Intelligent monitoring method, system and medium for special equipment damage based on working condition data

Abstract

The present invention relates to the field of monitoring data fusion technology, specifically to a method, system, and medium for intelligent monitoring of special equipment damage based on operating condition data. The method employs an isolation forest algorithm for high-sensitivity anomaly pre-screening; applies a patch embedding mechanism to generate time series feature vectors; retrieves related mechanism vectors to generate mixed sequences; generates state feature vectors and attribution logs through multiple rounds of self-attention calculation and information fusion; calculates damage probability prediction distribution vectors; applies a natural language processing model to quantify the severity of failure consequences of each damage mode; conducts risk assessment, obtains risk levels, and performs graded processing; obtains difference reports, and performs feedback optimization. This method enables a deep fusion of physical mechanism knowledge and real-time operating condition data in special equipment monitoring methods, enabling reliable diagnosis of the root cause of damage.

Description

Technical Field

[0001] The present invention relates to the technical field of monitoring data fusion, and in particular to a method, system and medium for intelligently monitoring damage of special equipment based on working condition data. Background Art

[0002] Special equipment often operates under complex conditions, often facing potential threats from a variety of damage mechanisms. However, the formation and development of special equipment damage is often hidden and difficult for personnel to detect. Therefore, the market urgently needs a special equipment damage monitoring method.

[0003] Currently, the most advanced equipment damage diagnosis methods primarily utilize data-driven analytical techniques, using pattern recognition from multiple sources to perform damage classification. A key feature of this method is its ability to deeply process and understand unstructured textual information. Using text analysis techniques, it extracts key features related to damage mechanisms from textual content such as technical specifications and historical maintenance records. While processing textual information, the method also analyzes a series of quantifiable operating parameters, such as pressure and temperature, acquired from the equipment in real time, converting these parameters into feature data for classification. The final step of the method is a comprehensive decision-making process. The mechanism features extracted from the text are combined with the state features extracted from the operating parameters. Through a comprehensive classification calculation, a precise judgment regarding the specific damage pattern is ultimately output.

[0004] However, this approach has exposed serious limitations in practice. Existing diagnostic techniques often provide a diagnosis without understanding the underlying causes. While they can provide a damage classification, the lack of a clear physical and logical explanation renders the diagnostic process opaque and its credibility difficult to assess. Furthermore, the long-term monitoring performance of existing methods degrades over time, and their ability to capture weak damage signals gradually decreases. Existing special equipment monitoring methods lack a deep fusion of knowledge about physical mechanisms and real-time operating data, making it difficult to reliably diagnose the root cause of damage.

[0005] To this end, an intelligent monitoring method, system and medium for special equipment damage based on working condition data are proposed. Summary of the Invention

[0006] The purpose of the present invention is to provide a method, system and medium for intelligent monitoring of special equipment damage based on working condition data. The present invention relates to the field of monitoring data fusion technology, and the specific steps are: using the isolation forest algorithm to perform high-sensitivity anomaly pre-screening; applying the patch embedding mechanism to generate a time series feature vector; retrieving the relevant mechanism vector to generate a mixed sequence; generating a state feature vector and an attribution log through multiple rounds of self-attention calculation and information fusion; calculating the damage probability prediction distribution vector; applying the natural language processing model to quantify the severity of the failure consequences of each damage mode; performing risk assessment, obtaining the risk level and performing graded processing; obtaining a difference report and performing feedback optimization. The existing special equipment monitoring method does not achieve deep fusion between physical mechanism knowledge and real-time working condition data, and it is difficult for the existing technology to make a reliable diagnosis of the root cause of the damage. To this end, a method, system and medium for intelligent monitoring of special equipment damage based on working condition data are proposed.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] The intelligent monitoring method for special equipment damage based on working condition data includes:

[0009] The isolation forest algorithm is used to pre-screen the real-time cableway operating condition data for highly sensitive anomalies and record suspected abnormal operating condition data. The patch embedding mechanism is applied to segment the suspected abnormal operating condition data into time series patches and output time series feature vectors.

[0010] Insert the CLS vector at the front end of the mixed sequence and send it to the multi-layer network; through multiple rounds of self-attention calculation and information fusion within the network, output the state feature vector, and save the attention weight of the middle layer of the network as the attribution log;

[0011] Inputting the state feature vector into a prediction head to obtain a damage probability prediction distribution vector;

[0012] A natural language processing model is applied to quantify the severity of the failure consequences of each damage mode; a risk assessment is performed based on the damage probability prediction distribution vector and the failure consequence severity to obtain a risk report, and the attribution log is used as a diagnosis basis for graded processing; the actual damage condition is obtained and compared with the damage probability prediction distribution vector, a difference report is output, and feedback optimization is performed.

[0013] Preferably, the highly sensitive abnormal pre-screening is:

[0014] An isolation forest algorithm is used to construct multiple isolation trees on historical normal operating condition data to form a pre-trained isolation forest model. Real-time cableway operating condition data is input into the pre-trained isolation forest model, and the average path length of the real-time cableway operating condition data in the multiple isolation trees is calculated, and the average path length is converted into an anomaly score. Based on the anomaly score distribution of the historical normal operating condition data, a judgment value lower than the conventional statistical threshold is set as a high-sensitivity judgment basis. The real-time cableway operating condition data and the data within the context time window whose anomaly scores exceed the high-sensitivity judgment basis are recorded as the suspected abnormal condition data.

