed by the one or more processors, cause the configured device to perform the method as described in the first aspect.

[0027] Fourthly, a computer storage medium is provided, the computer storage medium storing instructions that, when executed by a computer, cause the computer to perform the method described in the first aspect or the second aspect.

[0028] Fifthly, a computer program product is provided, the computer program product storing instructions that, when executed by a computer, cause the computer to perform the method described in the first aspect or the second aspect.

[0029] The beneficial effects of the second to fifth aspects can be referred to the introduction of the beneficial effects of the first aspect above, and will not be repeated here. Attached Figure Description

[0030] Figure 1 This is an overall architecture diagram of the fault root cause localization method for equipment value decay provided in the embodiments of the present invention; Figure 2 This is a flowchart illustrating the internal structure of a state-log correlation analysis model according to an embodiment of the present invention; Figure 3 This is a comparison diagram of the symmetric fusion architecture and the asymmetric bootstrapping architecture shown in the embodiments of the present invention; Figure 4 This is an example diagram illustrating a device failure root cause analysis report according to an embodiment of the present invention; Figure 5 This is a flowchart illustrating a fault root cause localization method for equipment value degradation according to an embodiment of the present invention. Figure 6 This is a schematic diagram of a fault root cause localization device for equipment value degradation that can be used to implement the method of the present invention. Figure 7 This is a schematic diagram of an electronic device provided by the present invention. Detailed Implementation

[0031] The solutions provided by the embodiments of the present invention will now be described with reference to the accompanying drawings. In the embodiments of the present invention, "multiple" refers to two or more objects, and "various kinds" refers to two or more types. Terms such as "first," "second," etc., are only used to distinguish similar objects and are not necessarily used to describe a specific order or number of objects.

[0032] Currently, industrial equipment undergoes performance degradation and value decline during long-term operation. The Residual Value Index (RVI) comprehensively reflects the equipment's residual value, predicts maintenance costs, and assesses technological advancements, serving as a key indicator for guiding asset maintenance and disposal decisions. However, when the system detects a significant decrease in the RVI, maintenance personnel still need to manually review massive amounts of unstructured operation logs and maintenance reports to pinpoint the root cause. Existing technologies have the following significant limitations.

[0033] Root cause analysis based on correlation discovery mines the probability of fault correlation by analyzing the time-series information between components and the textual features of alarm logs. However, this method solves the problem of "fault correlation discovery between components". Its core is correlation statistics rather than interpretive root cause localization. It does not establish a correlation reasoning link between the macro-value status of the equipment and the micro-log semantics, and cannot answer the business question of "what specific fault caused the current value decay".

[0034] The fault diagnosis scheme based on multimodal spatiotemporal dependency perception uses the UGFormer algorithm to fuse the Transformer and graph neural networks for fault detection. However, this scheme adopts a symmetrical fusion architecture, in which features of each modality, such as performance indicators, logs, and topology, participate equally in attention calculation. It lacks technical means to actively guide the direction of log analysis with value states, resulting in key log signals that are strongly correlated with value decay being overwhelmed by a large amount of irrelevant information.

[0035] In summary, existing technologies do not use indicators reflecting the economic value of equipment, such as the Residual Value Index (RVI), as the core guiding signal for analysis. They cannot achieve automated reasoning from value anomalies to explainable causes of failures. As a result, maintenance personnel still need to spend a lot of time manually troubleshooting when faced with low RVI values, making it impossible to accurately locate the root cause of the failure. This seriously restricts the level of intelligence in predictive maintenance.

[0036] To address the technical problems in existing technologies, such as the lack of an automated explanation link between value decay state and fault root cause, and the inability of multimodal fusion architectures to proactively guide log analysis based on value state, this invention proposes a fault root cause localization method and apparatus for equipment value decay. This method uses the equipment Residual Value Index (RVI) as the query vector. Through an asymmetric attention mechanism, it uses the state features encoded by the RVI as the query, and the text features of the equipment's operating logs from the same period as the key and value. It calculates the explanatory contribution of each log entry to the current value decay state, weights and fuses them, and then outputs the specific fault root cause category through a classifier. Log entries with high attention weights are automatically extracted as supporting evidence. This invention constructs a unidirectional proactive guidance inference link from value anomaly to fault explanation, avoiding the defect of key signals being submerged by noise in existing symmetric fusion architectures. It achieves automated and interpretable root cause localization, significantly improving the accuracy of predictive maintenance and operational efficiency.

