A device label and multi-source data fusion health assessment method and system

Abstract

The application discloses a kind of equipment label and multi-source data fusion health assessment method and system, belong to industrial equipment state monitoring technical field.Method includes: to the multi-source heterogeneous monitoring data pre-processing obtains feature matrix, equipment label is mapped as label vector;Through label modulation deep heterogeneous cross attention fusion network, cross attention fusion is carried out using the control parameter generated by label, and fusion feature vector is output;Variable and component health index, equipment comprehensive score and long-term deviation cumulative factor are generated by label enhanced layered health decoding network;Label modulation deep boundary mapping evaluator dynamically determines health grade and percentage score;Multi-scale time series convolution trend evaluation network outputs trend risk index and state according to historical sequence;Output evaluation report and according to maintenance feedback online fine-tuning misjudgment related parameter subset.System is composed of corresponding module.The application realizes label-driven personalized evaluation and dynamic self-optimization.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment condition monitoring technology, specifically to a health assessment method and system that integrates equipment tags and multi-source data. Background Technology

[0002] Industrial equipment online monitoring systems are widely used in pumps, fans, motors, and power distribution cabinets, capable of collecting various parameters such as temperature, vibration, and current. Some systems can also provide threshold alarms and simple reports. However, there are still some limitations in health assessment.

[0003] First, the depth of multi-source data fusion is limited. Existing systems mostly use single-variable threshold judgment or simple parameter superposition, which makes it difficult to comprehensively reflect the overall status of the equipment. Data with different sampling frequencies, dimensions, and structures, such as vibration, temperature, current, and maintenance records, lack a unified fusion mechanism.

[0004] Secondly, individual differences in equipment are not fully reflected. Even for the same type of equipment, the health assessment logic should differ due to variations in installation location, operating conditions, key components, historical faults, and maintenance status. However, existing methods typically apply a uniform model and fixed thresholds, leading to a disconnect between assessment results and the actual operating status of the equipment.

[0005] Furthermore, the discrepancy between the assessment results and the actual on-site situation is quite prominent. Due to the lack of effective integration of equipment tags, operating condition information, and historical records, situations often arise where monitoring indicators are normal but the equipment already has potential hazards, or indicators are high but the equipment is controllable, reducing the credibility of the assessment.

[0006] In addition, most existing health assessment mechanisms are static or set only once, lacking the ability to continuously correct based on maintenance feedback and the deterioration process, making it difficult to adapt to changes in the condition of equipment during long-term operation.

[0007] Therefore, it is necessary to provide a health assessment solution that can deeply integrate multi-source heterogeneous data and perform personalized assessment, dynamic adjustment and continuous optimization based on device tags. Summary of the Invention

[0008] The purpose of this invention is to provide a health assessment method and system that integrates device tags and multi-source data, thereby achieving tag-driven personalized assessment and dynamic self-optimization.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A health assessment method that integrates device tags and multi-source data includes: Step 1: Collect multi-source heterogeneous monitoring data of the target device, including real-time monitoring data, historical operation data, and business record data; Step 2: Perform cleaning, time alignment, normalization, feature extraction, and windowing processing on the multi-source heterogeneous monitoring data to obtain the feature matrix; Step 3: Obtain the tag set of the target device and map the tag set to a tag vector. The tag set includes device category, operating condition, key components and historical fault tags. Step 4: Input the feature matrix and label vector into the label modulation deep heterogeneous cross-attention fusion network. The label deep modulation module generates multiple fusion control parameters from the label vector, including the multi-source attention temperature coefficient vector, the cross-channel interaction gating matrix, and the feature importance modulation vector. The multi-source heterogeneous cross-attention fusion module uses the fusion control parameters to perform temperature-adjusted and gated cross-attention fusion on the multi-source channels in the feature matrix and outputs the fused feature vector. Step 5: Input the component query matrix and component importance vector generated by the fusion feature vector and label deep modulation module into the label-enhanced hierarchical health decoding network. Generate variable-level health indicators, component-level health indicators and equipment comprehensive health scores through variable health decoder, component aggregation encoder and equipment comprehensive scorer respectively. Update the long-term deviation cumulative factor based on the equipment comprehensive health score. Step 6: Input the device's comprehensive health score and the deep modulation features output by the tag deep modulation module into the tag modulation deep boundary mapping evaluator. Generate a dynamic health level boundary vector based on the deep modulation features through the tag boundary generation network. Determine the health level by comparing the device's comprehensive health score with the health level boundary vector, and calculate a percentage score. Step 7: Input the device comprehensive health score sequence and long-term deviation cumulative factor sequence from the current time and multiple preceding time windows into the multi-scale temporal convolutional trend evaluation network. Extract multi-scale degradation features through convolutional branches containing different scales, and output the trend risk index and trend state category after fusion by the scale attention mechanism. Step 8: The structured output includes a health assessment report containing health level, percentage score, trend risk index, trend status category, and maintenance recommendations; when maintenance feedback data containing verified component status is received, only the subset of parameters related to misjudgment in the fusion network, hierarchical health decoding network, boundary mapping evaluator, and trend assessment network are fine-tuned and updated online.

[0010] Further, in step 1, the real-time monitoring data includes vibration acceleration data collected by a piezoelectric accelerometer, temperature data collected by a Pt100 thermal resistor, current data collected by a Hall current sensor, voltage data collected by a voltage transformer, pressure data collected by a piezoresistive pressure transmitter, flow data collected by a turbine flow meter, rotational speed data collected by a photoelectric encoder, and active power data collected by a power transmitter; the business record data includes inspection records, spot inspection records, maintenance records, and work order records.

[0011] Further, in step 4, the tag deep modulation module consists of a linear projection layer and multiple cascaded tag residual enhancement blocks. Each enhancement block contains a main branch and residual connections, and outputs deep modulation features. A parameter decoding array maps the deep modulation features in parallel into a multi-source attention temperature coefficient vector, a cross-channel interactive gating matrix, a feature importance modulation vector, a fusion layer weight synthesis vector, a component query matrix, and a component importance vector.

[0012] Further, in step 5, the variable health decoder includes a common feature compression layer shared by all sub-networks and multiple independent variable health decoding sub-networks; each sub-network generates corresponding variable-level health indicators from fused feature vectors and variable embeddings through a residual fully connected structure; the component aggregation encoder uses the row vectors of the component query matrix as queries and the encoded features of the variable-level health indicators as keys and values, and performs label-guided attention aggregation under the mask of the variable-component association matrix to obtain component-level health indicators; the device comprehensive scorer uses the component importance vector to perform a weighted harmonic average of the component-level health indicators to calculate the device comprehensive health score.

[0013] Further, in step 6, the label boundary generation network is a deep fully connected mapping network with skip connections. Its input is a deep modulation feature, and its output, after recursive multiplication and upper limit calibration, generates four monotonically decreasing boundary values ​​located in the interval [0, 1], which correspond to the boundary points of the five health levels from A to E, respectively. The percentage score is obtained by nonlinear mapping based on the relative position of the device's comprehensive health score in the corresponding level interval and the preset score range of that level.

[0014] Furthermore, the multi-scale temporal convolutional trend evaluation network includes short-period branches, medium-period branches, and long-period branches. The three branches use ordinary one-dimensional convolutional kernels or dilated one-dimensional convolutional kernels of different sizes to adapt to degradation patterns over different time spans. The scale attention mechanism automatically learns fusion weights from the global pooling features of the three branches through a gating network.

[0015] Furthermore, the label-modulated deep heterogeneous cross-attention fusion network, the hierarchical health decoding network, the label-modulated deep boundary mapping evaluator, and the multi-scale temporal convolutional trend evaluation network are jointly trained in an end-to-end manner during the offline stage. The loss function of the joint training includes device percentage score regression loss, component state classification loss, level boundary penalty loss, and trend analysis loss.