[0015] Preferably, the high-sensitivity abnormality pre-screening process further comprises:

[0016] Each sensor of special equipment is modeled as a node of a dynamic information flow network. Based on the information theory method of transfer entropy, the structural health of information transfer within the system is quantified to obtain the system topology entropy of the entire network. By continuously monitoring the system topology entropy, structural cumulative risks are identified.

[0017] Preferably, the process of dividing the real-time cableway operating condition data into time series patches and projecting them into time series feature vectors is as follows:

[0018] A patch embedding mechanism is applied to segment the suspected abnormal operating condition data into overlapping data segments with a preset length along the time dimension to generate a time series patch. The time series patch is flattened into a one-dimensional original vector, and the one-dimensional original vector is linearly projected through a weight matrix to generate an initial time series feature vector. The position encoding vector is superimposed on the initial time series feature vector to output a time series feature vector.

[0019] Preferably, the process of outputting the state feature vector through multiple rounds of self-attention calculation and information fusion within the network is:

[0020] The CLS vector is spliced ​​to the starting position of the mixed sequence; and the spliced ​​sequence is sent to a multi-layer network, which is composed of multiple stacked Transformer encoder layers; in each layer of the multi-layer network, the attention weight between any two vectors in the current input sequence is calculated through a multi-head self-attention mechanism, and all vectors in the current input sequence are weighted summed according to the attention weight to generate an intermediate representation; at the same time, the multi-head self-attention mechanism enables the CLS vector to continuously aggregate information from all vectors in the mixed sequence; the intermediate representation and the current input sequence are then subjected to residual connection and layer normalization processing; the processed result is then sent to a feedforward neural network, and residual connection and layer normalization processing are performed again to generate an output sequence of the current layer, which is used as the input sequence of the next layer; after the spliced ​​sequence passes through all layers of the multi-layer network, a final output sequence is obtained; from the final output sequence, the final CLS output vector at the starting position is extracted and output as the state feature vector.

[0021] Preferably, the process of saving the attention weights of the middle layer of the network as attribution logs is:

[0022] Save the multi-head attention weight matrix calculated by at least one intermediate layer in the multi-layer network; parse the multi-head attention weight matrix to identify at least one related mechanism vector of the attention weight assigned by the CLS vector; trace back to the dynamic mechanism knowledge base based on the identity of the related mechanism vector to extract the corresponding original damage mechanism text description; record the suspected abnormal working condition data, the original damage mechanism text description and the corresponding attention weight value together as an attribution log.

[0023] Preferably, the generation of the attribution log further includes: considering each layer of the multi-layer network as a level of the graph, taking the input vector of the current layer as a node, and establishing directed weighted edges between nodes of adjacent levels based on the attention weights calculated by the self-attention mechanism; by recursively multiplying the attention matrices between layers and combining the influence of the residual connections of each layer, the cumulative attention contribution of the state feature vector from any initial input vector to the final output is calculated to form an attention flow that runs through the multi-layer network.

[0024] Preferably, the risk report is obtained by:

[0025] Calculate the occurrence probability of each damage mode in the damage probability prediction distribution vector and multiply it by the corresponding failure consequence severity value to obtain the risk score of each damage mode; arrange the risk scores in descending order to obtain a diagnosis list; transmit the diagnosis list to a preset risk level judgment model to obtain the risk level of the primary risk item; the diagnosis list and the risk level of the primary risk item constitute a risk report.

[0026] Preferably, the process of obtaining the risk report further includes: when potential damage is identified, starting a deep generative model combined with physical simulation constraints; the deep generative model uses the potential damage as a counterfactual intervention and generates multiple future scenario sequences through random sampling; integrating the multiple future scenario sequences to generate a set of possible future scenarios and make scenario-based robust decisions.

[0027] The intelligent monitoring system for special equipment damage based on working condition data includes:

[0028] The data processing module uses the isolation forest algorithm to pre-screen the real-time cableway operating condition data for highly sensitive anomalies and record suspected abnormal operating condition data. It also applies a patch embedding mechanism to segment the suspected abnormal operating condition data into time series patches and output a time series feature vector.

[0029] The data fusion module inserts the CLS vector at the front end of the mixed sequence and sends it to the multi-layer network. Through multiple rounds of self-attention calculation and information fusion within the network, it outputs the state feature vector and saves the attention weights of the middle layer of the network as attribution logs.

[0030] A probability prediction module inputs the state feature vector into a prediction head to obtain a damage probability prediction distribution vector;

[0031] The hierarchical decision-making module applies a natural language processing model to quantify the severity of the failure consequences of each damage mode; performs a risk assessment based on the damage probability prediction distribution vector and the failure consequence severity to obtain a risk level, and performs hierarchical processing in combination with the attribution log as a diagnosis basis; obtains the actual damage status, compares it with the damage probability prediction distribution vector, outputs a difference report, and performs feedback optimization.

