A method for intelligent diagnosis of paper machine operating status by integrating multi-sensor data

Abstract

This invention relates to an intelligent diagnostic method for the operating status of a paper machine that integrates multi-sensor data, comprising the following steps: S1: Constructing a multi-sensor acquisition network for key parts of the paper machine to acquire multi-source raw data streams; S2: Establishing a unified time reference and preprocessing the multi-source raw data streams to obtain a standardized dataset; S3: Constructing a working condition identification model and, based on the standardized dataset, obtaining the working condition label corresponding to the current time window; S4: Extracting and fusing multi-dimensional state features based on the standardized dataset with working condition labels to obtain a comprehensive vector representation with working condition information; S5: Constructing an intelligent diagnostic model for the paper machine's operating status and, based on the comprehensive vector representation with working condition information, identifying the operating status and obtaining the paper machine's operating status. This invention effectively improves the efficiency and reliability of intelligent diagnostics for the operating status of paper machines.

Description

Technical Field

[0001] This invention relates to intelligent equipment diagnosis, and more particularly to an intelligent diagnostic method for the operating status of a paper machine that integrates data from multiple sensors. Background Technology

[0002] The paper machine is the core continuous equipment in a paper production line. Its operation involves multiple stages, including mechanical transmission, pulp conveying, pressing and dewatering, drying control, winding and forming, electrical drive, and paper web quality inspection. Because paper machines typically operate in a high-speed, continuous, heavy-load, and multi-variable coupled environment, changes in the status of various critical components can affect each other. Problems such as bearing wear, abnormal roller vibration, tension fluctuations, vacuum system malfunctions, unbalanced drying temperatures, abnormal motor loads, or fluctuations in paper web quality can lead to increased paper defects, paper breaks, equipment damage, or even complete line shutdowns. Therefore, timely and accurate monitoring and diagnosis of the paper machine's operating status is a crucial technical requirement for ensuring the stability of paper production and product quality.

[0003] Existing paper machine condition monitoring methods often rely on single-type sensors or local parameters for judgment, such as identifying anomalies based solely on vibration signals, temperature signals, motor current, or paper web quality detection results. While these methods can reflect the status of local equipment or single process steps, they struggle to comprehensively characterize the multi-source coupling relationships in the complex operation of the paper machine. Furthermore, differences in sampling frequency, data format, timestamp accuracy, and data quality among different sensors can easily lead to misalignment of multi-source data, inconsistent feature representation, and delayed or misjudged diagnostic results. In addition, the normal operating characteristics of a paper machine vary significantly under different paper types, speeds, basis weights, moisture content, press pressures, and drying loads. Directly performing condition diagnosis without distinguishing specific operating conditions can easily misidentify normal transitions as abnormalities. Summary of the Invention

[0004] To address the aforementioned issues, the present invention aims to provide an intelligent diagnostic method for the operating status of paper machines that integrates data from multiple sensors. This method can identify whether the paper machine is currently in a normal, abnormal, or faulty state, thereby facilitating early warning, accurate diagnosis, and predictive maintenance of the paper machine's operating status, reducing unplanned downtime, and improving the continuity of paper machine production, the stability of paper web quality, and the efficiency of equipment operation and maintenance.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for intelligent diagnosis of paper machine operating status by integrating data from multiple sensors includes the following steps: S1: Construct a multi-sensor acquisition network for key parts of the paper machine to acquire multi-source raw data streams including the mechanical status, process status, electrical status and paper web quality status of the paper machine; S2: Establish a unified time benchmark and preprocess the multi-source raw data streams to obtain a standardized dataset; S3: Construct a working condition identification model and obtain the working condition label corresponding to the current time window based on a standardized dataset; S4: Based on a standardized dataset with working condition labels, multi-dimensional state features are extracted and fused to obtain a comprehensive vector representation with working condition information. S5: Construct an intelligent diagnostic model for the paper machine's operating status, and based on a comprehensive vector representation with operating condition information, identify the operating status and obtain the paper machine's operating status.

[0006] Furthermore, a multi-sensor data acquisition network is constructed for key parts of the paper machine, specifically as follows: vibration sensors are arranged in bearing housings, press rolls, guide rolls, drying cylinders, and motor ends; temperature sensors are arranged in bearings, drying cylinders, lubrication circuits, and motor windings; pressure sensors are arranged in the press hydraulic system, steam system, and vacuum system; current, power, and torque data are collected on the main drive motor and branch drive motors; and machine speed, tension, moisture content, and paper web image data are collected along the paper web running path.

[0007] Furthermore, a unified time reference is established, and the multi-source raw data streams are preprocessed to obtain a standardized dataset. Specifically, the time reference is set using the paper machine control system clock, and the timestamps of the data collected by each sensor are uniformly corrected using a timestamp calibration method. For data streams with different sampling frequencies, they are sliced ​​according to a preset time window, and resampling compression is used to map data from different sources to the same time scale, so that the mechanical status data, process status data, electrical status data, and paper web quality status data within the same time window have a consistent time correspondence. After time synchronization is completed, the original data streams from multiple sources are cleaned, and the cleaned data is formatted and standardized to obtain a standardized dataset.

