Intelligent factory fault diagnosis method and system based on AI prediction model

Abstract

The invention provides an intelligent factory fault diagnosis method and system based on an AI prediction model, and the method comprises the steps: obtaining an equipment monitoring data flow of a target production line of an intelligent factory, carrying out the diagnosis feature construction processing of the equipment monitoring data flow, generating a state evolution feature and a component correlation feature, and carrying out the fault diagnosis of the target production line of the intelligent factory; and inputting the state evolution characteristics and the component association characteristics into a pre-trained fault prediction model for fault prediction, and generating diagnosis result data containing fault risk levels. And the potential fault type and the propagation characteristic information are identified according to the diagnosis result data, and finally the maintenance guidance data containing the fault positioning identifier are generated based on the potential fault type and the propagation characteristic information and are transmitted to the factory operation and maintenance system to trigger the fault intervention operation, so that the accuracy and the maintenance efficiency of intelligent factory fault diagnosis are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and more particularly, to an intelligent factory fault diagnosis method and system based on an AI prediction model. Background Art

[0002] In the production operation of an intelligent factory, the stable operation of the production line is crucial for ensuring product quality, improving production efficiency, and reducing production costs. With the continuous improvement of the automation level of intelligent factories, the number and complexity of devices on the production line are increasing day by day, and the occurrence frequency and influence range of device failures are also expanding accordingly.

[0003] Currently, the commonly used fault diagnosis methods in intelligent factories mainly rely on manual experience judgment and simple rule-based monitoring systems. Although manual experience judgment can quickly locate some common faults based on the experience of professionals, this method is not only inefficient but also easily affected by personal subjective factors, making it difficult to comprehensively and accurately identify all potential faults. The simple rule-based monitoring system has the problem of fixed rule settings and lack of flexibility, and cannot adapt to the complex operating environment and diverse fault modes of the production line, and it is often difficult to effectively identify some new or complex faults. Therefore, there is an urgent need for a more intelligent and efficient fault diagnosis method that can accurately predict and diagnose the faults of the intelligent factory production line, take preventive measures in advance, and ensure the stable operation of the production line. Summary of the Invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, embodiments of the present invention provide an intelligent factory fault diagnosis method based on an AI prediction model, the method comprising: Obtain the device monitoring data stream of the target production line in the intelligent factory, the device monitoring data stream comprising multiple groups of status data segments with timestamp identifiers collected continuously; Perform diagnostic feature construction processing on the device monitoring data stream to generate a state evolution feature reflecting the device operation state and a component association feature reflecting the interaction relationship between device components; Input the state evolution feature and the component association feature into a pre-trained fault prediction model for fault prediction to generate diagnostic result data including a fault risk level; Identify the potential fault types currently existing in the target production line and the propagation feature information of the faults during the device operation according to the diagnostic result data; Generate maintenance guidance data including fault location identifiers based on the potential fault types and the propagation feature information, and transmit the maintenance guidance data to the factory operation and maintenance system to trigger fault intervention operations.

[0005] In another aspect, an embodiment of the present invention further provides an intelligent factory fault diagnosis system based on an AI prediction model, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0006] Based on the above aspects, the embodiment of the present invention obtains the device monitoring data stream of the target production line in the intelligent factory, and performs diagnostic feature construction processing on it to generate a state evolution feature reflecting the device operation state and a component association feature reflecting the interaction relationship between device components, effectively improving the accuracy of fault diagnosis. Inputting the constructed state evolution feature and component association feature into a pre-trained fault prediction model for fault prediction can make full use of the powerful learning and analysis capabilities of the AI model to generate diagnostic result data including the fault risk level, realizing accurate prediction and risk assessment of faults. Identifying the current potential fault types and the propagation feature information of the faults during the device operation based on the diagnostic result data helps to deeply understand the essence and development trend of the faults. Generating maintenance guidance data including fault location identifiers based on the potential fault types and propagation feature information and transmitting it to the factory operation and maintenance system to trigger fault intervention operations realizes the timely discovery, accurate diagnosis and effective intervention of faults, greatly improving the operation stability and production efficiency of the intelligent factory production line and reducing the losses caused by faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 is a schematic execution flow diagram of the intelligent factory fault diagnosis method based on the AI prediction model provided by the embodiment of the present invention.