[0037] Figure 1 This is an overall architecture diagram of the fault root cause localization method for equipment value degradation provided in an embodiment of the present invention. Figure 1 As shown, the overall architecture consists of three modules: a data input module, a status-log correlation analysis engine, and a report generation module.

[0038] The data input module is used to receive two types of input data: one is the residual value index (RVI) of the target equipment, which is a continuous value between [0,1], comprehensively reflecting the residual value of the equipment, the predicted operation and maintenance cost, and the technological gap; the other is the equipment operation log text that is in the same period as the RVI, which includes various alarms, errors and status information generated during the operation of the equipment.

[0039] The state-log correlation analysis engine includes a text encoder, a state feature encoder, an asymmetric attention fusion layer, and a classification output layer. For example, the text encoder can use a pre-trained BERT model to convert each log text into a high-dimensional text feature vector. The state feature encoder can use a multilayer perceptron (MLP) to map RVI values ​​to state feature vectors. The asymmetric attention fusion layer uses the state feature vector as the query vector and all log text feature vectors as keys and values. Through an attention mechanism, it calculates the explanatory contribution of each log entry to the current value decay state and weights them to obtain a context feature vector. The classification output layer takes the context feature vector as input to the classifier and outputs the specific fault root cause category.

[0040] The report generation module outputs a readable root cause analysis report. The report includes: the root cause name and confidence level, such as "Cooling system failure" with a confidence level of 92%; and high-attention-weighted log entries as supporting evidence, such as "Low fan speed log". In some embodiments, a visualization of the attention weight distribution may also be included to help operations personnel understand the reasoning behind the model.

[0041] The following describes a fault root cause localization method for equipment value degradation, as illustrated in an embodiment of the present invention, specifically including the following steps: In step S1, the residual value index (RVI) of the target device and its corresponding device operation log text are obtained.

[0042] In this embodiment of the invention, the data input module acquires two types of input data: one type is the Residual Value Index (RVI). This index comprehensively reflects the residual value of the equipment, predicted maintenance costs, and technological gaps. The lower the RVI, the more severe the decline in the economic value of the equipment. There are many specific methods for calculating RVI, which will not be elaborated upon in this invention. Furthermore, in this embodiment, pre-calculated RVI values ​​can be directly used. For example, taking a 4G base station BBU as an example, its current RVI value is 0.28.

[0043] The second input data is the device operation log text. The log is the device operation log within a time window corresponding to the RVI calculation cycle. The log can be semi-structured or unstructured text data, and each log entry includes a timestamp, log level (e.g., Info, Warning, Error), and a detailed description.

[0044] For example, the device collected a total of 127 log records. Key log entries are shown below: "Warning: Fan speed below threshold 1500 RPM".

[0045] "Error: CPU core temperature exceeded 85°C, throttling initiated".

[0046] "Info: Power supply unit output voltage fluctuation detected +5% to -8%".

[0047] "Info: Memory usage normal, 45% utilized".

[0048] The log text is preprocessed, including removing information irrelevant to semantic analysis such as timestamps, IP addresses, and device serial numbers, converting them to lowercase characters, and organizing each log entry into an independent text string.

[0049] In step S2, feature encoding is performed on the RVI and log text to obtain state feature vectors and text feature vector sets, respectively.

[0050] In this embodiment of the invention, firstly, a text encoder can be used to extract semantic features from the preprocessed device operation logs. This embodiment uses a pre-trained bidirectional encoder representation model (BERT) as the text encoder. Based on the Transformer architecture, the BERT model can capture deep semantic information in the log text through bidirectional context modeling, converting each natural language log entry into a dense semantic vector of fixed dimensions.

[0051] Specifically, the BERT-base model can also be used. This model contains 12 Transformer encoder layers, 12 attention heads, and an output dimension of 768. For a device log text containing n logs, each log is input into the BERT model. The model takes the vector representation of the last layer's token as the semantic feature of that log, or pools the output of all tokens. After encoding, a set of text feature vectors is obtained. These text feature vectors will simultaneously play the roles of key and value in the subsequent asymmetric attention mechanism, both for calculating similarity with the state feature vector and for carrying log content information to participate in weighted fusion.

[0052] In other alternative embodiments, recurrent neural networks such as Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU) can be used to encode the log sequence, or lightweight or enhanced pre-trained models such as DistilBERT and RoBERTa can be used to balance efficiency and accuracy. This embodiment does not limit this.