[0032] A medium stores computer program instructions, wherein the computer program is executed by a processor to perform steps of a method for intelligently monitoring damage of special equipment based on working condition data.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. Segmenting suspected abnormal operating condition data into overlapping time series patches can not only accurately capture the dynamic evolution characteristics of damage in the short term, but also ensure the continuity of time series information and avoid the loss of key signals at segment boundaries; linearly projecting the patches and superimposing position codes can convert unstructured time series data into structured feature vectors that are easy for the model to process, while fully preserving the contextual time series relationship of the signal; this processing method provides high-quality, high-information-density input for subsequent knowledge fusion and in-depth analysis, thus laying a solid data foundation for accurate damage diagnosis.

[0035] 2. By splicing the CLS vector as an information aggregator at the front end of the mixed sequence and processing it with a multi-layer Transformer encoder, the model can be effectively driven to conduct in-depth and global interaction and fusion of features from real-time working conditions and mechanism features from the knowledge base; its internal multi-head self-attention mechanism enables the CLS vector to dynamically and selectively aggregate the most important information in the entire sequence, while the residual connection and layer normalization ensure the stable and efficient transmission and refinement of information in the deep network; the final generated single state feature vector comprehensively summarizes the current state of the equipment and its potential physical causes, greatly improving the accuracy and reliability of subsequent diagnosis.

[0036] 3. By saving and parsing the multi-head attention weight matrix of the middle layer of the network, it is possible to trace the formation path of the diagnostic conclusion and identify which physical mechanism vectors the model focused on when making decisions; further tracing back to the knowledge base to extract the original text description of the damage mechanism, the abstract attention weight is associated with specific, readable physical knowledge, and an attribution log containing data, mechanism and weight values ​​is generated; this method transforms the deep learning model from a "black box" to an interpretable "white box", providing a clear "chain of evidence" for the diagnostic results, solving the trust problem of traditional methods, and significantly enhancing the credibility and auditability of the diagnostic conclusions. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A flow chart of a method for intelligently monitoring damage of special equipment based on operating condition data provided by an embodiment of the present invention;

[0038] Figure 2 A schematic structural diagram of a special equipment damage intelligent monitoring system based on working condition data provided by an embodiment of the present invention;

[0039] Figure 3 A schematic diagram of a high-sensitivity anomaly pre-screening process provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0041] Example 1

[0042] In order to better carry out intelligent damage monitoring of the reciprocating passenger ropeway in a tourist attraction, an intelligent damage monitoring method for special equipment based on working condition data is proposed. The overall process is as follows: Figure 1 As shown:

[0043] The damage of a reciprocating passenger ropeway in a tourist attraction was monitored intelligently. The operating data processed included the drive wheel speed, motor current, tensioning cylinder pressure, and load-bearing cable vibration frequency. The details are as follows:

[0044] The isolation forest algorithm is used to pre-screen real-time cableway operating data for highly sensitive anomalies and record suspected abnormal operating condition data. A patch embedding mechanism is applied to segment the suspected abnormal operating condition data into time series patches, and a time series feature vector is output. Based on the time series feature vector, relevant mechanism vectors are retrieved from the dynamic mechanism knowledge base and concatenated with the time series feature vector to generate a hybrid sequence.

[0045] The CLS vector is inserted at the front end of the mixed sequence and sent to the multi-layer network; through multiple rounds of self-attention calculation and information fusion within the network, the state feature vector is output, and the attention weight of the middle layer of the network is saved as the attribution log.

[0046] Inputting the state feature vector into a prediction head, calculating and outputting a damage probability prediction distribution vector;

[0047] A natural language processing model is applied to quantify the severity of the failure consequences of each damage mode; a risk assessment is performed based on the damage probability prediction distribution vector and the failure consequence severity to obtain a risk report, and the attribution log is used as a diagnosis basis for graded processing; the actual damage condition is obtained and compared with the damage probability prediction distribution vector, a difference report is output, and feedback optimization is performed.

[0048] Furthermore, the high sensitivity abnormal pre-screening, such as Figure 2 As shown:

[0049] An isolation forest algorithm is used to construct multiple isolation trees on historical normal operating condition data to form a pre-trained isolation forest model. Real-time cableway operating condition data is input into the pre-trained isolation forest model, and the average path length of the real-time cableway operating condition data in the multiple isolation trees is calculated, and the average path length is converted into an anomaly score. Based on the anomaly score distribution of the historical normal operating condition data, a judgment value lower than the conventional statistical threshold is set as a high-sensitivity judgment basis. The real-time cableway operating condition data and the data within the context time window whose anomaly scores exceed the high-sensitivity judgment basis are recorded as the suspected abnormal condition data.

[0050] In this monitoring event, the system received an operating condition data point. The pre-trained isolation forest model calculated an anomaly score of 0.68 for this data point, exceeding the preset high sensitivity threshold of 0.65. Therefore, the system recorded the 20-second window containing this data point as suspected abnormal operating condition data, characterized by small but frequent, non-periodic fluctuations in the tensioning cylinder pressure.

[0051] Setting a judgment value lower than the conventional statistical threshold as the basis for high-sensitivity judgment can not only filter massive operating condition data, reduce unnecessary data processing steps, and improve data processing efficiency, but also effectively capture subtle abnormal signals that are difficult to detect, avoid misjudgment, and achieve early warning of faults.