[0008] Furthermore, the working condition identification model comprises a three-tier architecture consisting of a multi-branch feature extraction network, a cross-modal attention fusion module, and a working condition classification head, with the specific structure as follows: ; in, To predict operating condition labels, Classify the working conditions. For cross-modal attention fusion modules, z mech , z proc , z elec , z qual These are feature vectors representing mechanical state, technological state, electrical state, and paper web quality state, respectively. For the t-th time window, let the modal inputs in the standardized dataset be... , where m represents the mode, m∈{mech, proc, elec, qual}, where mech represents the mechanical state data mode, proc represents the process state data mode, elec represents the electrical state data mode, and qual represents the paper web quality state data mode; For high-frequency time-series signals such as mechanical and electrical states, a one-dimensional temporal convolutional network is used to extract local temporal features. :

[0009] in, This represents a dilated causal convolution operation with a dilation rate of r. l =2 l-1 , and Here are the parameters of the l-th convolutional kernel, BN(·) is the batch normalization function, and σ(·) is the activation function. After L layers are stacked, high-frequency modal features z are obtained through global average pooling. m : ; in, This represents the local temporal features after L layers are stacked; For process status and paper web quality, a two-way gated loop unit is used to capture timing dependencies: ; in, This indicates vector concatenation; This represents the normalized data vector input to the recurrent neural network at time step t for the m-th mode; This represents the positive GRU hidden state of the m-th mode at time step t; This represents the positive GRU hidden state at the previous time step; This represents the reverse GRU hidden state of the m-th mode at time step t; This indicates the reverse GRU hidden state at the next time step; GRU(·) represents a gated loop unit used to extract the sequential dependencies between process state data and paper web quality state data. This represents a bidirectional temporal hidden feature formed by concatenating forward and reverse hidden states; A temporal self-attention mechanism is introduced to perform weighted aggregation of the hidden states at each time step:

[0010] in, and v m These are learnable attention parameters; Let represent the temporal attention weight of the m-th mode at time step t; exp(·) represents the exponential function; tanh(·) represents the hyperbolic tangent activation function; v m This represents the learnable attention score vector corresponding to the m-th modality; Indicates v m Transpose of; This represents the learnable attention weight matrix corresponding to the m-th modality; This represents the bidirectional hidden features at time step t′; ∑t′=1T in the denominator represents the normalization of the attention scores for all time steps within the current time window. The cross-modal attention fusion module integrates the feature vectors z from four modalities. m The modal feature matrix Z is stacked; a cross-modal multi-head self-attention mechanism is used to model the interrelationships between modalities. First, linear projection is used to map each modal feature into a query matrix Q, a key matrix K, and a value matrix V: ; Among them, W Q W K W V For learnable projection matrix; Multi-head attention is calculated as follows: ; Where H represents the number of attention heads, and W... O To output the projection matrix; Z represents the output of the i-th attention head; fuse The fused modal feature matrix; Softmax represents the normalized exponential function; d k Represents the key vector; Subsequently, the fused modal feature matrix Z was analyzed. fuse Weighted pooling along the modal dimension yields the comprehensive fusion vector: ; Where, w p Let β be the modality fusion weight vector. m This reflects the degree to which each mode contributes to the identification of operating conditions; The fusion vector z fused The input is a condition classification header, which consists of two fully connected layers and incorporates Dropout regularization. ; Where o represents the unnormalized score vector output by the working condition classification head; W1 and W2 represent the weight matrices of the first and second fully connected networks in the working condition classification head, respectively; b1 and b2 represent the bias vectors of the corresponding fully connected layers, respectively; σ(·) represents the nonlinear activation function; Dropout(·) represents the random deactivation regularization operation; The posterior probability of each working condition category is obtained using the Softmax function: ; Among them, o c Let X(t) represent the score corresponding to the c-th working condition category in the unnormalized score vector o; C represents the total number of preset working condition categories; c and c′ represent the working condition category indices; X(t) represents the standardized multi-source input data corresponding to the t-th time window; This indicates that under the condition of input X(t) Final output predicted load condition labels: .

[0011] Furthermore, the working condition identification model is trained using a labeled smooth cross-entropy loss function to improve the model's generalization ability and suppress overfitting. ; The smoothed label is : ; in, The labels represent the actual working conditions, and ε is the label smoothing coefficient. Introducing modal consistency regularization terms Encourage the fusion of weights β across modalities m The distribution remains stable to prevent the model from becoming overly reliant on a single modality: ; The final total loss is: Where λ is the regularization coefficient.