[0053] Secondly, the state encoder maps the Residual Value Index (RVI) to a high-dimensional state feature vector. RVI is a continuous numerical value ranging from [0,1], with lower values ​​indicating more severe degradation of the equipment's economic value. To match the high-dimensional text features with the query vector in the attention mechanism, the scalar RVI needs to be extended to a high-dimensional semantic space.

[0054] This embodiment uses a multilayer perceptron (MLP) to implement this mapping. The structure of the MLP is as follows: the input layer receives the scalar value of RVI; the hidden layer contains 128 neurons, using the modified linear unit ReLU as the activation function to introduce nonlinear transformation capability, enabling the model to learn the complex mapping relationship between RVI and log semantics; the output layer further transforms the 128-dimensional features of the hidden layer and outputs the state feature vector.

[0055] In the example, RVI=0.28 is processed by the state encoder to obtain the state feature vector V_status. This state feature vector will serve as the sole query vector Query in the subsequent attention mechanism, used to actively query and focus on the log features most relevant to the current value decay state. The size of the state feature vector can be designed by combining the dimension of the text feature vector and the computational cost; 128 dimensions ensure sufficient expressive power without significantly increasing model complexity.

[0056] In other alternative embodiments, multidimensional state vectors, such as combinations of RVI and health scores, time-series state sequences, RVI values ​​at multiple time points, or state change rates, can also be input as encoder inputs. This embodiment does not limit this.

[0057] Through the two encoding steps described above, this embodiment of the invention unifies the structured macro-value indicator RVI and the unstructured micro-log text into the same feature space.

[0058] In step S3, the explanatory contribution of each log to the current value decay state is calculated through an asymmetric state-guided attention mechanism, and a fusion context feature vector is generated.

[0059] In this embodiment of the invention, an asymmetric attention architecture is constructed, using the state feature vector of the Residual Value Index (RVI) as the unique query and the text feature vectors of all logs as both the key and value. This architecture breaks the pattern of equal interaction among modalities in traditional symmetric fusion, realizing a unidirectional attention flow that actively guides log analysis through state, thereby enabling the model to focus on the log information that is most explanatory to value decay.

[0060] First, the state feature vector This vector serves as the query in the attention mechanism. It characterizes the current macroscopic value decay state of the target device. The text feature vector set... It also serves as both the key and value in the attention mechanism.

[0061] This achieves a unidirectional attention flow that proactively guides log analysis through state. The query originates solely from the state modality, while the key and value originate solely from the log modality; their roles are strictly asymmetrical. This asymmetrical structure ensures that the value state is dominant, while the log content is in a subordinate position, guided and focused.

[0062] Next, regarding the first i The attention weights for each log entry are calculated using a scaled dot product attention mechanism, as shown in the following formula: in: This is the state feature vector (Query). For the first i Key, the text feature vector of a log entry. This represents the dimension of the feature vector. In this embodiment, it is set to 128, but in actual calculations, it can be calculated first. and Projected onto the same dimension. The dot product similarity measures the degree of matching between the log and the current value state; the denominator is the normalization factor of the similarity of all logs, ensuring that the sum of all weights is 1. This is a scaling factor to prevent the softmax gradient from saturating due to excessively large dot product values.

[0063] Calculated attention weights Direct characterization of the first i The weight of each log entry contributes to the interpretation of the current RVI value. A higher weight indicates a stronger causal relationship between the log entry and value decay, and it is more likely to be the root cause of the RVI decline.

[0064] Taking 127 log entries from a 4G base station BBU as an example, the calculated partial attention weight distribution is as follows: Table 1. Attention weight distribution corresponding to logs The distribution in Table 1 clearly shows that, of the two logs related to heat dissipation, low fan speed and CPU overheating contributed the most to the explanation of value decay, accounting for over 75%, while the normal memory log contributed almost nothing. This is highly consistent with the final diagnostic conclusion of heat dissipation system failure, verifying the interpretability of attention weights.

[0065] Then, the text feature vector of each log entry is... We obtain a context feature vector that incorporates state guidance information by weighting and summing the features according to their attention weights. This weighted summation operation significantly enhances the semantic information of high-contribution logs while effectively suppressing information from low-contribution or irrelevant logs. C Essentially, it is a selective aggregation result of the original log semantic features under the guidance of value status. It not only retains the key log content related to value decay, but also provides quantifiable explanatory basis through weight allocation.