[0016] This invention also provides a health assessment system that integrates device tags and multi-source data, applied to the aforementioned health assessment method that integrates device tags and multi-source data, comprising: The data acquisition module is used to collect multi-source heterogeneous monitoring data from the target device; The data preprocessing module is used to perform cleaning, time alignment, normalization, feature extraction and windowing processing on the multi-source heterogeneous monitoring data to obtain a feature matrix; The tag acquisition module is used to acquire the tag set of the target device and map it into a tag vector. The data label fusion module is used to input the feature matrix and the label vector into the label modulation deep heterogeneous cross-attention fusion network and output the fused feature vector. The health indicator construction module is used to generate variable-level health indicators, component-level health indicators, equipment comprehensive health scores, and long-term deviation accumulation factors by utilizing the fused feature vector and label control parameters through a label-enhanced hierarchical health decoding network. The health assessment module is used to calculate the comprehensive health score and depth modulation features of the device through the tag modulation depth boundary mapping evaluator, and output the health level and percentage score. The trend analysis module is used to analyze historical health score sequences and cumulative factor sequences through a multi-scale temporal convolutional trend assessment network, and output trend risk index and trend status category. The results output and update module is used to generate and output structured health assessment reports and perform selective online fine-tuning updates to network parameters based on maintenance feedback data.

[0017] In summary, the present invention has at least one of the following beneficial technical effects: First, by unifying the preprocessing of multi-source heterogeneous monitoring data and cross-attention fusion of tag modulation, the depth of information interaction between different types of data is improved, making the overall health assessment more reflective of the comprehensive status of the equipment.

[0018] Second, by introducing device tags to deeply participate in the generation of fusion control parameters, attention modulation, and boundary generation, the evaluation logic is adaptively adjusted according to the differences in device tags, and the evaluation results are closer to the actual working conditions of specific devices.

[0019] Third, a hierarchical health indicator system covering variable, component, and equipment layers was constructed, and combined with trend risk factors, to output hierarchical and traceable health conclusions, enhancing the interpretability and engineering operability of the assessment.

[0020] Fourth, by using a multi-scale temporal convolutional trend evaluation network, it is possible to identify slow deterioration or accelerated deterioration trends before the health score is seriously exceeded, providing early warning for maintenance decisions.

[0021] Fifth, through the selective parameter online fine-tuning mechanism, only a subset of relevant parameters are updated based on maintenance feedback and field verification data, enabling the evaluation model to be continuously optimized as the equipment is operated, maintained, and feedback is received. This accumulates experience while avoiding catastrophic amnesia, making it more suitable for long-term industrial applications. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a comparison chart of the health levels of water pump P-105A under different evaluation methods in an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0024] like Figure 1 As shown, the present invention proposes a health assessment method that integrates device tags and multi-source data, comprising: Step 1: Collect multi-source heterogeneous monitoring data of the target device, including real-time monitoring data, historical operation data, and business record data; Step 2: Perform cleaning, time alignment, normalization, feature extraction, and windowing processing on the multi-source heterogeneous monitoring data to obtain the feature matrix; Step 3: Obtain the tag set of the target device and map the tag set to a tag vector. The tag set includes device category, operating condition, key components and historical fault tags. Step 4: Input the feature matrix and label vector into the label modulation deep heterogeneous cross-attention fusion network. The label deep modulation module generates multiple fusion control parameters from the label vector, including the multi-source attention temperature coefficient vector, the cross-channel interaction gating matrix, and the feature importance modulation vector. The multi-source heterogeneous cross-attention fusion module uses the fusion control parameters to perform temperature-adjusted and gated cross-attention fusion on the multi-source channels in the feature matrix and outputs the fused feature vector. Step 5: Input the component query matrix and component importance vector generated by the fusion feature vector and label deep modulation module into the label-enhanced hierarchical health decoding network. Generate variable-level health indicators, component-level health indicators and equipment comprehensive health scores through variable health decoder, component aggregation encoder and equipment comprehensive scorer respectively. Update the long-term deviation cumulative factor based on the equipment comprehensive health score. Step 6: Input the device's comprehensive health score and the deep modulation features output by the tag deep modulation module into the tag modulation deep boundary mapping evaluator. Generate a dynamic health level boundary vector based on the deep modulation features through the tag boundary generation network. Determine the health level by comparing the device's comprehensive health score with the health level boundary vector, and calculate a percentage score. Step 7: Input the device comprehensive health score sequence and long-term deviation cumulative factor sequence from the current time and multiple preceding time windows into the multi-scale temporal convolutional trend evaluation network. Extract multi-scale degradation features through convolutional branches containing different scales, and output the trend risk index and trend state category after fusion by the scale attention mechanism. Step 8: The structured output includes a health assessment report containing health level, percentage score, trend risk index, trend status category, and maintenance recommendations; when maintenance feedback data containing verified component status is received, only the subset of parameters related to misjudgment in the fusion network, hierarchical health decoding network, boundary mapping evaluator, and trend assessment network are fine-tuned and updated online.

[0025] Next, the above method will be elaborated in detail with specific implementation examples and parameter settings, as follows: In step 1, multi-source heterogeneous monitoring data of the target device is collected. This multi-source heterogeneous monitoring data includes real-time monitoring data, historical operational data, and business record data, specifically: Acquire real-time monitoring data, historical operation data, and business record data related to the target device.

[0026] The real-time monitoring data is collected through a sensor network deployed on the device, specifically including: vibration acceleration data collected by a piezoelectric accelerometer, temperature data collected by a Pt100 thermal resistor, current data collected by a Hall current sensor, voltage data collected by a voltage transformer, pressure data collected by a piezoresistive pressure transmitter, flow data collected by a turbine flow meter, rotational speed data collected by a photoelectric encoder, and active power data collected by a power transmitter.

[0027] The historical operating data is a collection of time-series data from the aforementioned sensors over a preset period of time, such as a snapshot of minute-level historical data from the past year.

[0028] The business record data includes structured or semi-structured text data from the computerized maintenance management system, specifically: inspection records, spot check records, maintenance records, and work order records. These records contain descriptions of discovered problems, handling measures, spare parts replacement information, and handling conclusions.

[0029] In step 2, the multi-source heterogeneous monitoring data undergoes cleaning, time alignment, normalization, feature extraction, and windowing to obtain the feature matrix, specifically: Perform cleaning, alignment, normalization, feature extraction, and windowing processing on data from different sources, with different frequencies, and different dimensions.

[0030] First, for each data source At any moment raw collected values Perform cleaning, removing outliers based on the Raida criterion: If If the value is not found, it is marked as abnormal and replaced with the nearest valid value from the same time point. For data source The average value over long-term operation. The standard deviation is denoted as .

[0031] Next, the cleaned data is processed according to a preset unified reference time series. Perform time alignment. For high-sampling-rate data sources such as vibration data, at each alignment time... The time window that falls Statistical features are extracted from the data. For business record data with low sampling rate or asynchronous arrival, a time alignment technique based on Lagrange interpolation is used to continuously map the impact of the occurrence time onto the adjacent alignment time label.

[0032] Then, the time-aligned data is subjected to min-max normalization to unify the value range of the heterogeneous data. Interval: ; in For alignment time The value, and Data sources The historical minimum and maximum values.

[0033] Next, feature extraction is performed on the normalized data. A feature extraction engine with a parallel multi-branch structure is adopted, which includes: a Fast Fourier Transform branch for vibration data, which outputs the vibration mean and peak frequency; a sliding window statistics branch for slowly variable data, which outputs the mean, standard deviation, and slope of change; and a semantic encoding branch for business record data, which converts the text into numerical vectors based on a pre-trained word embedding model and calculates the defect quantification factor based on the handling code.