[0052] The entire pre-screening and even diagnostic logic is based on the analysis of "signals," assuming that damage will always appear in the form of observable signals. This methodology cannot handle the accumulation of "structural risks" that appear "quiet" at the signal level due to the reorganization of the internal coupling relationships of complex systems. To overcome this problem, the high-sensitivity anomaly pre-screening process is further improved as follows: each sensor of special equipment is modeled as a node in a dynamic information flow network; based on the information theory method of transfer entropy, the "structural health" of information transfer within the system is quantified to obtain the system topological entropy of the entire network; by continuously monitoring the system topological entropy, structural cumulative risks are identified; the above-mentioned method elevates the monitoring dimension from the physical signal level to the system organizational structure level, enabling the system to detect structural functional degradations that are completely invisible in physical signal monitoring.

[0053] Furthermore, the process of dividing the real-time cableway operating condition data into time series patches and projecting them into time series feature vectors is as follows:

[0054] A patch embedding mechanism is applied to segment the suspected abnormal working condition data into overlapping data segments with preset length along the time dimension to generate time series patches; the time series patches are flattened into one-dimensional original vectors, and the one-dimensional original vectors are linearly projected through a weight matrix to generate an initial time series feature vector; a position encoding vector is superimposed on the initial time series feature vector to output a time series feature vector.

[0055] During this monitoring event, the 20 seconds of recorded four-dimensional operating data were segmented into multiple 5-second time series patches with 2-second overlaps. Each patch was flattened, linearly projected, and position-encoded, resulting in a 768-dimensional time series feature vector. This vector mathematically represents the dynamic pattern of small, non-periodic pressure fluctuations.

[0056] By segmenting suspected anomaly data into overlapping time-series patches, we can not only precisely capture the dynamic evolution of damage over a short period of time, but also ensure the continuity of temporal information and avoid the loss of key signals at segment boundaries. The addition of position encoding preserves complete contextual information for subsequent analysis.

[0057] Furthermore, the construction process of the dynamic mechanism knowledge base is as follows:

[0058] First, a pre-trained large-scale natural language processing model is applied to perform deep semantic encoding on each text unit. The core information contained in it, including the causal logic of the damage, key influencing factors, and specific failure modes, is mapped and converted into a high-dimensional "knowledge vector" that can accurately capture its internal logic. Subsequently, the original text, the newly generated knowledge vector, and metadata containing information such as identity and damage type are systematically associated and uniformly stored in the database. At the same time, an efficient vector search algorithm is used to establish a similarity retrieval engine for all knowledge vectors. Finally, to ensure the "dynamic" nature of the knowledge base, a periodic automated process continuously scans the knowledge source and automatically repeats the above vectorization and indexing process for any new or revised content, thereby achieving incremental and continuous updating of the knowledge base content, keeping it always up to date and most comprehensive.

[0059] In this monitoring event, the system searches the knowledge base based on the time series feature vectors generated above. It retrieves the three mechanism vectors with the highest similarity. These three mechanism vectors constitute the related set.

[0060] Furthermore, the process of outputting the state feature vector through multiple rounds of self-attention calculation and information fusion within the network is:

[0061] The CLS vector is spliced ​​to the starting position of the mixed sequence; and the spliced ​​sequence is sent to a multi-layer network, which is composed of multiple stacked Transformer encoder layers; in each layer of the multi-layer network, the attention weight between any two vectors in the current input sequence is calculated through a multi-head self-attention mechanism, and all vectors in the current input sequence are weighted summed according to the attention weight to generate an intermediate representation; at the same time, the multi-head self-attention mechanism enables the CLS vector to continuously aggregate information from all other vectors in the mixed sequence; the intermediate representation and the current input sequence are then subjected to residual connection and layer normalization processing; the processed result is then sent to a feedforward neural network, and residual connection and layer normalization processing are performed again to generate the output sequence of this layer, which is used as the input sequence of the next layer; after the spliced ​​sequence passes through all layers of the multi-layer network, a final output sequence is obtained; from the final output sequence, the final CLS output vector at the starting position is extracted and output as the state feature vector.

[0062] In this monitoring event, a mixed sequence consisting of a time series feature vector and the three mechanism vectors was prepended with a CLS vector and fed into a 12-layer Transformer encoder network. After the complete computational process, the CLS output vector at the starting position was extracted as the state feature vector representing the current device state.

[0063] The combination of multi-head self-attention, feedforward networks, residual connections, and layer normalization ensures stable and deep information fusion and refinement within multi-layer networks. The application of CLS vectors aggregates the scattered information in the entire mixed sequence into a single state feature vector, providing high-quality input for subsequent accurate diagnosis.

[0064] Furthermore, the process of saving the attention weights of the middle layer of the network as attribution logs is as follows:

[0065] Save the multi-head attention weight matrix calculated by at least one intermediate layer in the multi-layer network; parse the multi-head attention weight matrix to identify at least one related mechanism vector of the attention weight assigned by the CLS vector; trace back to the dynamic mechanism knowledge base based on the identity of the related mechanism vector to extract the corresponding original damage mechanism text description; record the suspected abnormal working condition data, the corresponding original damage mechanism text description and the corresponding attention weight value together as an attribution log.