[0066] The entire calculation process described above can be uniformly represented in the form of a standard attention mechanism: Among them, matrix From text feature vectors Stacked together. This expression clearly shows that the embodiments of the present invention use state feature vectors. For a unique query, using all log features as keys and values, the algorithm actively retrieves the information most relevant to value decay from massive logs through scaling dot product attention operations, and aggregates it into a context feature vector for subsequent root cause classification.

[0067] Thus, through the embodiments of the present invention, an end-to-end interpretable correlation is established between macroeconomic value indicators and microlog semantics, providing core computational support for automated and intelligent fault root cause localization.

[0068] Figure 2 This is a flowchart illustrating the internal structure of a state-log correlation analysis model according to an embodiment of the present invention. Figure 2As shown, the model consists of five parts: state feature encoding, text feature encoding, attention weight calculation, context feature fusion, and classification output layer.

[0069] In the state feature coding section, the residual value index (RVI) of the target device is set to 0.28, which is then mapped into a state feature vector by a multilayer perceptron. This vector serves as the unique query vector Query in the attention mechanism.

[0070] The text feature encoding part receives device operation logs, such as log 1, log 2, log 3, ..., log N, and encodes each log entry into a text feature vector using a pre-trained BERT model. This forms a set of text feature vectors, which serve as both keys and values ​​in the attention mechanism.

[0071] The attention weight calculation part will use the state feature vector As the unique query vector, it is used in conjunction with the feature vectors of each text. Perform dot product operations and calculate the attention weight for each log entry using a scaled dot product attention mechanism. Attention weights It directly represents the contribution of the corresponding log to the explanation of the current device value decay state.

[0072] The context feature fusion part is based on the calculated attention weights. For the corresponding text feature vector A weighted summation is performed to generate a contextual feature vector that incorporates value state guidance information. .

[0073] The classification output layer will use the context feature vector C Input a fully connected classifier and output the probability distribution of each root cause category. Example output: probability of heat dissipation system failure 0.92, probability of power system failure 0.05, probability of network connection problem 0.02, probability of other categories 0.01, and finally output heat dissipation system failure as the root cause category.

[0074] The entire process uses the Residual Value Index (RVI) as a guiding signal, actively focuses on relevant log semantics through an asymmetric attention mechanism, and finally outputs a complete inference chain that can explain the root cause of the failure.

[0075] Figure 3 This is a comparison diagram of the symmetric fusion architecture and the asymmetric bootstrap architecture shown in the embodiments of the present invention.

[0076] The left side shows a symmetric fusion architecture in existing technologies. This architecture extracts features from multiple modalities, such as metrics, logs, and images, and processes them equally through a symmetric fusion network. There is no hierarchy among the features of each modality. The fused features are then input into a classifier for decision-making. This architecture fails to highlight the guiding role of one modality on another, leading to the suppression of key information and making it difficult to achieve targeted root cause explanations.

[0077] The right side illustrates the asymmetric guidance architecture proposed in this embodiment of the invention. It uses only the Remaining Value Index (RVI) as the sole guiding signal, obtaining the query vector Query through state encoding. Simultaneously, it uses the device operation log as the object of analysis, obtaining the key and value through text encoding. Then, through an asymmetric attention mechanism, the RVI state actively guides the analysis direction of log features, and the finally fused contextual features are used for classification output. This architecture achieves unidirectional active guidance of log semantics by the value state, focusing on key logs that explain value decay, thereby outputting explainable root causes of failures.

[0078] As can be seen from the comparison, the embodiments of the present invention break through the limitation of equal treatment of each modality in traditional symmetric fusion, establish an asymmetric reasoning link dominated by economic value indicators, and significantly improve the pertinence and interpretability of fault root cause localization.

[0079] In step S4, the context feature vector is input into the classifier, the root cause category of the fault is output, and an analysis report is generated.

[0080] In this embodiment of the invention, the context feature vector is input into a fully connected classifier to determine the specific fault root cause category under the current value decay state.

[0081] Specifically, the classifier can be structured as follows: the input layer receives a context feature vector C, contains a fully connected hidden layer with 128 neurons, and uses the ReLU activation function to introduce non-linearity; the output layer is a Softmax layer, with the number of nodes corresponding to the preset total number of fault root cause categories. The Softmax function converts the raw score of each output node into a probability distribution, where the sum of the probabilities of all categories is 1.