[0034] Finally, a fixed-length sliding window is applied to segment the extracted continuous feature sequence, forming a series of feature matrices. Each matrix serves as a basic input unit for a subsequent fusion model. Each feature matrix is ​​denoted as... ,in This represents the total number of data source channels after feature extraction. The feature length within each sliding window.

[0035] In step 3, the tag set of the target device is obtained and mapped to a tag vector. The tag set includes device category, operating condition, key components, and historical fault tags, specifically: Read the target device's equipment attributes, operating condition attributes, environmental attributes, component attributes, historical attributes, and risk attribute tag information.

[0036] First, a standardized equipment tag library is constructed, which categorizes tags into six main types: equipment category tags, installation location and process section tags, operating condition tags, environmental condition tags, critical component tags, and historical failure and risk level tags. Each tag is defined as a key-value pair, with its value being a real-valued initial influence factor.

[0037] The reading process begins with receiving the unique identifier of the target device. The system then performs a join query in the device management database. The query operations are performed on the device basic information table, operating process table, environmental monitoring table, bill of materials table, historical maintenance record summary table, and risk assessment table, extracting the corresponding key-value pairs.

[0038] Aggregate all the key-value pairs retrieved to form a label vector. Specifically, for binary tags, such as whether a bearing is a critical component, the value is mapped to 1 or 0; for polymorphic tags, such as high, medium, and low risk level tags, the value is mapped to a continuous value. The real numbers within the interval. The final tag vector, as a core set of control parameters, fully defines the individual characteristics and operating context of the target device.

[0039] In step 4, the feature matrix and label vector are input into the label-modulated deep heterogeneous cross-attention fusion network. The label deep modulation module generates multiple fusion control parameters from the label vector, including a multi-source attention temperature coefficient vector, a cross-channel interaction gating matrix, and a feature importance modulation vector. The multi-source heterogeneous cross-attention fusion module then uses these fusion control parameters to perform temperature-adjusted and gated cross-attention fusion on the multi-source channels in the feature matrix, outputting a fused feature vector. Specifically: The preprocessed multi-source heterogeneous monitoring data is associated with the target device tag set, and the interpretation method, weight configuration, and sensitivity setting of different data are adjusted according to the tag information. This step uses a tag-modulated deep heterogeneous cross-attention fusion network, whose overall architecture consists of four parts: a data initial encoding module, a tag deep modulation module, a multi-source heterogeneous cross-attention fusion module, and a fusion feature output module. A detailed description of this module follows: 1. Preliminary Data Encoding Module The input to this module is a feature matrix obtained after windowing in step 2. ,in Number of data source channels The feature length within each sliding window. This module aims to uniformly encode and augment the representation of single-dimensional feature sequences from different physical sources.

[0040] First, for each data source channel Set up a data source-specific convolutional encoder. This encoder consists of stacked one-dimensional convolutional layers, batch normalization layers, and activation function layers, with the following structure: First layer: One-dimensional convolution, with 1 input channel and 1 output channel. The kernel size is Step length ,filling Then, batch normalization and ReLU activation are performed.

[0041] Second layer: One-dimensional convolution, 16 input channels, [number of output channels not specified] kernel size Step length ,filling Then, batch normalization and ReLU activation are performed.

[0042] Third layer: One-dimensional convolution, 32 input channels, [number of output channels not specified] kernel size Step length ,filling Then, batch normalization and ReLU activation are performed.

[0043] After the above three convolutional layers, each data source channel The original sequence Encoded as a time series containing 64 feature maps ,in This represents the length of the downsampled sequence. The number of channels, 64, is designed as a uniform feature dimension across all data sources.

[0044] Then, to enable the model to subsequently recognize the physical properties of different data sources, a data source type encoding is introduced. A learnable type embedding matrix is ​​defined. For each data source channel Assign a type embedding vector Add it via broadcast mechanism. At each step of time: ; After that, all channels By concatenating the components along the time dimension, a large feature tensor is formed. This serves as the input for subsequent cross-attention fusion.

[0045] 2. Tag depth modulation module The input to this module is the tag vector of the target device. Vector dimension For the total number of labels, each component has been normalized to [value]. The function of this module is to nonlinearly map low-dimensional discrete-continuous hybrid label information into a set of high-dimensional, structured fusion control parameters, guiding all aspects of the subsequent attention fusion process.

[0046] The module's specific structure is a deep residual mapping network, and its processing flow is as follows: First, the label vector After passing through a linear projection layer, the dimension is increased to : ; in , .

[0047] Subsequently, This is achieved through three cascaded tag residual enhancement blocks. Each enhancement block has the same structure, containing two branches: The main branches are, in order: linear layer ReLU, linear layer ; The residual connection branch adds the input of the enhancement block directly to the output of the main branch, followed by a layer normalization operation. The operation of an enhancement block can be represented as: ; in , , After three enhancement blocks, deep modulation features are obtained. .

[0048] Finally, a parameter decoding array will The mapping is used to obtain all the control parameters required for subsequent fusion and evaluation. The array contains six parallel linear projection heads, each generating the following parameter sets: Multi-source attention temperature coefficient vector : by linear layer It is obtained by activation via Softplus, ensuring that each component is positive.

[0049] Cross-channel interactive gating matrix : by linear layer After being activated by sigmoid, it is recombined into a matrix to control the flow of information between sources.

[0050] Fusion layer weight synthesis vector : by linear layer It is then activated by sigmoid and used to synthesize the query base for multi-source attention.

[0051] Feature Importance Modulation Vector : by linear layer The weights are then normalized using softmax and used as initial prior weights.

[0052] Component Query Matrix : by linear layer And obtained through output recombination, wherein This represents the preset total number of components. Each row of this matrix... Represents the label for the component The specific query vector for the attention method is passed to step 5 for tag-guided component aggregation.

[0053] Component Importance Vector : by linear layer And obtained by softmax normalization, ensuring This vector is passed to step 5 for calculating the weighted harmonic average of the overall equipment score.

[0054] The above six sets of parameters are all uniquely determined by the tag content, enabling the tags to have in-depth and comprehensive control over the subsequent fusion and evaluation process.

[0055] 3. Multi-source heterogeneous cross-attention fusion module This module is the core computing unit, receiving the output from the preliminary data encoding module. The control parameters output by the label depth modulation module are used to perform a multi-step, heterogeneous cross-attention fusion process.

[0056] Step 1: Construct a heterogeneous query matrix. Each data source... The corresponding feature sequences are first shrunk into vectors through a shared time pooling operation. The pooling method is global average pooling along the time dimension. Then, the feature importance modulation vector obtained from the label modulation module is used. and fusion layer weight synthesis vector Generate a heterogeneous query matrix that is aware of the tag context. , of which The row is constructed as: ; This design allows the query vector to contain information from the data source itself, as well as a shared base that is cross-source and controlled by labels, thus breaking the homogeneity of traditional self-attention queries.

[0057] Step 2: Perform cross-attention with temperature regulation and gating. Rearranged as key tensors Sum tensor All shapes This involves concatenating the feature vectors from all time steps across all data sources into a long sequence. When calculating the attention score, a temperature coefficient vector generated from the labels is introduced. and gating matrix To control cross-source interactions.

[0058] Specifically, for a row of the query matrix Corresponding data source The query first calculates the dot product similarity between the query and all key vectors, and then divides it by the temperature coefficient. : ; in The first key tensor in the key tensor A 64-dimensional vector of rows, Traverse all Index of a key vector. Temperature coefficient. The introduction of this feature allows the sharpness of attention distribution across different data sources to be dynamically determined by the labels.