[0012] Furthermore, based on a standardized dataset with working condition labels, multi-dimensional state features are extracted and fused to obtain a comprehensive vector representation with working condition information, as follows: Based on a standardized dataset with operating condition labels, multi-dimensional state features are extracted from the mechanical state data, process state data, electrical state data, and paper web quality state data within each time window to form the original feature vector f for each mode. m, where m∈{mech, proc, elec, qual}, and each modal feature vector describes the state information of the paper machine at different operating levels within the current time window; After extracting the original features of each modality, a multimodal feature fusion process adaptively based on the operating condition is performed to obtain a comprehensive feature representation. The adaptive fusion process takes the operating condition label of the current time window as a conditional input and converts the operating condition label into an operating condition embedding vector e through the operating condition embedding layer. c The working condition embedding vector and the feature vectors of each mode are input together into the working condition gating fusion network for modulation; After concatenating the gated modal feature vectors, the resulting vectors are nonlinearly fused through a multi-layer feedforward network and residual connections to obtain the comprehensive fused feature vector f. fuse and the working condition embedding vector e c The data is then spliced ​​together again to form an intermediate representation that takes into account both multimodal state information and operating condition background information. ; After completing the adaptive feature fusion for operating conditions, the intermediate representation f is... mid Further mapping and compression are performed to construct the final comprehensive vector representation v with operating condition information. (t) .

[0013] Furthermore, the intelligent diagnostic model for paper machine operation status adopts a three-level diagnostic architecture consisting of an adaptive feature enhancement module, a hierarchical diagnostic decoder, and a status confidence calibration module.

[0014] Furthermore, the operating condition adaptive feature enhancement module modulates the input representation according to operating conditions, as follows: Label the working conditions The diagnostic side operating condition embedding vector is generated again through the embedding layer. Channel-level scaling and offset of the synthetic vector representation are performed through affine transformation: ; ; ; Nonlinear enhancement is then achieved through a residual feedforward network: ; To further capture local feature responses associated with specific fault modes, a channel attention mechanism is introduced to recalibrate the enhanced features: ; in , Here, r is the weighting coefficient, and r is the channel compression ratio. After the above processing, This is the diagnostic input feature vector after the working condition adaptive enhancement.

[0015] Furthermore, the hierarchical diagnostic decoder performs state category diagnosis, conditional fine-grained state diagnosis, and multi-granularity joint reasoning to obtain the final joint diagnostic output, as detailed below: The state category diagnosis inputs the enhanced feature v^(t) into the state category diagnosis head, and outputs the category probability distribution through a two-layer fully connected network: ; ; The conditional fine-grained state diagnosis uses the major category diagnosis result as a condition to predict the major category probability vector p. (1) Transformed into class conditional embedding vectors via linear projection And after being concatenated with the enhanced features, it is input into the fine-grained diagnostic head: ; To ensure that the fine-grained diagnostic results are consistent with the results of the major categories at the hierarchical level, a hierarchical consistency constraint mask M is introduced. y The mask is based on the prediction results of the major categories. Suppress output from subclasses that do not belong to this main category: ; The multi-granularity joint reasoning performs joint probability modeling on broad-category diagnosis and fine-granular diagnosis, and defines a joint diagnosis score: ; in, For the major category y (1) The corresponding set of subclasses, λ h The coefficient for hierarchical consistency reward is 1, and 1[·] is the indicator function. The final combined diagnostic output is: .

[0016] Furthermore, the state confidence calibration module introduces temperature scaling calibration to perform post-processing calibration on the diagnostic probability, resulting in a reliable confidence estimate: ; Among them, the temperature parameter T s On the validation set, optimization is achieved by minimizing the negative log-likelihood: ; Based on this, an uncertainty quantification mechanism is introduced, and MC-Dropout is used to perform K random forward inferences on the model to obtain an estimate of the cognitive uncertainty of the diagnostic results:

[0017] Where U is the prediction entropy, reflecting the uncertainty of the model's diagnosis of the current sample; when U exceeds the preset threshold U th When this happens, the sample is marked as a low-confidence sample.

[0018] The present invention has the following beneficial effects: 1. The present invention establishes a working condition identification model to obtain the working condition label corresponding to the current time window, so that subsequent state feature extraction and diagnostic judgment can be combined with the actual operating conditions of the paper machine for analysis, avoiding the misjudgment of normal working condition changes such as changes in machine speed, paper type switching, and load adjustment as abnormal conditions, thereby improving the accuracy and stability of diagnostic results. 2. This invention extracts and fuses multi-dimensional state features from a standardized dataset with working condition labels to obtain a comprehensive vector representation with working condition information. Then, it uses an intelligent diagnostic model for paper machine operating status to identify the operating status, which can fully explore the correlation between multi-source data and the law of fault evolution. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: refer to Figure 1 In this embodiment, a method for intelligent diagnosis of paper machine operating status by fusing data from multiple sensors is provided, including the following steps: S1: Construct a multi-sensor acquisition network for key parts of the paper machine to acquire multi-source raw data streams including the mechanical status, process status, electrical status and paper web quality status of the paper machine; S2: Establish a unified time benchmark and preprocess the multi-source raw data streams to obtain a standardized dataset; S3: Construct a working condition identification model and obtain the working condition label corresponding to the current time window based on a standardized dataset; S4: Based on a standardized dataset with working condition labels, multi-dimensional state features are extracted and fused to obtain a comprehensive vector representation with working condition information. S5: Construct an intelligent diagnostic model for the paper machine's operating status, and based on a comprehensive vector representation with operating condition information, identify the operating status and obtain the paper machine's operating status.