[0082] By mapping contextual feature vectors that incorporate value state guidance information to a predefined fault category space, the classifier can automatically learn the deep correlation between log semantics and value decay, thereby outputting the most interpretable fault category.

[0083] The specific root cause categories preset in this embodiment are: heat dissipation system failure, such as fan failure or heat sink blockage; power system failure, such as unstable voltage or damaged power supply module; network connection problems, such as link interruption or signal interference; storage medium damage, such as hard drive bad sectors or flash memory failure; software configuration errors, such as improper parameter settings or version incompatibility; hardware aging, such as capacitor aging or chip performance degradation; external environmental interference, such as temperature fluctuations or electromagnetic interference; and other undefined fault conditions.

[0084] For the 4G base station BBU in this example, the probability distribution output by the model is as follows: heat dissipation system failure 0.92, power system failure 0.05, and other categories, including network connectivity issues and storage media damage, totaling 0.03. Since the probability of heat dissipation system failure is much higher than other categories, the model ultimately outputs this root cause as the diagnostic conclusion. Simultaneously, several log entries with the highest attention weights are retained, such as "Fan speed below threshold 1500 RPM" with a weight of 0.42 and "CPU core temperature exceeded 85°C" with a weight of 0.35, as direct evidence supporting this diagnostic conclusion.

[0085] Furthermore, this classifier is trained through supervised learning. During the training phase, a large number of labeled context feature vectors and real failure root cause samples are used, and the network parameters are optimized using the cross-entropy loss function and backpropagation algorithm. After training, the classifier can quickly and accurately identify the root cause of new input context feature vectors, achieving automated and intelligent localization of the causes of equipment value degradation.

[0086] Figure 4 This is an example diagram illustrating a device failure root cause analysis report according to an embodiment of the present invention. Figure 4 As shown, the report comprises five parts: equipment information, diagnostic conclusions, supporting evidence, attention weight visualization, and recommended actions.

[0087] The equipment information section displays basic information about the analyzed equipment, including equipment type (4G base station BBU), equipment ID (BBU-2021-0372), current residual value index (RVI) of 0.28, and analysis time. The diagnostic conclusion section outputs the system's primary failure root cause category as "heat dissipation system failure," with a 92% confidence level. The supporting evidence section lists log entries that contribute most to the explanation of value decay, sorted by explanatory contribution from highest to lowest. Each log entry includes the original text, timestamp, and corresponding attention weight. For example, a log entry with a contribution of 0.42 is "Warning: Fan speed belowthreshold 1500 RPM," a log entry with a contribution of 0.35 is "Error: CPU core temperature exceeded 85°C," and a log entry with a contribution of 0.15 is "Info: Thermal throttling initiated."

[0088] The attention weight visualization section displays each key log and its corresponding attention weight in a list format, making it easy for users to intuitively understand the degree of influence of each log on the diagnostic conclusion.

[0089] The recommended actions section generates specific handling instructions based on the root cause category, including actions such as checking the cooling fan, cleaning dust, and replacing the damaged cooling module.

[0090] The report transforms the abstract attention weights and classification results output by the model into structured information that operations and maintenance personnel can directly understand, achieving a seamless connection from root cause localization to handling guidance, and significantly improving operations and maintenance efficiency and operability.

[0091] In summary, this embodiment of the invention encodes the Residual Value Index (RVI), which reflects the decline in the economic value of equipment, into a state feature vector and uses it as the unique query vector for the attention mechanism. Simultaneously, it encodes the equipment operation log into a text feature vector as a key and value. With an asymmetric attention architecture, it actively guides the semantic analysis of logs and the association between value state. Finally, it outputs a specific fault root cause category that can directly explain the value decline. This achieves automated reasoning from "value anomaly" to "explainable fault cause," overcoming the shortcomings of existing technologies that lack value state guidance and cannot establish a correlation between macro indicators and micro logs. This significantly improves the intelligence level and accuracy of fault root cause localization.

[0092] The specific process of the fault root cause localization method for equipment value degradation provided by the embodiments of the present invention will be described below with reference to the accompanying drawings.

[0093] Figure 5 This is a flowchart illustrating a method for locating the root cause of equipment value degradation, as shown in an embodiment of the present invention. Figure 5 As shown, it includes the following steps: 510: Obtain the Residual Value Index (RVI) and equipment operation logs of the target equipment. The Residual Value Index (RVI) is used to quantify the degree of economic value decay of the equipment.