[0059] Next, label-based gating is applied to regulate the cross-source information flow. A gating mask of the same size as the scoring matrix is ​​generated. Its elements Defined as follows: If key The data source belongs to ,but ,in Gating matrix generated for the label The Middle Line 1 The column element, whose value indicates the labels allowed from the data source. To data source The intensity of information transmission. Applying a gating mask to the score: ; Then to Attention weights are obtained by applying softmax to each row. : ; Step 3: Weighted Aggregation and Fusion Output. The obtained attention weights are used to perform a weighted summation of the value tensors to obtain the output from each data source. The output represents : ; in The value of the tensor in the first A 64-dimensional vector of rows.

[0060] All these output vectors are concatenated to generate a fused feature vector. Then, it is projected through a linear layer to the final feature dimension of 128, yielding the input required for step 5. : ; in , , This means flattening the matrix into a vector.

[0061] This step outputs the fused feature vector. Step 5: Output depth modulation features Step 6: Output the component query matrix. and component importance vector Proceed to step 5.

[0062] In step 5, the component query matrix and component importance vector generated by the fusion feature vector and label deep modulation module are input into the label-enhanced hierarchical health decoding network. The variable-level health index, component-level health index, and equipment comprehensive health score are generated through the variable health decoder, component aggregation encoder, and equipment comprehensive scorer, respectively. The long-term deviation cumulative factor is then updated based on the equipment comprehensive health score. Specifically: Based on the fusion results, health indicators reflecting the status of single variables, components, and equipment are established, forming the basis for comprehensive health assessment. This step implements a label-enhanced hierarchical health decoding network, whose overall architecture consists of three parts: a variable health decoder, a component aggregation encoder, and an equipment comprehensive scorer. The input is the fusion feature vector output from step 4. Component query matrix and component importance vector The output includes a set of variable health indicators, a set of component health indicators, a comprehensive equipment health score, and trend risk factors. These will be discussed in detail below: 1. Variable Health Decoder The function of this decoder is to fuse device-wide features. The analysis is performed as a health status indicator for each independent monitoring variable. To enhance feature representation and robustness, a compact residual fully connected network is used for the health indicator calculation branch of each variable.

[0063] First, define the set of variables to be evaluated. ,in For a typical rotating machine, the preset total number of monitored variables is taken as follows: These correspond to the mean vibration value, peak vibration frequency, mean temperature, rate of temperature rise, mean current, current fluctuation coefficient, mean pressure, and rotational speed deviation rate, respectively. This set of variables does not correspond one-to-one with the data source collected in step 1, but rather are semantic variables after feature extraction, determined by the system configuration based on the equipment type during deployment.

[0064] For each variable Construct an independent variable health decoding subnetwork The subnetwork adopts a residual structure, and its specific components are as follows: Common Feature Compression Layer: A linear layer that compresses... The 128-dimensional features are compressed to 64-dimensionality, followed by batch normalization and the LeakyReLU activation function with a slope of 0.01. This layer is shared by all sub-networks, outputting common compressed features. This reduces the number of parameters and extracts common features among variables.

[0065] Variable-Specific Embedding Injection: Define a learnable variable embedding matrix Sub-network Read its corresponding embedding vector ,and By performing element-wise addition, we obtain the variable-perceptual features. .

[0066] The residual block contains two fully connected layers. The first layer maps 64 dimensions to 128 dimensions, followed by batch normalization and ReLU activation; the second layer maps 128 dimensions back to 64 dimensions, followed by batch normalization. After undergoing a linear projection transformation to the same dimension as the second layer output, the results are used as residual connections, summed, and then activated by ReLU to obtain... This operation can be represented as: ; in , , For the projection matrix, This indicates batch normalization operation.

[0067] Output layer: A linear layer maps 64 dimensions to 1 dimension, followed by a sigmoid activation function, directly outputting the health index of the variable. The closer the value is to 1, the healthier the condition.

[0068] ; in For the Sigmoid function, , .

[0069] all The subnetworks are processed in parallel to obtain the variable layer health indicator vector. .

[0070] 2. Component-integrated encoder This module aims to aggregate variable-level health indicators into component-level health indicators according to the equipment's physical structure. The aggregation process does not use fixed weights, but rather a label-guided graph attention aggregation mechanism, allowing the assessment of component importance and the allocation of variable contributions to components to be dynamically adjusted based on equipment labels.

[0071] First, define the component set. , The total number of components, compared with the component query matrix in step 4. The number of rows is consistent. Based on the knowledge of the equipment bill of materials, establish a binary variable-component association matrix. If variable The measurement location belongs to the component Or can directly reflect the components In this state, Otherwise, it is 0.

[0072] Vector of variable health indicators As input, the variable feature vector is first obtained by dimensionality enhancement through a positional encoding layer. The lifting method is a linear mapping plus ReLU activation: ; in , .

[0073] Then, the component query matrix passed in step 4. each line For the query, use the feature vectors of all variables. For keys and values, perform label-guided attention aggregation on a component-by-component basis. The polymerization process is as follows: When calculating attention scores, an association matrix is ​​introduced. As an attention mask, it ensures that only variables physically relevant to the component are noticed. The attention score is calculated as follows: ; in The denominator represents the dot product of two 32-dimensional vectors. This is the scaling factor.

[0074] Applying a related mask: If Then let .

[0075] Attention weights are obtained through softmax normalization. : ; part Aggregated feature vectors We obtain the result through weighted summation: ; Finally, the component health indicators are obtained through an output layer. : ; in , , This is a sigmoid function. This design allows query vectors for important or risky components to automatically focus on the most relevant variables during the learning process, and the intensity of this focus is modulated by the labels.

[0076] all The components are computed in parallel, ultimately yielding a component-level health indicator vector. .

[0077] 3. Equipment Comprehensive Scoring Device This module integrates component-level health indicators into a comprehensive equipment health score. The integration method uses a weighted harmonic mean to fully reflect the negative impact of severe degradation of any component on the overall equipment health. The weights directly use the component importance vector passed in step 4. This vector is dynamically determined by the label information.

[0078] Equipment comprehensive health score The calculation is as follows: ; in To prevent division by zero. The closer the value is to 1, the healthier the overall condition of the device. This score is output to step 6 for final rating determination and is also recorded by the trend analysis module.

[0079] 4. Trend Risk Factor Introduce a long-term deviation cumulative factor This factor is used to detect slow deterioration in device health. It is updated on each assessment using an exponentially weighted moving average. ; in The decay factor is the initial time. . The value is passed to step 7 as one of the core inputs for trend analysis.

[0080] This step outputs a vector of variable health indicators. Component health index vector Comprehensive equipment health score and long-term deviation cumulative factor . and Passed to step 8 for report generation. Pass to steps 6 and 7. Proceed to step 7. The system maintains a length of The historical evaluation results queue stores the results from the most recent 64 time windows. and The value is used for trend analysis in step 7.

[0081] In step 6, the device's overall health score and the depth modulation features output by the tag depth modulation module are input into the tag modulation depth boundary mapping evaluator. The tag boundary generation network generates a dynamic health level boundary vector based on the depth modulation features. The health level is determined by comparing the device's overall health score with the health level boundary vector, and a percentage score is calculated. Specifically: Based on the health indicator system, the target equipment is assessed for its health level or health score, and the current health status is output. This step implements a tag modulation depth boundary mapping evaluator, whose input is the comprehensive health score of the equipment output in step 5. and the depth modulation features output by the tag depth modulation module in step 4 The output is a health level and a percentage score. The evaluator consists of three parts: a label boundary generation network, a rating determiner, and a rating mapper. Each part will be described in detail below: 1. Label Boundary Generation Network The function of this network is to dynamically generate a set of asymmetric health level boundary vectors specific to each device based on device tag information. Fixed thresholds are replaced with a set of dynamic values ​​calculated by the deep network, allowing the threshold distribution and intervals to be arbitrarily shaped according to device characteristics.