[0021] In this embodiment, a multi-sensor acquisition network is constructed for key parts of the paper machine, specifically as follows: vibration sensors are arranged at the bearing housing, press roll, guide roll, drying cylinder, and motor end; temperature sensors are arranged at the bearing, drying cylinder, lubrication oil circuit, and motor winding positions; pressure sensors are arranged in the press hydraulic system, steam system, and vacuum system; current, power, and torque data are collected on the main drive motor and branch drive motors; and machine speed, tension, moisture content, and paper web image data are collected along the paper web running path.

[0022] In this embodiment, a unified time reference is established, and the multi-source raw data streams are preprocessed to obtain a standardized dataset. Specifically, the time reference is set using the paper machine control system clock, and the timestamps of the data collected by each sensor are uniformly corrected using a timestamp calibration method. For data streams with different sampling frequencies, they are sliced ​​according to a preset time window, and resampling compression is used to map data from different sources to the same time scale, so that the mechanical status data, process status data, electrical status data, and paper web quality status data within the same time window have a consistent time correspondence. After time synchronization is completed, the multi-source raw data streams undergo data cleaning, which includes missing value handling, outlier identification, noise suppression, duplicate data removal, communication interruption data marking, and invalid data filtering. For short-term missing data, linear interpolation is used for completion. For long-term missing data or data from sensor failures, missing data is marked and its weight in subsequent fusion diagnostics is reduced. For data significantly exceeding physical or process constraints, anomaly removal or correction is performed using threshold rules. For easily disturbed signals such as vibration and current, wavelet denoising and other methods can be used to remove environmental noise and transient interference. The cleaned data is then formatted and standardized to obtain a standardized dataset. Specifically, data output from different sensors is uniformly converted to a preset data format and stored according to a unified field structure. This field structure can include time window number, sensor number, location identifier, data type, original value, processed value, unit, sampling frequency, validity identifier, and quality identifier. For data with different units and large numerical ranges, normalization is used to make different types of data comparable. For categorical information, such as paper type, production batch, equipment operating mode, alarm type, etc., encoding methods can be used to convert it into a data format that the model can recognize.

[0023] In this embodiment, the working condition recognition model includes a three-level architecture consisting of a multi-branch feature extraction network, a cross-modal attention fusion module, and a working condition classification head. The specific structure is as follows: ; in, To predict operating condition labels, Classify the working conditions. For cross-modal attention fusion modules, z mech , z proc , z elec , z qual These are feature vectors representing mechanical state, technological state, electrical state, and paper web quality state, respectively. For the t-th time window, let the modal inputs in the standardized dataset be... , where m represents the mode, m∈{mech, proc, elec, qual}, where mech represents the mechanical state data mode, proc represents the process state data mode, elec represents the electrical state data mode, and qual represents the paper web quality state data mode; For high-frequency time-series signals such as mechanical and electrical states, a one-dimensional temporal convolutional network (1D-TCN) is used to extract local temporal features. :

[0024] in, This represents a dilated causal convolution operation with a dilation rate of r. l =2 l-1 , and Here are the parameters of the l-th convolutional kernel, BN(·) is the batch normalization function, and σ(·) is the activation function. After L layers are stacked, high-frequency modal features z are obtained through global average pooling. m : ; in, This represents the local temporal features after L layers are stacked; For low-frequency and mid-frequency multivariable timing signals such as process status and paper web quality status, a bidirectional gated cyclic unit (Bi-GRU) is used to capture timing dependencies: ; in, This indicates vector concatenation; This represents the normalized data vector input to the recurrent neural network at time step t for the m-th mode; This represents the positive GRU hidden state of the m-th mode at time step t; This represents the positive GRU hidden state at the previous time step; This represents the reverse GRU hidden state of the m-th mode at time step t; This indicates the reverse GRU hidden state at the next time step; GRU(·) represents a gated loop unit used to extract the sequential dependencies between process state data and paper web quality state data. This represents a bidirectional temporal hidden feature formed by concatenating forward and reverse hidden states; A temporal self-attention mechanism is introduced to perform weighted aggregation of the hidden states at each time step:

[0025] in, and v m These are learnable attention parameters; Let represent the temporal attention weight of the m-th mode at time step t; exp(·) represents the exponential function; tanh(·) represents the hyperbolic tangent activation function; v m This represents the learnable attention score vector corresponding to the m-th modality; Indicates v m Transpose of; This represents the learnable attention weight matrix corresponding to the m-th modality; This represents the bidirectional hidden features at time step t′; ∑t′=1T in the denominator represents the normalization of the attention scores for all time steps within the current time window. The cross-modal attention fusion module integrates the feature vectors z from four modalities. m The features are stacked into a modal feature matrix Z; a cross-modal multi-head self-attention (CMSA) mechanism is used to model the inter-modal relationships: First, linear projection is used to map each modal feature into a query matrix Q, a key matrix K, and a value matrix V: ; Among them, W Q W K W V For learnable projection matrix; Multi-head attention is calculated as follows: ; Where H represents the number of attention heads, and W... O To output the projection matrix; Z represents the output of the i-th attention head; fuse The fused modal feature matrix; Softmax represents the normalized exponential function; d k Represents the key vector; Subsequently, the fused modal feature matrix Z was analyzed.fuse Weighted pooling along the modal dimension yields the comprehensive fusion vector: ; Where, w p Let β be the modality fusion weight vector. m This reflects the degree to which each mode contributes to the identification of operating conditions; The fusion vector z fused The input is a condition classification header, which consists of two fully connected layers and incorporates Dropout regularization. ; Where o represents the unnormalized score vector output by the working condition classification head; W1 and W2 represent the weight matrices of the first and second fully connected networks in the working condition classification head, respectively; b1 and b2 represent the bias vectors of the corresponding fully connected layers, respectively; σ(·) represents the nonlinear activation function; Dropout(·) represents the random deactivation regularization operation; The posterior probability of each working condition category is obtained using the Softmax function: ; Among them, o c Let X(t) represent the score corresponding to the c-th working condition category in the unnormalized score vector o; C represents the total number of preset working condition categories; c and c′ represent the working condition category indices; X(t) represents the standardized multi-source input data corresponding to the t-th time window; This indicates that under the condition of input X(t) Final output predicted load condition labels: .