[0094] 520: Obtain the state feature vector based on the residual value index (RVI), and obtain the text feature vector based on the device operation log.

[0095] 530: Using the state feature vector as the query vector and the text feature vector as the key and value vectors, the context feature vector is obtained through an attention mechanism.

[0096] 540: Based on the context feature vector, obtain the root cause category of the change in the residual value index (RVI) of the target equipment.

[0097] By encoding the Residual Value Index (RVI), which reflects the decline in the economic value of equipment, into a state feature vector and using it as the unique query vector for the attention mechanism, and simultaneously encoding the equipment operation log into a text feature vector as a key and value, an asymmetric attention architecture is used to proactively guide the semantic analysis of logs and the association between value state. The final output is a specific fault root cause category that can directly explain the value decline. This achieves automated reasoning from "value anomaly" to "explainable fault cause," overcoming the shortcomings of existing technologies that lack value state guidance and cannot establish a correlation between macro indicators and micro logs. This significantly improves the intelligence level and accuracy of fault root cause localization.

[0098] In some embodiments, step 530 includes: The attention mechanism is used to calculate the similarity between the state feature vector and each text feature vector, and the similarity is normalized to obtain the attention weight.

[0099] The context feature vector is obtained by weighting and summing the text feature vectors according to the attention weights.

[0100] By calculating the similarity between the value state features and each log feature and normalizing it to obtain the attention weight, and then performing a weighted summation of the log features, the log information with high attention weight is significantly enhanced, while the logs with low weight are suppressed. This generates a context feature vector that focuses on the explanation of value decay, providing high-quality feature input for subsequent root cause classification.

[0101] In some embodiments, step 530 involves calculating the similarity between the state feature vector and each text feature vector using an attention mechanism, and normalizing the similarity to obtain the attention weights, including: The original similarity score is obtained by calculating the dot product between the status feature vector and the text feature vector of each device operation log.

[0102] The original similarity score is scaled based on the dimension of the text feature vector to obtain the scaled score.

[0103] Based on the scaled score, the attention weight of each device operation log is determined.

[0104] In this embodiment of the invention, attention weights are calculated by scaling the dot product. On the one hand, the dot product efficiently measures the similarity between the state vector and the log vector. On the other hand, scaling by dividing by the square root of the dimension avoids the softmax gradient from becoming too large, thus ensuring the rationality of the attention weight distribution and the stability of training.

[0105] In some embodiments, step 530 includes: The attention weights are calculated using the following formula: Where Q is the state feature vector. For the first i The text feature vector corresponding to each device operation log. For feature dimension, For the first i Attention weights for each device operation log.

[0106] The calculation process of attention weight is defined by a clear mathematical formula, which makes the contribution of each log to the explanation of value decay quantifiable and traceable, providing a foundation for the extraction of supporting evidence and the interpretability of root cause localization.

[0107] In some embodiments, step 540 includes: The context feature vector is input into the classifier, which outputs the root cause category of the fault and its confidence level. The output category of the classifier includes at least one of the following: heat dissipation system failure, power system failure, network connection problem, storage medium damage, software configuration error, hardware aging, and external environmental interference.

[0108] By pre-setting root cause categories covering common equipment failures, the classifier automatically identifies the most likely failure type and confidence level from the context feature vector that incorporates value state guidance information. This achieves end-to-end mapping from multi-source heterogeneous data to specific failure root causes, avoiding the tedious process of manually checking logs and making experience-based judgments.

[0109] In some embodiments, step 520, obtaining the state feature vector based on the residual value index (RVI), includes: The residual value index RVI is mapped to a state feature vector of a preset dimension through a multilayer perceptron, wherein the multilayer perceptron includes at least one fully connected layer and a nonlinear activation function.

[0110] By mapping the single-value residual value index (RVI) to a high-dimensional state feature vector through a multilayer perceptron, the feature representation capability of the value state is enhanced, enabling it to perform effective dot product similarity calculation with the text feature vector. At the same time, the introduction of a non-linear activation function improves the model's ability to fit non-linear changes in value decay.

[0111] In some embodiments, step 520, which obtains the text feature vector based on the device operation log, includes: The device operation logs are input into a pre-trained language model to extract the semantic feature vector of each device operation log. The pre-trained language model is BERT.