[0082] The network structure employs a deep fully connected mapping network with skip connections, and its input is... The output consists of four upper boundary values ​​for five state levels. ,satisfy The boundary meaning is: when At that time, it was grade A. At that time, it was grade B. At that time, it was grade C. At that time, it was classified as Grade D. It was at level E.

[0083] The specific network structure is as follows: Input layer: As input.

[0084] First fully connected block: Linear layer This is followed by a batch normalization layer and the GELU activation function. The output is denoted as... .

[0085] Second fully connected block: Linear layer This is followed by a batch normalization layer and the GELU activation function. The output is denoted as... .

[0086] Third fully connected block: Linear layer This is followed by a batch normalization layer and the GELU activation function. The output is denoted as... .

[0087] Jump connection fusion: Through a linear projection layer Transform to Same dimension, then with Element-wise addition, followed by LayerNorm normalization, yields the fused features. This design ensures that deep networks can take into account low-level features and prevents over-compression of label information.

[0088] Boundary regression output layer: a linear layer Output a four-dimensional vector To ensure that the boundary is decreasing and located at... Follow these steps for post-processing: First, Compressed using the Sigmoid function to The interval is obtained .

[0089] Then, we force a monotonically decreasing trend using recursive multiplication: Let ,right Set recursively: ; Right now , , .because Naturally satisfied .

[0090] Finally, perform upper bound calibration on the top boundary: ; ; ; ; Ensure that even the healthiest devices maintain a certain level of differentiation. Final boundary vector. That is the output.

[0091] 2. Grading and Scoring Mapping Obtain the boundary vector Afterwards, The algorithm compares the value with four boundary values ​​successively to determine the grade range it falls into. The decision logic is as follows: like Grade A; Otherwise, if The grade is B; Otherwise, if The grade is C; Otherwise, if The grade is D; Otherwise, the grade is E.

[0092] While outputting the grade, calculate the percentage score. It adopts a nonlinear mapping method within the level.

[0093] First, define the lower and upper bounds of the interval. For the selected level: Level A, , Grade B , Grade C , Grade D , Class E , .

[0094] Then, calculate Relative position within this interval : ; in, . Cut off at between.

[0095] Then, the mapped score is calculated based on the preset percentage range for each level. The percentage ranges for A to E are as follows: , , , , Let the lower bound of the grade range be . The upper boundary is The final score is: ; This rating is given directly to the user and serves as a digital summary of the device's overall status.

[0096] 3. Summary of forward reasoning process The complete forward inference process for step 6 is as follows: First step, receive from step 5 Received from step 4 .

[0097] The second step, Input label boundary generation network, through A depth map with skip connections, producing .

[0098] The third step is comparison. and Determine the health level.

[0099] Fourth step, utilize Calculate the relative position within the grade range. .

[0100] Step 5: Output health level and At the same time The data is then passed to the trend analysis module in step 7.

[0101] The health level output in this step and Passed to step 8 for report generation. The current value and historical sequence are passed to step 7.

[0102] In step 7, the device comprehensive health score sequence and long-term deviation cumulative factor sequence from the current time and multiple preceding time windows are input into a multi-scale temporal convolutional trend assessment network. Multi-scale degradation features are extracted through convolutional branches containing different scales, and then fused using a scale attention mechanism to output a trend risk index and trend state category. Specifically: This step compares the health results of the target device across multiple time windows to identify trends and deterioration tendencies in its health status. A multi-scale temporal convolutional trend evaluation network is implemented, consisting of a sequence construction layer, a multi-scale temporal feature extraction backbone, a cross-scale attention fusion layer, and a trend risk decoder. The network input is the current time step. and the preceding sequence The comprehensive health score sequence of the equipment calculated within a time window and the corresponding long-term deviation cumulative factor sequence ,in The system uses a preset observation sequence length to cover a sufficiently long historical interval to capture slow degradation. In step 5, the system maintains a historical evaluation result queue of length 64 to provide the aforementioned sequence. The output is a trend risk index. The trend state categories include four types: stable, slow deterioration, accelerated deterioration, and step change. These will be explained in detail below: 1. Sequence Construction Layer Align the two input sequences along the time steps to form a two-channel time series. The first channel is The sequence, the second channel is Sequence. Due to The sequence itself is The exponential moving average of the bias, along with the complementary information between the two, helps the network capture the coupling pattern between instantaneous fluctuations and long-term cumulative effects of the scores. Before inputting the convolutional network, [the following is likely a separate, unrelated sentence:] ... Transpose This is based on the channel dimension convention for matching one-dimensional convolution.

[0103] 2. Multi-scale temporal feature extraction backbone To simultaneously capture short-term fluctuations, medium-term shifts, and long-term downward trends, a multi-scale feature extraction backbone with three parallel convolutional branches is designed. Each branch employs a one-dimensional convolutional kernel of different sizes to achieve specialized responses to patterns across different time spans.

[0104] The short-period branch consists of two cascaded one-dimensional convolutional blocks. The first convolutional block has 2 input channels, 16 output channels, and a kernel size of [missing information]. Step length ,filling This is followed by batch normalization and GELU activation. The second convolutional block has 16 input channels and 32 output channels, with a kernel size of [size missing]. Step length ,filling This is followed by batch normalization and GELU activation. This branch has a smaller receptive field and is good at detecting high-frequency fluctuations or sudden jumps in scores over a short period of time.

[0105] The intermediate periodic branch consists of two one-dimensional convolutional blocks. The first convolutional block has 2 input channels and 16 output channels, with a kernel size of [missing information]. Step length ,filling This is followed by batch normalization and GELU activation. The second convolutional block has 16 input channels and 32 output channels, with a kernel size of [size missing]. Step length ,filling This is followed by batch normalization and GELU activation. This branch can capture phased, sustained declines in the device over several weeks or months.

[0106] The long-period branch employs dilated convolutions to achieve an exponentially increased receptive field without increasing the number of parameters. This branch contains three dilated convolutional layers. The first layer has 2 input channels, 16 output channels, and a kernel size of [missing information]. Expansion rate Step length This is followed by batch normalization and GELU activation. Second layer: 16 input channels, 32 output channels, kernel size... Expansion rate Step length This is followed by batch normalization and GELU activation. Third layer: 32 input channels, 32 output channels, kernel size... Expansion rate Step length This is followed by batch normalization and GELU activation. The effective receptive field after the three layers are stacked can cover the entire The sequence length can be used to directly model the monotonic degradation trend over long time spans.

[0107] Each branch ultimately compresses the sequence length. To maintain a uniform length for fusion, adaptive average pooling is used to normalize the time dimension to a fixed size. Therefore, short-cycle branch output Mid-cycle branch output Long-cycle branch output .

[0108] 3. Cross-scale attention fusion layer Different devices exhibit different dominant timescales during the degradation process. This layer uses a scale attention mechanism to automatically determine whether the current trend is mainly dominated by short-term fluctuations, medium-term shifts, or long-term declines, and integrates the characteristics of the three branches accordingly.

[0109] First, the outputs of the three branches are concatenated along the channel dimension to obtain a joint feature map. The number of channels is 96, which is 32+32+32. Then, for... Perform global average pooling to obtain a 96-dimensional global representation vector. .

[0110] The vector is then fed into a scale-attention gating network, which is a two-layer fully connected structure: ; in , ; , . The three components correspond to the weights of the short, medium, and long scales, respectively, and their sum is 1.