[0026] In this embodiment, the working condition identification model is trained using a labeled smooth cross-entropy loss function to improve the model's generalization ability and suppress overfitting. ; The smoothed label is : ; in, The labels represent the actual working conditions, and ε is the label smoothing coefficient. Introducing modal consistency regularization terms Encourage the fusion of weights β across modalities m The distribution remains stable to prevent the model from becoming overly reliant on a single modality: ; The final total loss is: Where λ is the regularization coefficient. After the model is trained, forward inference is performed on the standardized data samples for each time window to output the corresponding working condition label. They are then written into a standardized dataset to form a standardized dataset with working condition labels.

[0027] In this embodiment, based on a standardized dataset with working condition labels, multi-dimensional state features are extracted and fused to obtain a comprehensive vector representation with working condition information, as follows: Based on a standardized dataset with working condition labels, multi-dimensional state features are extracted from the mechanical state data, process state data, electrical state data, and paper web quality state data within each time window. The feature extraction process is constrained by the working condition labels, that is, the feature extraction parameters and benchmark reference values ​​corresponding to different working conditions are used to ensure that the extracted features can accurately reflect the degree of state deviation under the current working condition, rather than the normal fluctuation differences between different working conditions. For mechanical condition data, features can be extracted from three levels: time domain, frequency domain, and time-frequency domain. Time domain features include mean, root mean square value, peak value, peak-to-peak value, waveform indicators, kurtosis, skewness, and margin indicators. Frequency domain features include fundamental frequency amplitude, harmonic amplitude, sideband energy, frequency centroid, and spectral entropy. Time-frequency domain features can be obtained by using short-time Fourier transform, continuous wavelet transform, or empirical mode decomposition to obtain the energy distribution characteristics of each frequency band. For process condition data, statistical characteristics, trend characteristics, and fluctuation characteristics of each process parameter within the current time window can be extracted, including mean, variance, rate of change, maximum deviation, sliding standard deviation, and the correlation coefficient matrix between each process parameter. For electrical condition data, the fundamental amplitude, harmonic components, power spectrum characteristics, current fluctuation indicators, and statistical characteristics of inverter operating parameters of the motor current signal can be extracted. For paper web quality condition data, the statistical distribution characteristics of quality indicators such as basis weight, moisture content, and thickness, the uniformity of web width distribution, and the type, frequency, and distribution location characteristics of surface defects can be extracted. The original feature vector f for each mode is then formed. m , where m∈{mech, proc, elec, qual}, and each modal feature vector describes the state information of the paper machine at different operating levels within the current time window; After extracting the original features of each modality, a multimodal feature fusion process adaptively based on the operating condition is performed to obtain a comprehensive feature representation. The adaptive fusion process takes the operating condition label of the current time window as a conditional input and converts the operating condition label into an operating condition embedding vector e through the operating condition embedding layer. c The working condition embedding vector and the feature vectors of each mode are input together into the working condition gating fusion network for modulation; The gated fusion network calculates the dynamic fusion gating coefficient under the operating conditions for each modal feature vector. Specifically, the gating coefficient for the m-th modality is calculated as follows:

[0028] in, and Here are the learnable parameters, and σ(·) is the Sigmoid activation function. This indicates vector concatenation; Then, gated modulation is performed on the features of each mode: , where ⊙ represents element-wise product; Through the aforementioned gating mechanism, the contribution weights of each modal feature are adaptively adjusted under different operating conditions, enabling the fusion process to highlight the modal information most relevant to the current operating state based on the operating background. After concatenating the gated modal feature vectors, they are nonlinearly fused through a multi-layer feedforward network and residual connections to obtain the comprehensive fused feature vector f. fuse and the working condition embedding vector e c The data is then spliced ​​together again to form an intermediate representation that takes into account both multimodal state information and operating condition background information. This ensures that the subsequent integrated vector representation always carries clear operating condition information; After completing the adaptive feature fusion for operating conditions, the intermediate representation f is... mid Further mapping and compression are performed to construct the final comprehensive vector representation v with operating condition information. (t) Specifically, the intermediate representations are mapped to a unified low-dimensional representation space through a projection network to obtain the comprehensive vector representation v. (t) The projection network includes a linear transformation layer, a batch normalization layer, and a nonlinear activation layer, where d v This is a pre-defined comprehensive representation dimension.