[0066] Attribution logs provide intuitive explanations for diagnostic conclusions, resolving the "black box" problem of traditional deep learning models. By tracing the attention weights of the CLS vector back to the original mechanism text, it clearly reveals the specific physical mechanism knowledge the model relied on when making its judgment, greatly enhancing the credibility of the diagnostic results and user trust in the model.

[0067] Attribution logs are more like static "evidence snapshots" that tell us what the model saw, but fail to fully depict how it linked these clues step by step to form the "chain of reasoning" for the final diagnosis. To overcome this limitation, the attribution log generation process was further improved as follows: each layer of the multi-layer network is regarded as a level of the graph, the input vector of the current layer is used as a node, and directed weighted edges are established between nodes of adjacent layers based on the attention weights calculated by the self-attention mechanism; by recursively multiplying the attention matrices between layers and combining the influence of the residual connections of each layer, the cumulative attention contribution from any initial input vector to the final output state feature vector is calculated, thus forming an "attention flow" that runs through the multi-layer network. The above solution allows operation and maintenance personnel to not only see the basis for the diagnosis, but also understand the complete thinking process behind it, thereby enhancing the trust and auditability of the AI ​​diagnostic conclusions.

[0068] Furthermore, the process of obtaining the risk report is as follows:

[0069] Calculate the occurrence probability of each damage mode in the damage probability prediction distribution vector and multiply it by the corresponding failure consequence severity value to obtain the risk score of each damage mode; arrange the risk scores in descending order to obtain a diagnosis list; transmit the diagnosis list to a preset risk level judgment model to obtain the risk level of the primary risk item; the diagnosis list and the risk level of the primary risk item constitute a risk report.

[0070] In this monitoring event, the state feature vector was fed into the prediction head, which output a predicted damage probability distribution vector. Combined with the preset failure consequence severity, a risk score was calculated. The risk level assessment model determined this score as "medium risk," generating a risk report containing a diagnostic list and risk level.

[0071] By multiplying the probability of occurrence by the severity of the consequences of failure, we can effectively identify and prioritize potential damage that, while not the most likely, may have extremely serious consequences, thus avoiding major safety incidents. The resulting risk report provides a clear basis for decision-making for operations and maintenance personnel.

[0072] The aforementioned failure consequence severity is a static value unrelated to the operating scenario. It cannot dynamically reflect the distinct hazard levels of the same damage under different operating conditions (e.g., during peak load and no-load testing). This results in a lack of scenario-specific adaptability in risk assessment results, potentially misjudging the true urgency of the risk. Therefore, the process for generating risk reports is further improved as follows: when a potential damage is identified, a deep generative model incorporating physical simulation constraints is activated; this deep generative model uses the potential damage as a counterfactual intervention and generates multiple future scenario sequences through random sampling; these multiple future scenario sequences are integrated to generate a set of "possible future scenarios" for robust scenario-based decision-making. This approach completely transforms risk assessment from a deterministic "prediction task" to a "narrative task" that explores uncertainty.

[0073] Furthermore, the specific process of performing the classification process is as follows:

[0074] Integrate risk reports and attribution logs into intelligent decision-making recommendations. Based on the risk level in the report, the automated grading process is automatically triggered, as follows:

[0075] When low risk is identified, enhanced monitoring and logging operations are performed without manual intervention;

[0076] When a medium risk is identified, a warning is sent to the staff member, including a suggested action. The action is then executed after receiving the staff member's approval.

[0077] When a high risk is identified, an emergency alarm is issued, and auxiliary operations can only continue after staff issue key instructions (such as "execute shutdown").

[0078] Throughout the entire process, all operations, whether automatically executed or manually triggered, will have their complete decision-making basis and execution records archived, providing a traceability basis for subsequent closed-loop feedback optimization.

[0079] Furthermore, the feedback optimization process is as follows:

[0080] After the inspection is completed, the actual damage condition is manually submitted. The actual damage condition is converted into a standard "ground truth label". Next, the damage probability prediction distribution vector output by the model is quantitatively compared with the true label, and a detailed difference report is automatically generated by calculating indicators such as cross entropy loss and accuracy. This report not only shows the gap between the prediction and reality, but also makes a preliminary attribution of the cause of the difference. Finally, based on this difference report, a full-link feedback optimization process is systematically triggered: according to the attribution conclusion, the case data is used to fine-tune the diagnosis model, optimize the retrieval model, adjust the sensitivity of the pre-screening model, and submit a request to the expert to revise the knowledge base, so as to ensure that each link of the entire method can continue to learn and improve itself in practice.

[0081] Example 2

[0082] In order to better monitor the damage of a certain ski resort's passenger aerial ropeway, a special equipment damage intelligent monitoring system based on working condition data is proposed. The system structure is as follows: Figure 3 As shown:

[0083] In a specific monitoring event of this embodiment, this method performs real-time monitoring on the cableway, and the processed working condition data include vibration acceleration of key towers, cableway running speed, wind speed, and motor current.