[0111] Perform global average pooling along the time dimension on the feature maps of each branch to obtain three 32-dimensional vectors: ; ; ; The fused trend characteristics are calculated as follows: ; in This refers to the trend characteristics after multi-scale fusion.

[0112] 4. Trend Risk Decoder The decoder uses fusion features Long-term deviation cumulative factor at the current moment As input, explicit knowledge is injected, and the output is a trend risk index and a state category.

[0113] First, and By splicing them together, we can obtain Then, decoding is performed through two fully connected layers: ; ; ; in , ; , ; , . The Sigmoid function outputs a trend risk index. A higher value indicates a clearer and more urgent trend of deterioration. The categories are probabilities of four trend states, and the category with the highest probability is taken as the trend state: 0 is stable, 1 is slow deterioration, 2 is accelerated deterioration, and 3 is a step change.

[0114] 5. Output and Connection This step outputs the trend risk index. and trend status category, with step 5 Step 6 Health Level and The results are then sent to step 8 for result encapsulation and feedback processing.

[0115] The following is a detailed explanation of the end-to-end joint training process for the entire network: All trainable parameters in the label-modulated deep heterogeneous cross-attention fusion network of step 4, the label-enhanced hierarchical health decoding network of step 5, the label-modulated deep boundary mapping evaluator of step 6, and the multi-scale temporal convolutional trend evaluation network of step 7 are trained using an end-to-end joint optimization approach, which specifically includes the following: 1. Training dataset construction A large number of equipment evaluation cases were extracted from the equipment historical operation database. Each case It includes the following components: A time window The feature matrix obtained after preprocessing the multi-source heterogeneous monitoring data in step 2. .

[0116] The tag vector of the device at the corresponding time. .

[0117] Multi-layered authenticity labels calibrated by on-site experts based on maintenance reports, operation logs, and post-inspection analysis: True value of equipment rating (out of 100) .

[0118] True values ​​of the binary states of each component , 0 indicates normal operation, while 1 indicates a deterioration or malfunction that requires attention.

[0119] Trend State Truth Value The score is determined based on whether a failure occurred within 30 days after the case and the score evolution curve before the failure. If there is no failure and the score is stable, it is marked as 0; if there is a slow monotonous decline and a failure eventually occurs, it is marked as 1; if the decline accelerates and a failure occurs in a short period of time, it is marked as 2; if the score drops sharply due to sudden factors and a failure occurs in a short period of time, it is marked as 3.

[0120] For cases with trend status 1, 2, or 3, also record the number of days since the failure. , , used for soft label calculation of trend risk index.

[0121] The dataset is grouped by device and divided into training, validation and test sets in an 8:1:1 ratio to ensure that data from different time periods on the same device does not appear in different groups and to prevent data leakage.

[0122] 2. Definition of Joint Loss Function Joint loss function Regression loss from equipment rating Component status classification loss Level boundary penalty loss and trend analysis loss It is composed of four weighted combinations: ; in , , , To balance the hyperparameters.

[0123] The specific definitions of each loss item are as follows: Equipment rating regression loss : We employ Huber loss to enhance robustness to anomalous labels. Let... : ; in , For batch size.

[0124] Component state classification loss We employ weighted focus loss to mitigate class imbalance and address hard-to-class samples. ; in , , This refers to the component health indicators output by Sigmoid.

[0125] Level boundary penalty loss : Boundary order loss and boundary rationality loss composition.

[0126] Boundary order loss forces a decrease across the four boundaries: ; in , , , .

[0127] Boundary reasonableness loss based on the true level Apply hinge penalty to push It falls within the correct interval. Set the interval parameter. ,but: like : ; like : ; like : ; like : ; like : ; Summary: ; Trend analysis loss Cross-entropy loss for trend state classification and trend risk index regression loss composition.

[0128] The standard multi-class cross-entropy loss is applied to... and the true trend status .

[0129] Using mean squared error, the target value is defined as follows: If Target value ;like The target value is based on the number of days until the failure. calculate: ; That is, the closer to the fault, the closer the target value is to 1. The calculation is as follows: ; in This represents the number of samples in the batch where the trend state is non-zero. Final: ; 3. Training process and parameter settings Joint training was performed using the AdamW optimizer, with the weight decay coefficient set to... Initial learning rate It uses a cosine annealing strategy with a linear warmup of 5 epochs. Batch size Total training rounds Calculate the validation set after each epoch. An early stopping strategy is adopted: if the validation loss does not improve within 30 consecutive epochs, training is terminated and the model parameters corresponding to the optimal validation loss are restored.

[0130] Data augmentation during training: Random time shifts are applied to the input sequence, with the shift amount ranging from... Uniform sampling is performed within each time step; Gaussian noise with a standard deviation of 0.01 is injected into the sequence values. This enhancement applies only to the input sequence in step 7, improving the robustness of the trend analysis network.

[0131] After training, save the complete model parameters. When deploying for inference, directly execute the forward propagation from steps 4 to 7, without needing to calculate gradients.

[0132] In step 8, the structured output includes a health assessment report containing health level, percentage score, trend risk index, trend status category, and maintenance recommendations. When maintenance feedback data containing verified component status is received, only the subset of parameters related to misjudgment in the fusion network, hierarchical health decoding network, boundary mapping evaluator, and trend assessment network are fine-tuned online. Specifically: Output health assessment results, trend results, and recommendations, and dynamically update assessment parameters or strategies based on on-site verification and maintenance feedback. The input for this step is the variable health indicator vector output from step 5. Component health index vector Comprehensive equipment health score and long-term deviation cumulative factor Step 6 outputs the health level and percentage score. The trend risk index output in step 7 The input includes trend status categories, and optionally structured feedback data returned from the maintenance work order system. The output is a structured report containing diagnostic results and maintenance recommendations, as well as updated model parameters. Step 8 will be explained in detail below: 1. Structured output of evaluation results This module summarizes, interprets, and diagnoses the multi-dimensional health information generated from each upstream step, producing a clear report for operations and maintenance personnel. The report generation process follows this workflow: First, from step 5 Identify the lowest scorer among them Each component Among these components, from step 5... In the middle, based on the component-variable correlation matrix The lowest-scoring variable belonging to each low-scoring component is identified as the main source of degradation for that component.

[0133] The report engine then combines the above data with the tag information to generate readable text. The report contains the following structural blocks: Equipment Overview: Label information such as equipment code, process section, and operating conditions.

[0134] Current status assessment: Health level, ranging from A to E, and a percentage score. .

[0135] Component-level health analysis: Lists the health indicators for each component. The report includes a brief analysis of the condition. For low-scoring components, the report identifies the main deterioration variables and their corresponding explanations. For example, a bearing component with a health score of 0.52 has a health score of 0.31, with the main deterioration source being the peak frequency of high-frequency vibration, which highly matches the characteristics of historical fault labels.

[0136] Trend Analysis: Provides the trend status, including stable, slow deterioration, accelerated deterioration, step change, and a trend risk index. When the trend is unstable and At that time, an early warning message will be generated.

[0137] Maintenance recommendations: Based on low-scoring components and identified sources of degradation, combined with component labels and historical maintenance records, match corresponding operation guidelines from a pre-set recommendation library. For example, recommend scheduling a shutdown for inspection and replacement of bearings, or recommend shortening the vibration monitoring cycle to once every 4 hours.

[0138] The final structured report data is encapsulated in JSON format and provided to upper-level monitoring interfaces, mobile applications, or alarm systems for presentation.

[0139] 2. Dynamic update mechanism based on maintenance feedback This mechanism is the core of the system's continuous learning, aiming to refine the model using real equipment conditions verified through on-site maintenance, making its assessments of similar equipment and operating conditions increasingly accurate. The update process is triggered by maintenance feedback events.