[0029] In this embodiment, a hierarchical labeling system for the paper machine's operating status is first established. The paper machine's operating status is divided into two levels: the first level is a broad category, including normal operation, performance degradation, and fault conditions; the second level is a fine-grained subcategory, further subdividing each broad category into specific status types. Let the broad category label be... Fine-grained state subclass tags are C1 and C2 represent the total number of categories in the major category and subcategory, respectively. The fine-grained condition subcategories cover typical faults and degradation states such as early bearing wear, excessive roller vibration, vacuum system leakage, drying cylinder temperature imbalance, loose drive chain, abnormal paper web tension, paper breakage warning, motor overload, and lubrication failure. The final diagnostic output is... The binary tuple enables joint state recognition at both coarse and fine granular levels.

[0030] In this embodiment, the intelligent diagnostic model for paper machine operation status adopts a three-level diagnostic architecture consisting of an adaptive feature enhancement module, a hierarchical diagnostic decoder, and a status confidence calibration module.

[0031] In this embodiment, the working condition adaptive feature enhancement module modulates the input representation according to working condition conditions, as follows: Label the working conditions The diagnostic side operating condition embedding vector is generated again through the embedding layer. Channel-level scaling and offset of the synthetic vector representation are performed through affine transformation: ; ; ; Nonlinear enhancement is then achieved through a residual feedforward network: ; To further capture local feature responses associated with specific fault modes, a channel attention mechanism is introduced to recalibrate the enhanced features: ; in , Here, r is the weighting coefficient, and r is the channel compression ratio. After the above processing, This is the diagnostic input feature vector after the working condition adaptive enhancement.

[0032] In this embodiment, the hierarchical diagnostic decoder performs state category diagnosis, conditional fine-grained state diagnosis, and multi-granularity joint reasoning to obtain the final joint diagnostic output, as detailed below: The state category diagnosis inputs the enhanced feature v^(t) into the state category diagnosis head, and outputs the category probability distribution through a two-layer fully connected network: ; ; The conditional fine-grained state diagnosis uses the major category diagnosis result as a condition to predict the major category probability vector p. (1) Transformed into class conditional embedding vectors via linear projection And after being concatenated with the enhanced features, it is input into the fine-grained diagnostic head: ; To ensure that the fine-grained diagnostic results are consistent with the results of the major categories at the hierarchical level, a hierarchical consistency constraint mask M is introduced. y The mask is based on the prediction results of the major categories. Suppress output from subclasses that do not belong to this main category: ; The multi-granularity joint reasoning performs joint probability modeling on broad-category diagnosis and fine-granular diagnosis, and defines a joint diagnosis score: ; in, For the major category y (1) The corresponding set of subclasses, λ h The coefficient for hierarchical consistency reward is 1, and 1[·] is the indicator function. The final combined diagnostic output is: .

[0033] In this embodiment, the state confidence calibration module introduces temperature scaling calibration to perform post-processing calibration on the diagnostic probability, thereby obtaining a reliable confidence estimate: ; Among them, the temperature parameter T s On the validation set, optimization is achieved by minimizing the negative log-likelihood: ; Based on this, an uncertainty quantification mechanism is introduced, and MC-Dropout is used to perform K random forward inferences on the model to obtain an estimate of the cognitive uncertainty of the diagnostic results:

[0034] Where U is the prediction entropy, reflecting the uncertainty of the model's diagnosis of the current sample; when U exceeds the preset threshold U th When this happens, the sample is marked as a low-confidence sample, triggering a further data acquisition process, thereby improving the reliability of the diagnostic system under complex operating conditions and boundary states.

[0035] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can 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.

[0036] 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.

[0037] 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.

[0038] 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 1 The steps of the function specified in one or more boxes.

[0039] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for intelligent diagnosis of paper machine operating status by integrating data from multiple sensors, characterized in that, Includes the following steps: S1: Construct a multi-sensor acquisition network for key parts of the paper machine to acquire multi-source raw data streams including the mechanical status, process status, electrical status and paper web quality status of the paper machine; S2: Establish a unified time benchmark and preprocess the multi-source raw data streams to obtain a standardized dataset; S3: Construct a working condition identification model and obtain the working condition label corresponding to the current time window based on a standardized dataset; S4: Based on a standardized dataset with working condition labels, multi-dimensional state features are extracted and fused to obtain a comprehensive vector representation with working condition information. S5: Construct an intelligent diagnostic model for the paper machine's operating status, and based on a comprehensive vector representation with operating condition information, identify the operating status and obtain the paper machine's operating status.

2. The intelligent diagnostic method for paper machine operating status fusion based on multi-sensor data according to claim 1, characterized in that, The construction of a multi-sensor acquisition network for key parts of the paper machine is specifically as follows: vibration sensors are arranged in the bearing housing, press roll, guide roll, drying cylinder, and motor end; temperature sensors are arranged in the bearing, drying cylinder, lubrication oil circuit, and motor winding positions; pressure sensors are arranged in the press hydraulic system, steam system, and vacuum system; current, power, and torque data are collected on the main drive motor and branch drive motors; and machine speed, tension, moisture content, and paper web image data are collected along the paper web running path.