[0084] The data processing module uses the isolation forest algorithm to pre-screen the real-time cableway operating condition data for highly sensitive anomalies and record suspected abnormal operating condition data. It also applies a patch embedding mechanism to segment the suspected abnormal operating condition data into time series patches and project them into time series feature vectors. Based on the time series feature vectors, it retrieves relevant mechanism vectors and performs vector splicing to generate a hybrid sequence.

[0085] The data fusion module inserts the CLS vector at the front end of the mixed sequence and sends it to the multi-layer network. Through multiple rounds of self-attention calculation and information fusion within the network, it outputs the state feature vector and saves the attention weight of the middle layer of the network as the attribution log.

[0086] A probability prediction module inputs the state feature vector into a prediction head, calculates and outputs a damage probability prediction distribution vector;

[0087] The hierarchical decision-making module applies a natural language processing model to quantify the severity of the failure consequences of each damage mode; performs a risk assessment based on the damage probability prediction distribution vector and the failure consequence severity to obtain a risk level, and performs hierarchical processing in combination with the attribution log as a diagnosis basis; obtains the actual damage status, compares it with the damage probability prediction distribution vector, outputs a difference report, and performs feedback optimization.

[0088] Furthermore, the highly sensitive abnormal pre-screening is:

[0089] An isolation forest algorithm is used to construct multiple isolation trees on historical normal operating condition data to form a pre-trained isolation forest model. Real-time cableway operating condition data is input into the pre-trained isolation forest model, and the average path length of the real-time cableway operating condition data in the multiple isolation trees is calculated, and the average path length is converted into an anomaly score. Based on the anomaly score distribution of the historical normal operating condition data, a judgment value lower than the conventional statistical threshold is set as a high-sensitivity judgment basis. The real-time cableway operating condition data and the data within the context time window whose anomaly scores exceed the high-sensitivity judgment basis are recorded as the suspected abnormal condition data.

[0090] Furthermore, the process of dividing the real-time cableway operating condition data into time series patches and projecting them into time series feature vectors is as follows:

[0091] A patch embedding mechanism is applied to segment the suspected abnormal operating condition data into overlapping data segments with a preset length along the time dimension to generate a time series patch. The time series patch is flattened into a one-dimensional original vector, and the one-dimensional original vector is linearly projected through a weight matrix to generate an initial time series feature vector. The position encoding vector is superimposed on the initial time series feature vector to output a time series feature vector.

[0092] Furthermore, the construction process of the dynamic mechanism knowledge base is as follows:

[0093] First, a pre-trained large-scale natural language processing model is applied to perform deep semantic encoding on each text unit, mapping and converting the core information contained therein, including the causal logic of the damage, key influencing factors, and specific failure modes, into a high-dimensional "knowledge vector" that accurately captures its internal logic. Subsequently, the original text, the newly generated knowledge vector, and metadata containing information such as identity and damage type are systematically associated and uniformly stored in the database. At the same time, a similarity retrieval engine is established for all knowledge vectors using an efficient vector search algorithm. Finally, to ensure the "dynamic" nature of the knowledge base, a periodic automated process continuously scans the knowledge source and automatically repeats the above vectorization and indexing process for any new or revised content, thereby achieving incremental and continuous updating of the knowledge base content, keeping it always up to date and most comprehensive.

[0094] In the knowledge retrieval step of the present method, a retrieval is performed based on the time series feature vectors generated above, and three mechanism vectors with the highest similarity are retrieved, and these three mechanism vectors constitute a related set.

[0095] Furthermore, the process of outputting the state feature vector through multiple rounds of self-attention calculation and information fusion within the network is:

[0096] The CLS vector is spliced ​​to the starting position of the mixed sequence; and the spliced ​​sequence is sent to a multi-layer network, which is composed of multiple stacked Transformer encoder layers; in each layer of the multi-layer network, the attention weight between any two vectors in the current input sequence is calculated through a multi-head self-attention mechanism, and all vectors in the current input sequence are weighted summed according to the attention weight to generate an intermediate representation; at the same time, the multi-head self-attention mechanism enables the CLS vector to continuously aggregate information from all other vectors in the mixed sequence; the intermediate representation and the current input sequence are then subjected to residual connection and layer normalization processing; the processed result is then sent to a feedforward neural network, and residual connection and layer normalization processing are performed again to generate the output sequence of this layer, which is used as the input sequence of the next layer; after the spliced ​​sequence passes through all layers of the multi-layer network, a final output sequence is obtained; from the final output sequence, the final CLS output vector at the starting position is extracted and output as the state feature vector.

[0097] In this method, a mixed sequence consisting of a time series feature vector and the three mechanism vectors is prepended with a CLS vector and fed into a 12-layer Transformer encoder network. After the complete computational process, the CLS output vector at the starting position is extracted as the state feature vector representing the current device state.

[0098] Furthermore, the process of saving the attention weights of the middle layer of the network as attribution logs is as follows:

[0099] Save the multi-head attention weight matrix calculated by at least one intermediate layer in the multi-layer network; parse the multi-head attention weight matrix to identify at least one related mechanism vector of the attention weight assigned by the CLS vector; trace back to the dynamic mechanism knowledge base based on the identity of the related mechanism vector to extract the corresponding original damage mechanism text description; record the suspected abnormal working condition data, the corresponding original damage mechanism text description and the corresponding attention weight value together as an attribution log.