[0140] Feedback Data Standards: After each maintenance operation, maintenance personnel or condition analysts enter feedback data in a structured manner through the maintenance management system. A feedback record includes: a unique equipment identifier; a timestamp corresponding to this maintenance, used to trace the assessment input at that time; and a verified list of component statuses for each component. , give 0 indicates normal operation, and 1 indicates confirmed deterioration or malfunction requiring repair; if determined by a comprehensive expert assessment, a true percentage score for the entire equipment may be included. Actual trend conclusions and text annotations can be converted into true trend status labels. .

[0141] Selective parameter update strategy: To avoid the model forgetting a large range of historical data due to online fine-tuning, a localized update method is adopted. Only a subset of network parameters relevant to the current feedback is updated, specifically: In step 4, the tag depth modulation module determines the weights and biases of the six linear projection heads of the parameter decoding array, as well as the parameters of the last layer of the tag residual enhancement block.

[0142] In step 5, the variable health decoder uses an association matrix to verify the faulty component. All parameters of the associated variable subnetwork.

[0143] In step 5, the attention query vector for the corresponding faulty component in the component aggregation encoder. and output layer parameters.

[0144] Step 6 generates the output layer of the label boundary and the parameters of its most recent fully connected block.

[0145] Step 7: Parameters of the last two fully connected layers of the trend risk decoder.

[0146] During the update, other parameters such as the common feature compression layer, multi-scale convolution backbone, and preliminary data encoding module remain frozen, while only the aforementioned local parameters related to misjudgment are temporarily unfrozen.

[0147] Online fine-tuning of the loss function: For single or multiple feedback samples, a small batch of feedback can be accumulated, with a batch size of [missing information]. The value ranges from 1 to 8, and a fine-tuning loss is constructed. : ; in, To focus on the weighted focus loss between the true component labels and model predictions on the feedback samples, the form and joint training... Consistent. exist Enabled if it exists; otherwise, the value is 0. Enabled when the true level can be inferred; otherwise, it is 0. exist Enabled if it exists, otherwise 0. Coefficient , , Used to control various influences.

[0148] Optimizer and Update Process: The Adam optimizer is used to update the unfrozen parameters, including the learning rate. The algorithm iterates for 5 to 10 epochs on the feedback samples. After the update is complete, the parameters are immediately frozen, and the system reverts to inference mode. To prevent perturbations caused by a single erroneous feedback, gradient pruning is employed, with the maximum gradient norm limited to 1.0.

[0149] Furthermore, a feedback replay buffer is introduced to reinforce memory. During offline training, a subset of samples representing various typical failure modes is retained from the original training set as a core set. During online updates, the current feedback samples are mixed with 16 representative samples from the core set in a mini-batch for fine-tuning, further ensuring the model does not suffer from memory loss. The replay weights of the core set samples are multiplied by a coefficient of 0.1, allowing the model to primarily learn new feedback while maintaining basic memory of past patterns.

[0150] Closed-loop update: After executing the parameter update, the system records the parameter change log for this update, including the L2 norm of the parameter difference before and after the update. Maintenance and management personnel can observe more accurate component degradation identification and more accurate grade determination in subsequent evaluations of similar equipment. This mechanism supports continuous iteration with hundreds of feedback events per year without requiring full retraining.

[0151] In summary, the input and output of step 8 are: Input: From step 5 , , , Health level from step 6 From step 7 Trend status; feedback data from external systems .

[0152] Output: A structured evaluation report, displayed in JSON format or via a user interface; updated model parameters, saved to the model library.

[0153] Through the above design, step 8 not only completes the final presentation of information, but also connects the entire system into a self-optimizing closed loop, ensuring continuous improvement in assessment accuracy in complex industrial environments.

[0154] The technical solution and its beneficial effects of the present invention will be further explained below with reference to specific application examples.

[0155] Example 1: Personalized Health Assessment of Critical Water Pumps The system focuses on a circulating water pump (equipment number P-105A) in the central pump room of a chemical plant. This equipment is equipped with a vibration acceleration sensor, a temperature sensor, a current sensor, and a speed sensor. The system can also query its maintenance work order records for the past two years.

[0156] The label vectors obtained in step 3 include: pump equipment, central pump house, high load conditions, key bearing components, humid environment, high historical vibration, and medium risk level. Among these, the risk level in the polymorphic labels is mapped to a numerical value. Step 2: Extracted monitoring variables This includes the average vibration value, peak vibration frequency, average temperature, temperature rise rate, average current, current fluctuation coefficient, average pressure, and rotational speed deviation rate.

[0157] In step 4, the tag depth modulation module increases the attention weight of current fluctuation characteristics and vibration peak frequency characteristics based on high-load operating conditions and the tags of key bearing components. In step 6, the tag boundary generation network outputs the dynamic health level boundary for the device. Table 1 compares the dynamic boundaries corresponding to devices with different tag characteristics. It can be seen that the low-level threshold for devices with historical fault and risk tags is significantly raised, and the evaluation is more stringent.

[0158] Table 1 Comparison of Device Dynamic Health Level Boundaries with Different Label Features

[0159] During a particular assessment window, the equipment's overall health score was... If based on fixed boundaries The rating falls into grade C; and according to the device's dynamic boundaries, , and It was rated as Grade D, out of 100. After interval mapping calculation, it is approximately This result more accurately reflects the actual condition of the pump, which urgently requires attention due to excessive bearing vibration.

[0160] Step 7's multi-scale temporal convolutional trend evaluation network uses 64 historical windows. Serial and long-term deviation cumulative factor The sequence is taken as input. A continuous decline in the health indicators of the bearing components is detected. The output trend status is "slow deterioration," and the trend risk index is [not specified]. Step eight provides a warning in the report: "The bearing components are showing an increasing trend of deterioration; it is recommended to arrange an inspection."

[0161] Example 2: Dynamic Boundary Assessment and Trend Indication of Distribution Cabinets A high-voltage switchgear (equipment number G-201) in a substation is equipped with busbar temperature sensors and current transformers. A tag acquisition module reads tags indicating high-load operation, critical busbar components, and terminal risks. During the fusion process in step 4, these tags increase the gating weight of the temperature channel, facilitating the expression of the cross-source impact of temperature anomalies on electrical components.

[0162] In one assessment, the health index of the mother drain temperature rise rate variable was only [missing information]. According to the component aggregation encoder, the health indicators of the busbar components are as follows: Overall equipment rating The label boundary generation network will generate labels based on the terminal risk labels. Upgraded to ,therefore Grade B, out of 100. The system output a potential overheating risk warning for the wiring terminals. On-site infrared temperature measurement confirmed that the wiring terminals were slightly oxidized.

[0163] Figure 2 This is a comparison chart of the health levels of water pump P-105A under different assessment methods in an embodiment of the present invention. The results of the fixed boundary method and the dynamic boundary method of the present invention are compared over 12 consecutive assessment windows. It can be seen that the dynamic boundary method causes the level to drop from C to D earlier, reflecting a strict assessment strategy for equipment with risk labels.

[0164] The two embodiments above demonstrate that the present invention, through tag-driven dynamic boundary and multi-scale trend analysis, can provide personalized assessments and early warnings for different devices, significantly improving the engineering applicability of health assessments.