3. The intelligent diagnostic method for paper machine operating status fusion based on multi-sensor data according to claim 1, characterized in that, The establishment of a unified time reference and the preprocessing of multi-source raw data streams to obtain a standardized dataset are as follows: using the paper machine control system clock as the unified time reference and a timestamp calibration method, the timestamps of the data collected by each sensor are uniformly corrected; for data streams with different sampling frequencies, they are sliced ​​according to a preset time window and resampling compression is used to map data from different sources to the same time scale, so that the mechanical status data, process status data, electrical status data and paper web quality status data within the same time window have a consistent time correspondence. After time synchronization is completed, the original data streams from multiple sources are cleaned, and the cleaned data is formatted and standardized to obtain a standardized dataset.

4. The intelligent diagnostic method for paper machine operating status fusion based on multi-sensor data according to claim 1, characterized in that, The operating condition identification model comprises a three-tier architecture consisting of a multi-branch feature extraction network, a cross-modal attention fusion module, and an operating condition classification head. The specific structure is as follows: ； in, To predict operating condition labels, Classify the working conditions. For cross-modal attention fusion modules, z mech ,z proc , z elec , z qual These are feature vectors representing mechanical state, technological state, electrical state, and paper web quality state, respectively. For the t-th time window, let the modal inputs in the standardized dataset be... , where m represents the mode, m∈{mech, proc, elec, qual}, where mech represents the mechanical state data mode, proc represents the process state data mode, elec represents the electrical state data mode, and qual represents the paper web quality state data mode; For high-frequency time-series signals such as mechanical and electrical states, a one-dimensional temporal convolutional network is used to extract local temporal features. : ； in, This represents a dilated causal convolution operation with a dilation rate of r. l =2 l-1 , and Here are the parameters of the l-th convolutional kernel, BN(·) is the batch normalization function, and σ(·) is the activation function. After L layers are stacked, high-frequency modal features z are obtained through global average pooling. m : ； in, This represents the local temporal features after L layers are stacked; For process status and paper web quality, a two-way gated loop unit is used to capture timing dependencies: ； in, This indicates vector concatenation; This represents the normalized data vector input to the recurrent neural network at time step t for the m-th mode; This represents the positive GRU hidden state of the m-th mode at time step t; This represents the positive GRU hidden state at the previous time step; This represents the reverse GRU hidden state of the m-th mode at time step t; This indicates the reverse GRU hidden state at the next time step; GRU(·) represents a gated loop unit used to extract the sequential dependencies between process state data and paper web quality state data. This represents a bidirectional temporal hidden feature formed by concatenating forward and reverse hidden states; A temporal self-attention mechanism is introduced to perform weighted aggregation of the hidden states at each time step: ； in, and v m These are learnable attention parameters; Let represent the temporal attention weight of the m-th mode at time step t; exp(·) represents the exponential function; tanh(·) represents the hyperbolic tangent activation function; v m This represents the learnable attention score vector corresponding to the m-th modality; Indicates v m Transpose of; This represents the learnable attention weight matrix corresponding to the m-th modality; This represents the bidirectional hidden features at time step t′; ∑t′=1T in the denominator represents the normalization of the attention scores for all time steps within the current time window. The cross-modal attention fusion module integrates the feature vectors z from four modalities. m The modal feature matrix Z is stacked; a cross-modal multi-head self-attention mechanism is used to model the interrelationships between modalities. First, linear projection is used to map each modal feature into a query matrix Q, a key matrix K, and a value matrix V: ； Among them, W Q W K W V For learnable projection matrix; Multi-head attention is calculated as follows: ; Where H represents the number of attention heads, and W... O To output the projection matrix; Z represents the output of the i-th attention head; fuse The fused modal feature matrix; Softmax represents the normalized exponential function; d k Represents the key vector; Subsequently, the fused modal feature matrix Z was analyzed. fuse Weighted pooling along the modal dimension yields the comprehensive fusion vector: ; Where, w p Let β be the modality fusion weight vector. m This reflects the degree to which each mode contributes to the identification of operating conditions; The fusion vector z fused The input is a condition classification header, which consists of two fully connected layers and incorporates Dropout regularization. ; Where o represents the unnormalized score vector output by the working condition classification head; W1 and W2 represent the weight matrices of the first and second fully connected networks in the working condition classification head, respectively; b1 and b2 represent the bias vectors of the corresponding fully connected layers, respectively; σ(·) represents the nonlinear activation function; Dropout(·) represents the random deactivation regularization operation; The posterior probability of each working condition category is obtained using the Softmax function: ; Among them, o c Let X(t) represent the score corresponding to the c-th working condition category in the unnormalized score vector o; C represents the total number of preset working condition categories; c and c′ represent the working condition category indices; X(t) represents the standardized multi-source input data corresponding to the t-th time window; This indicates that under the condition of input X(t) Final output predicted load condition labels: 。 5. The intelligent diagnostic method for paper machine operating status fusion based on multi-sensor data according to claim 4, characterized in that, The operating condition identification model is trained using a labeled smooth cross-entropy loss function to improve the model's generalization ability and suppress overfitting. ； The smoothed label is : ； in, The labels represent the actual working conditions, and ε is the label smoothing coefficient. Introducing modal consistency regularization terms Encourage the fusion of weights β across modalities m The distribution remains stable to prevent the model from becoming overly reliant on a single modality: ； The final total loss is: Where λ is the regularization coefficient.