[0100] During the execution of this method, the attention weight matrix of the 11th layer was saved and parsed, and it was found that the CLS vector allocated 85% of the attention weight to the mechanism vector corresponding to knowledge 1.

[0101] Furthermore, the process of obtaining the risk report is as follows:

[0102] Calculate the occurrence probability of each damage mode in the damage probability prediction distribution vector and multiply it by the corresponding failure consequence severity value to obtain the risk score of each damage mode; arrange the risk scores in descending order to obtain a diagnosis list; transmit the diagnosis list to a preset risk level judgment model to obtain the risk level of the primary risk item; the diagnosis list and the risk level of the primary risk item constitute a risk report.

[0103] During the implementation of this method, the state feature vector is fed into the prediction head, which outputs a predicted damage probability distribution vector. Combined with the preset failure consequence severity, a risk score is calculated. This risk score is determined as "medium-high risk" by the risk level judgment model, and a risk report is generated, including a diagnostic list and risk level.

[0104] Furthermore, the specific process of performing the classification process is as follows:

[0105] First, the generated risk report and attribution log are integrated into a comprehensive intelligent decision-making proposal. Subsequently, a hierarchical processing process with varying degrees of automation is automatically triggered based on the risk level in the report: for low-risk situations, early warning operations such as enhanced monitoring and logging are autonomously executed without human intervention; for medium-risk situations, a warning with clear disposal recommendations is generated and executed after staff approval; for high-risk situations, an emergency alert is issued, forcing the system to wait for authorized personnel to issue key instructions (such as "execute shutdown") before continuing with auxiliary operations. Throughout the entire process, all operations, whether executed automatically or manually, are archived with their complete decision-making basis and execution records, providing a traceable basis for subsequent closed-loop feedback optimization.

[0106] Furthermore, the feedback optimization process is as follows:

[0107] After the inspection is completed, the actual damage condition is manually submitted. The actual damage condition is converted into a standard "ground truth label". Next, the damage probability prediction distribution vector output by the model is quantitatively compared with the true label, and a detailed difference report is automatically generated by calculating indicators such as cross entropy loss and accuracy. This report not only shows the gap between the prediction and reality, but also makes a preliminary attribution of the cause of the difference. Finally, based on this difference report, a full-link feedback optimization process is systematically triggered: according to the attribution conclusion, the case data is used to fine-tune the diagnosis model, optimize the retrieval model, adjust the sensitivity of the pre-screening model, and submit a request to the expert to revise the knowledge base, so as to ensure that each link of the entire method can continue to learn and improve itself in practice.

[0108] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent monitoring method for special equipment damage based on working condition data, characterized in that: include: Use the isolation forest algorithm to pre-screen the real-time cableway operating data for highly sensitive anomalies and record suspected abnormal operating data; Applying a patch embedding mechanism, the suspected abnormal operating condition data is segmented into overlapping data segments of a preset length along the time dimension to generate a time series patch. The time series patch is flattened into a one-dimensional original vector, and the one-dimensional original vector is linearly projected through a weight matrix to generate an initial time series feature vector. The position encoding vector is superimposed on the initial time series feature vector to output a time series feature vector. Insert the CLS vector at the front end of the mixed sequence and send it to the multi-layer network; in each layer of the multi-layer network, through multiple rounds of self-attention calculation and information fusion within the network, calculate the attention weight between any two vectors in the current input sequence, and perform weighted summation on all vectors in the current input sequence according to the attention weight to generate an intermediate representation; at the same time, the multi-head self-attention mechanism enables the CLS vector to continuously aggregate information from all vectors in the mixed sequence; then perform residual connection and layer normalization processing on the intermediate representation and the current input sequence; then send the processed result to the feedforward neural network, and perform residual connection and layer normalization processing again to generate the output sequence of the current layer, and use it as the next layer An input sequence is obtained after the spliced ​​sequence passes through all layers of the multi-layer network; a final output sequence is obtained; from the final output sequence, the final CLS output vector at the starting position is extracted as the state feature vector, the state feature vector is output, and a multi-head attention weight matrix calculated by at least one intermediate layer in the multi-layer network is saved; the multi-head attention weight matrix is ​​parsed to identify at least one related mechanism vector of the attention weight assigned by the CLS vector; according to the identity identifier of the related mechanism vector, the dynamic mechanism knowledge base is traced back to extract the corresponding original damage mechanism text description; and the suspected abnormal working condition data, the original damage mechanism text description and the corresponding attention weight of the network intermediate layer are saved as an attribution log; Inputting the state feature vector into a prediction head to obtain a damage probability prediction distribution vector; Applying a natural language processing model to quantify the severity of the failure consequences of each damage mode; performing a risk assessment based on the damage probability prediction distribution vector and the failure consequence severity to obtain a risk report, and performing a graded processing based on the attribution log as a diagnosis basis; The actual damage condition is obtained, compared with the damage probability prediction distribution vector, a difference report is output, and feedback optimization is performed.