[0112] By leveraging the powerful semantic understanding capabilities of pre-trained language models such as BERT, the original natural language logs are transformed into semantic feature vectors rich in contextual information, providing high-quality text representations for subsequent interactions with value state features and significantly improving the model's ability to extract hidden fault information from the logs.

[0113] Figure 6 This is a schematic diagram of a fault root cause localization device for equipment value degradation, as illustrated in an embodiment of the present invention. Figure 6 As shown, it includes: The first processing module is used to obtain the Residual Value Index (RVI) of the target equipment and the equipment operation log. The Residual Value Index (RVI) is used to quantify the degree of economic value decay of the equipment.

[0114] The second processing module is used to obtain the state feature vector based on the residual value index (RVI) and the text feature vector based on the device operation log.

[0115] The third processing module is used to take the state feature vector as the query vector and the text feature vector as the key vector and value vector, and obtain the context feature vector through an attention mechanism.

[0116] The fourth processing module is used to obtain the root cause category of the change in the residual value index (RVI) of the target device based on the context feature vector.

[0117] In some embodiments, the first processing module is specifically used to: obtain the residual value index (RVI) from the equipment management system, wherein the RVI is a continuous value between [0,1], which comprehensively reflects the residual value of the equipment, the predicted operation and maintenance cost, and the technological gap; and obtain the equipment operation log text from the equipment log library, wherein the log text includes at least one of operation status records, alarm information, and error reports.

[0118] In some embodiments, the second processing module is specifically used for: The residual value index (RVI) is input into a multilayer perceptron, which includes at least one fully connected layer and a nonlinear activation function, and mapped to a state feature vector of a preset dimension. The device operation logs are input into a pre-trained language model, which can be BERT, RoBERTa, or DistilBERT, and the semantic feature vectors of each log are extracted to form a set of text feature vectors.

[0119] In some embodiments, the third processing module is specifically used for: The state feature vector is used as the unique query vector in the attention mechanism, and the set of text feature vectors is used as the key and value in the attention mechanism. The dot product between the state feature vector and each text feature vector is calculated to obtain the original similarity score. The original similarity score is scaled according to the dimension of the text feature vector to obtain the scaled score. The scaled score is normalized by the softmax function to obtain the attention weight of each log. The text feature vectors are weighted and summed according to the attention weight to generate the context feature vector.

[0120] In some embodiments, the third processing module is specifically used to: calculate the attention weights using the following formula: Where Q is the state feature vector. For the first i The text feature vector corresponding to each device operation log. For feature dimension, For the first i Attention weights for each device operation log.

[0121] Context feature vector ,in Let the text feature vector corresponding to the i-th device operation log be used as the value vector.

[0122] In some embodiments, the fourth processing module is specifically used to: input the context feature vector into a fully connected classifier, the output layer of the classifier adopts the softmax activation function, and outputs the root cause category of the fault and its confidence level; the output category of the classifier includes at least one of the following: heat dissipation system failure, power system failure, network connection problem, storage medium damage, software configuration error, hardware aging, external environmental interference, and others / unknown.

[0123] In some embodiments, the apparatus further includes: a report generation module, configured to extract log entries with attention weights greater than a preset threshold as supporting evidence for the root cause category of the fault, and generate an analysis report containing the root cause category of the fault, confidence level, supporting evidence, and a visualization of attention weights.

[0124] Those skilled in the art will readily recognize that, based on the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein, the present invention can be implemented in hardware or a combination of hardware and computer software. Whether a function is implemented in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the present invention.

[0125] It should be noted that, Figure 6 The division of modules / units is illustrative and represents only one logical functional division; in actual implementation, other division methods are possible. For example, two or more functions can be integrated into a single data acquisition module. The integrated modules described above can be implemented either in hardware or as software functional modules.

[0126] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the aforementioned root cause localization methods for device value degradation. That is, an electronic device according to an embodiment of the present invention may include, but is not limited to: a processor and a memory; the memory is used to store the computer program; the processor is used to execute a root cause localization method for device value degradation as shown in any embodiment of the present invention by calling the computer program.

[0127] In one alternative embodiment, an electronic device is provided, such as Figure 7 As shown, Figure 7 The illustrated electronic device 700 includes a processor 701 and a memory 703. The processor 701 and the memory 703 are connected, for example, via a bus 702. Optionally, the electronic device 700 may further include a transceiver 704, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 704 is not limited to one type, and the structure of the electronic device 700 does not constitute a limitation on the embodiments of the present invention.