[0165] This invention also provides a health assessment system that integrates device tags and multi-source data, applied to the aforementioned health assessment method that integrates device tags and multi-source data, comprising: The data acquisition module is used to collect multi-source heterogeneous monitoring data from the target device; The data preprocessing module is used to perform cleaning, time alignment, normalization, feature extraction and windowing processing on the multi-source heterogeneous monitoring data to obtain a feature matrix; The tag acquisition module is used to acquire the tag set of the target device and map it into a tag vector. The data label fusion module is used to input the feature matrix and the label vector into the label modulation deep heterogeneous cross-attention fusion network and output the fused feature vector. The health indicator construction module is used to generate variable-level health indicators, component-level health indicators, equipment comprehensive health scores, and long-term deviation accumulation factors by utilizing the fused feature vector and label control parameters through a label-enhanced hierarchical health decoding network. The health assessment module is used to calculate the comprehensive health score and depth modulation features of the device through the tag modulation depth boundary mapping evaluator, and output the health level and percentage score. The trend analysis module is used to analyze historical health score sequences and cumulative factor sequences through a multi-scale temporal convolutional trend assessment network, and output trend risk index and trend status category. The results output and update module is used to generate and output structured health assessment reports and perform selective online fine-tuning updates to network parameters based on maintenance feedback data.

[0166] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0167] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0168] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0169] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the function specified in one or more boxes.

[0170] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A health assessment method that integrates device tags and multi-source data, characterized in that, include: Step 1: Collect multi-source heterogeneous monitoring data of the target device, including real-time monitoring data, historical operation data, and business record data; Step 2: Perform cleaning, time alignment, normalization, feature extraction, and windowing processing on the multi-source heterogeneous monitoring data to obtain the feature matrix; Step 3: Obtain the tag set of the target device and map the tag set to a tag vector. The tag set includes device category, operating condition, key components and historical fault tags. Step 4: Input the feature matrix and label vector into the label modulation deep heterogeneous cross-attention fusion network. The label deep modulation module generates multiple fusion control parameters from the label vector, including the multi-source attention temperature coefficient vector, the cross-channel interaction gating matrix, and the feature importance modulation vector. The multi-source heterogeneous cross-attention fusion module uses the fusion control parameters to perform temperature-adjusted and gated cross-attention fusion on the multi-source channels in the feature matrix and outputs the fused feature vector. Step 5: Input the component query matrix and component importance vector generated by the fusion feature vector and label deep modulation module into the label-enhanced hierarchical health decoding network. Generate variable-level health indicators, component-level health indicators and equipment comprehensive health scores through variable health decoder, component aggregation encoder and equipment comprehensive scorer respectively. Update the long-term deviation cumulative factor based on the equipment comprehensive health score. Step 6: Input the device's comprehensive health score and the deep modulation features output by the tag deep modulation module into the tag modulation deep boundary mapping evaluator. Generate a dynamic health level boundary vector based on the deep modulation features through the tag boundary generation network. Determine the health level by comparing the device's comprehensive health score with the health level boundary vector, and calculate a percentage score. Step 7: Input the device comprehensive health score sequence and long-term deviation cumulative factor sequence from the current time and multiple preceding time windows into the multi-scale temporal convolutional trend evaluation network. Extract multi-scale degradation features through convolutional branches containing different scales, and output the trend risk index and trend state category after fusion by the scale attention mechanism. Step 8: The structured output includes a health assessment report containing health level, percentage score, trend risk index, trend status category, and maintenance recommendations; when maintenance feedback data containing verified component status is received, only the subset of parameters related to misjudgment in the fusion network, hierarchical health decoding network, boundary mapping evaluator, and trend assessment network are fine-tuned and updated online.

2. The health assessment method based on the fusion of device tags and multi-source data according to claim 1, characterized in that, In step 1, the real-time monitoring data includes vibration acceleration data collected by a piezoelectric accelerometer, temperature data collected by a Pt100 resistance thermometer, current data collected by a Hall current sensor, voltage data collected by a voltage transformer, pressure data collected by a piezoresistive pressure transmitter, flow data collected by a turbine flow meter, rotational speed data collected by a photoelectric encoder, and active power data collected by a power transmitter; the business record data includes inspection records, spot inspection records, maintenance records, and work order records.

3. The health assessment method based on the fusion of device tags and multi-source data according to claim 1, characterized in that, In step 4, the tag depth modulation module consists of a linear projection layer and multiple cascaded tag residual enhancement blocks. Each enhancement block contains a main branch and residual connections, and outputs depth modulation features. A parameter decoding array maps deep modulation features in parallel to a multi-source attention temperature coefficient vector, a cross-channel interactive gating matrix, a feature importance modulation vector, a fusion layer weight synthesis vector, a component query matrix, and a component importance vector.

4. The health assessment method based on the fusion of device tags and multi-source data according to claim 3, characterized in that, In step 5, the variable health decoder includes a common feature compression layer shared by all sub-networks and multiple independent variable health decoding sub-networks. Each sub-network generates corresponding variable-level health indicators from the fused feature vector and variable embeddings through a residual fully connected structure. The component aggregation encoder uses the row vector of the component query matrix as the query and the encoded features of the variable-level health indicators as the key and value. Under the mask of the variable-component association matrix, it performs label-guided attention aggregation to obtain component-level health indicators. The device comprehensive scorer uses the component importance vector to perform a weighted harmonic average on the component-level health indicators to calculate the device comprehensive health score.

5. The health assessment method based on the fusion of device tags and multi-source data according to claim 4, characterized in that, In step 6, the label boundary generation network is a deep fully connected mapping network with skip connections. Its input is deep modulation features, and its output, after recursive multiplication and upper limit calibration, generates four monotonically decreasing boundary values ​​located in the interval [0, 1], which correspond to the boundary points of the five health levels from A to E. The percentage score is obtained by nonlinear mapping based on the relative position of the device's comprehensive health score in the corresponding level interval and the preset score range of that level.

6. The health assessment method based on the fusion of device tags and multi-source data according to claim 5, characterized in that, The multi-scale temporal convolutional trend evaluation network includes short-period branches, medium-period branches, and long-period branches. The three branches use ordinary one-dimensional convolutional kernels or dilated one-dimensional convolutional kernels of different sizes to adapt to the degradation patterns of different time spans. The scale attention mechanism automatically learns fusion weights from the global pooling features of the three branches through a gating network.

7. The health assessment method based on the fusion of device tags and multi-source data according to claim 6, characterized in that, The label-modulated deep heterogeneous cross-attention fusion network, hierarchical health decoding network, label-modulated deep boundary mapping evaluator, and multi-scale temporal convolutional trend evaluation network are jointly trained in an end-to-end manner during the offline stage. The loss function of the joint training includes device percentage score regression loss, component state classification loss, level boundary penalty loss, and trend analysis loss.

8. A health assessment system that integrates device tags and multi-source data, applied to the health assessment method that integrates device tags and multi-source data as described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to collect multi-source heterogeneous monitoring data from the target device; The data preprocessing module is used to perform cleaning, time alignment, normalization, feature extraction and windowing processing on the multi-source heterogeneous monitoring data to obtain a feature matrix; The tag acquisition module is used to acquire the tag set of the target device and map it into a tag vector. The data label fusion module is used to input the feature matrix and the label vector into the label modulation deep heterogeneous cross-attention fusion network and output the fused feature vector. The health indicator construction module is used to generate variable-level health indicators, component-level health indicators, equipment comprehensive health scores, and long-term deviation accumulation factors by utilizing the fused feature vector and label control parameters through a label-enhanced hierarchical health decoding network. The health assessment module is used to calculate the comprehensive health score and depth modulation features of the device through the tag modulation depth boundary mapping evaluator, and output the health level and percentage score. The trend analysis module is used to analyze historical health score sequences and cumulative factor sequences through a multi-scale temporal convolutional trend assessment network, and output trend risk index and trend status category. The results output and update module is used to generate and output structured health assessment reports and perform selective online fine-tuning updates to network parameters based on maintenance feedback data.

Images