2. The special equipment damage intelligent monitoring method based on working condition data according to claim 1 is characterized in that: The highly sensitive abnormal pre-screening is: An isolation forest algorithm is used to construct multiple isolation trees on historical normal operating condition data to form a pre-trained isolation forest model. Real-time cableway operating condition data is input into the pre-trained isolation forest model, and the average path length of the real-time cableway operating condition data in the multiple isolation trees is calculated, and the average path length is converted into an anomaly score. Based on the abnormal score distribution of the historical normal operating condition data, a judgment value lower than the conventional statistical threshold is set as a high-sensitivity judgment basis; the real-time cableway operating condition data and the data within the context time window whose abnormal scores exceed the high-sensitivity judgment basis are recorded as the suspected abnormal operating condition data.

3. The special equipment damage intelligent monitoring method based on working condition data according to claim 2 is characterized in that: The process of high-sensitivity abnormal pre-screening further includes: Each sensor of special equipment is modeled as a node of a dynamic information flow network. Based on the information theory method of transfer entropy, the structural health of information transfer within the system is quantified to obtain the system topology entropy of the entire network. By continuously monitoring the system topology entropy, structural cumulative risks are identified.

4. The special equipment damage intelligent monitoring method based on working condition data according to claim 1 is characterized in that: The generation of the attribution log further includes: considering each layer of the multi-layer network as a level of the graph, taking the input vector of the current layer as a node, and establishing directed weighted edges between nodes of adjacent levels based on the attention weights calculated by the self-attention mechanism; by recursively multiplying the attention matrices between each layer and combining the influence of the residual connections of each layer, the cumulative attention contribution of the state feature vector from any initial input vector to the final output is calculated, forming an attention flow that runs through the multi-layer network.

5. The special equipment damage intelligent monitoring method based on working condition data according to claim 1 is characterized in that: The process of obtaining the risk report is as follows: Calculate the occurrence probability of each damage mode in the damage probability prediction distribution vector and multiply it by the corresponding failure consequence severity value to obtain the risk score of each damage mode; arrange the risk scores in descending order to obtain a diagnosis list; transmit the diagnosis list to a preset risk level judgment model to obtain the risk level of the primary risk item; the diagnosis list and the risk level of the primary risk item constitute a risk report.

6. The special equipment damage intelligent monitoring method based on working condition data according to claim 4 is characterized in that: The risk report acquisition process further includes: when potential damage is identified, starting a deep generative model combined with physical simulation constraints; the deep generative model uses the potential damage as a counterfactual intervention and generates multiple future scenario sequences through random sampling; integrating the multiple future scenario sequences to generate a set of possible future scenarios and make scenario-based robust decisions.

7. The intelligent monitoring system for special equipment damage based on working condition data is characterized by: include: The data processing module uses the isolation forest algorithm to pre-screen the real-time cableway operating data for highly sensitive anomalies and record suspected abnormal operating data; Applying a patch embedding mechanism, the suspected abnormal operating condition data is segmented into overlapping data segments of a preset length along the time dimension to generate a time series patch. The time series patch is flattened into a one-dimensional original vector, and the one-dimensional original vector is linearly projected through a weight matrix to generate an initial time series feature vector. The position encoding vector is superimposed on the initial time series feature vector to output a time series feature vector. The data fusion module inserts the CLS vector at the front end of the mixed sequence and sends it to the multi-layer network; in each layer of the multi-layer network, the attention weight between any two vectors in the current input sequence is calculated through multiple rounds of self-attention calculation and information fusion within the network, and all vectors in the current input sequence are weighted summed according to the attention weight to generate an intermediate representation; at the same time, the multi-head self-attention mechanism enables the CLS vector to continuously aggregate information from all vectors in the mixed sequence; the intermediate representation is then subjected to residual connection and layer normalization processing with the current input sequence; the processed result is then sent to the feedforward neural network, and residual connection and layer normalization processing are performed again to generate the output sequence of the current layer, and is the input sequence of the next layer; after the spliced ​​sequence passes through all layers of the multi-layer network, a final output sequence is obtained; from the final output sequence, the final CLS output vector at the starting position is extracted as the state feature vector, the state feature vector is output, and the multi-head attention weight matrix calculated by at least one intermediate layer in the multi-layer network is saved; the multi-head attention weight matrix is ​​parsed to identify at least one related mechanism vector of the attention weight assigned by the CLS vector; according to the identity of the related mechanism vector, the dynamic mechanism knowledge base is traced back to extract the corresponding original damage mechanism text description; and the suspected abnormal working condition data, the original damage mechanism text description and the corresponding attention weight of the network intermediate layer are saved as the attribution log; A probability prediction module inputs the state feature vector into a prediction head to obtain a damage probability prediction distribution vector; A hierarchical decision-making module applies a natural language processing model to quantify the severity of the failure consequences of each damage mode; performs a risk assessment based on the damage probability prediction distribution vector and the failure consequence severity to obtain a risk level, and performs hierarchical processing based on the attribution log as a diagnosis basis; The actual damage condition is obtained, compared with the damage probability prediction distribution vector, a difference report is output, and feedback optimization is performed.

8. A medium having computer program instructions stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for intelligently monitoring damage of special equipment based on operating condition data as described in any one of claims 1 to 6 are implemented.

Images