[0128] Among them, electronic devices can also be terminal devices, which can be any device that can install applications, including at least one of smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, smart TVs, and smart in-vehicle devices.

[0129] It should be noted that, Figure 7 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0130] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-mentioned methods for locating the root cause of equipment value degradation.

[0131] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0132] It should be noted that the terms "first," "second," etc., used in the specification and claims of this invention are used to distinguish similar objects and represent a limitation on a specific order or sequence. Where appropriate, the order of use for similar objects can be interchanged so that the embodiments of the invention described herein can be implemented in an order other than that shown or described.

[0133] Those skilled in the art will recognize that this invention can be implemented as a system, method, or computer program product. Therefore, this invention can be specifically implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, this invention can also be implemented as a computer program product contained in one or more computer-readable media, which includes computer-readable program code.

[0134] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for locating the root cause of equipment value degradation, characterized in that, include: Obtain the Residual Value Index (RVI) and equipment operation log of the target equipment. The Residual Value Index (RVI) is used to quantify the degree of economic value decay of the equipment. Based on the Residual Value Index (RVI), a state feature vector is obtained; based on the device operation log, a text feature vector is obtained. Using the state feature vector as the query vector and the text feature vector as the key and value vectors, an attention mechanism is used to obtain the context feature vector. Based on the context feature vector, the root cause category of the change in the residual value index (RVI) of the target device is obtained.

2. The method according to claim 1, characterized in that, The step of using the state feature vector as a query vector and the text feature vector as a key and value vector, and obtaining the context feature vector through an attention mechanism, includes: The similarity between the state feature vector and each of the text feature vectors is calculated using an attention mechanism, and the similarity is normalized to obtain the attention weights. The text feature vector is obtained by weighting and summing the text feature vectors according to the attention weights.

3. The method according to claim 2, characterized in that, The process involves calculating the similarity between the state feature vector and each of the text feature vectors using an attention mechanism, and then normalizing the similarity to obtain attention weights, including: Calculate the dot product between the state feature vector and the text feature vector of each of the device operation logs to obtain the original similarity score; The original similarity score is scaled according to the dimension of the text feature vector to obtain the scaled score; Based on the scaled score, the attention weight of each of the device operation logs is determined.

4. The method according to claim 3, characterized in that, The original similarity score is obtained by calculating the dot product between the state feature vector and the text feature vector of each of the device operation logs; the original similarity score is then scaled according to the dimension of the text feature vector to obtain a scaled score. Based on the scaled score, the attention weight of each of the device operation logs is determined, including: The attention weights are calculated using the following formula: Where Q is the state feature vector. For the first i The text feature vector corresponding to each device operation log. For feature dimension, For the first i Attention weights for each device operation log.

5. The method according to claim 1, characterized in that, The step of obtaining the fault root cause category of the residual value index (RVI) change of the target equipment based on the context feature vector includes: The context feature vector is input into the classifier, which outputs the root cause category of the fault and its confidence level. The output category of the classifier includes at least one of the following: heat dissipation system failure, power system failure, network connection problem, storage medium damage, software configuration error, hardware aging, and external environmental interference.

6. The method according to claim 1, characterized in that, The step of obtaining the state feature vector based on the residual value index (RVI) includes: The residual value index RVI is mapped to a state feature vector of a preset dimension through a multilayer perceptron, wherein the multilayer perceptron includes at least one fully connected layer and a nonlinear activation function.

7. The method according to claim 1, characterized in that, include: The device operation logs are input into a pre-trained language model to extract the semantic feature vector of each device operation log. The pre-trained language model is BERT.

8. A fault root cause localization device for equipment value degradation, characterized in that, include: The first processing module is used to obtain the Residual Value Index (RVI) of the target equipment and the equipment operation log. The Residual Value Index (RVI) is used to quantify the degree of economic value decay of the equipment. The second processing module is used to obtain a state feature vector based on the residual value index (RVI) and a text feature vector based on the device operation log. The third processing module is used to take the state feature vector as a query vector and the text feature vector as a key vector and a value vector, and obtain the context feature vector through an attention mechanism. The fourth processing module is used to obtain the root cause category of the change in the residual value index (RVI) of the target device based on the context feature vector.

9. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the fault root cause localization method for device value degradation as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one computer program, which is loaded and executed by a processor to enable the computer to implement the fault root cause localization method for equipment value degradation as described in any one of claims 1 to 7.

Images