6. The intelligent diagnostic method for paper machine operating status fusion based on multi-sensor data according to claim 1, characterized in that, The standardized dataset with working condition labels is used to extract and fuse multi-dimensional state features to obtain a comprehensive vector representation with working condition information, as detailed below: Based on a standardized dataset with operating condition labels, multi-dimensional state features are extracted from the mechanical state data, process state data, electrical state data, and paper web quality state data within each time window to form the original feature vector f for each mode. m , where m∈{mech, proc, elec, qual}, and each modal feature vector describes the state information of the paper machine at different operating levels within the current time window; After extracting the original features of each modality, a multimodal feature fusion process adaptively based on the operating condition is performed to obtain a comprehensive feature representation. The adaptive fusion process takes the operating condition label of the current time window as a conditional input and converts the operating condition label into an operating condition embedding vector e through the operating condition embedding layer. c The working condition embedding vector and the feature vectors of each mode are input together into the working condition gating fusion network for modulation; After concatenating the gated modal feature vectors, the resulting vectors are nonlinearly fused through a multi-layer feedforward network and residual connections to obtain the comprehensive fused feature vector f. fuse and the working condition embedding vector e c The data is then spliced ​​together again to form an intermediate representation that takes into account both multimodal state information and operating condition background information. ; After completing the adaptive feature fusion for operating conditions, the intermediate representation f is... mid Further mapping and compression are performed to construct the final comprehensive vector representation v with operating condition information. (t) .

7. The intelligent diagnostic method for paper machine operating status fusion based on multi-sensor data according to claim 6, characterized in that, The intelligent diagnostic model for the paper machine's operating status adopts a three-level diagnostic architecture consisting of an adaptive feature enhancement module, a hierarchical diagnostic decoder, and a status confidence calibration module.

8. The intelligent diagnostic method for paper machine operating status fusion based on multi-sensor data according to claim 7, characterized in that, The adaptive feature enhancement module modulates the input representation according to operating conditions, as follows: Label the working conditions The diagnostic side operating condition embedding vector is generated again through the embedding layer. Channel-level scaling and offset of the synthetic vector representation are performed through affine transformation: ; ; ; Nonlinear enhancement is then achieved through a residual feedforward network: ; To further capture local feature responses associated with specific fault modes, a channel attention mechanism is introduced to recalibrate the enhanced features: ; in , Here, r is the weighting coefficient, and r is the channel compression ratio. After the above processing, This is the diagnostic input feature vector after the working condition adaptive enhancement.

9. The intelligent diagnostic method for paper machine operating status fusion based on multi-sensor data according to claim 8, characterized in that, The hierarchical diagnostic decoder performs state category diagnosis, conditional fine-grained state diagnosis, and multi-granularity joint reasoning to obtain the final joint diagnostic output, as detailed below: The state category diagnosis inputs the enhanced feature v^(t) into the state category diagnosis head, and outputs the category probability distribution through a two-layer fully connected network: ； ； The conditional fine-grained state diagnosis uses the major category diagnosis result as a condition to predict the major category probability vector p. (1) Transformed into class conditional embedding vectors via linear projection And after being concatenated with the enhanced features, it is input into the fine-grained diagnostic head: ； To ensure that the fine-grained diagnostic results are consistent with the results of the major categories at the hierarchical level, a hierarchical consistency constraint mask M is introduced. y The mask is based on the prediction results of the major categories. Suppress output from subclasses that do not belong to this main category: ； The multi-granularity joint reasoning performs joint probability modeling on broad-category diagnosis and fine-granular diagnosis, and defines a joint diagnosis score: ； in, For the major category y (1) The corresponding set of subclasses, λ h The coefficient for hierarchical consistency reward is 1, and 1[·] is the indicator function. The final combined diagnostic output is: 。 10. The intelligent diagnostic method for paper machine operating status fusion based on multi-sensor data according to claim 9, characterized in that, The state confidence calibration module incorporates temperature scaling calibration to perform post-processing calibration on the diagnostic probability, resulting in a reliable confidence estimate. ； Among them, the temperature parameter T s On the validation set, optimization is achieved by minimizing the negative log-likelihood: ； Based on this, an uncertainty quantification mechanism is introduced, and MC-Dropout is used to perform K random forward inferences on the model to obtain an estimate of the cognitive uncertainty of the diagnostic results: ； Where U is the prediction entropy, reflecting the uncertainty of the model's diagnosis of the current sample; when U exceeds the preset threshold U th When this happens, the sample is marked as a low-confidence sample.

Images