[0008] Figure 2 is a schematic diagram of exemplary hardware and software components of the intelligent factory fault diagnosis system based on the AI prediction model provided by the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0009] The present invention will be specifically described below with reference to the accompanying drawings of the specification. Figure 1 is a schematic flow diagram of the intelligent factory fault diagnosis method based on the AI prediction model provided by an embodiment of the present invention. The intelligent factory fault diagnosis method based on the AI prediction model will be introduced in detail below.

[0010] Step S110: Obtain the device monitoring data stream of the target production line in the intelligent factory, where the device monitoring data stream includes multiple groups of state data segments continuously collected with timestamp identifiers.

[0011] Exemplarily, in the smart factory for papermaking felt production, the target production line includes a variety of key equipment, such as looms, setting machines, hot presses, etc. To accurately monitor the operating status of the equipment, sensors with different functions are installed on various equipment. A rotational speed sensor is installed on the loom to monitor the rotational speed of the loom's main shaft, which is directly related to the weaving efficiency of the felt; a tension sensor can detect the tension of the warp and weft during the weaving process in real time, and the stability of the tension is crucial for the flatness of the felt; a vibration sensor can sense the vibration of the loom during operation, and abnormal vibration may indicate wear or looseness of mechanical components. The setting machine is equipped with a temperature sensor to monitor the temperature during the setting process, and the appropriate temperature is a key factor to ensure the setting quality of the felt; a pressure sensor can measure the pressure applied during setting to ensure that the felt meets the specified physical properties. In addition to temperature and pressure sensors on the hot press, a displacement sensor can monitor the displacement change of the pressing plate during the hot pressing process, which is very important for controlling the thickness and density of the felt.

[0012] In this embodiment, the above sensors continuously collect equipment operation data according to their respective set collection frequencies, and each collected equipment operation data can be correspondingly attached with an accurate timestamp identifier. For example, the rotational speed sensor collects the loom rotational speed data every set short time interval and marks the collection moment for each data point. Through wired or wireless network transmission methods, the timestamped status data from different sensors are aggregated and integrated to form an equipment monitoring data stream, which contains multiple continuously collected status data segments with timestamp identifiers, thereby recording the operating status of each equipment on the target production line at different times.

[0013] Step S120: Perform diagnostic feature construction processing on the equipment monitoring data stream to generate a state evolution feature reflecting the equipment operating status and a component association feature reflecting the interaction relationship between equipment components.

[0014] In order to accurately diagnose possible faults in the papermaking felt production equipment, it is necessary to deeply process the obtained equipment monitoring data stream and construct features that can reflect the equipment operating status and component interaction relationship.

[0015] Step S121: Perform data alignment processing on the equipment monitoring data stream to uniformly map the status data segments with different collection frequencies to the standard time axis and generate a data set after time alignment.

[0016] During the papermaking felt production process, there are differences in the collection frequencies of different sensors. For example, the tension sensor may collect the tension data of the loom warp at a higher frequency, while the temperature sensor may collect the temperature data of the setting machine at a lower frequency. This frequency difference results in inconsistent status data in terms of time, which is not conducive to subsequent analysis.

[0017] To solve this problem, a standard time interval needs to be determined as a reference. For data with a high acquisition frequency, based on the time points of the standard time axis, only the status data corresponding to the standard time points are retained; for data with a low acquisition frequency, the interpolation method is used to supplement the missing data points. For example, when there are missing temperature data collected by a temperature sensor on the standard time axis, the temperature values of the missing points are calculated by linear interpolation according to the previously and subsequently collected data points. Taking the rotational speed and tension data of a loom as an example, the values corresponding to the standard time points are selected from the rotational speed data, and the missing values are supplemented by interpolation for the tension data, so as to unify the data with different acquisition frequencies onto the standard time axis, generate a data set after time alignment, make all status data consistent in time, and facilitate subsequent feature extraction and analysis.

[0018] Step S122: Perform state evolution extraction processing on the data set after time alignment, analyze the change trend of consecutive state data segments through a sliding time window, and generate state evolution features including state fluctuation patterns and periodic anomaly identifiers.

[0019] Step S1221: Set a sliding time window with a fixed length to segment the data set after time alignment, and obtain multiple data window units with time continuity.

[0020] In the production scenario of a papermaking wire blanket, the fixed length of the sliding time window needs to be set according to the operating characteristics of the equipment and the data change rules. Taking a loom as an example, its operating state changes rapidly. For example, the rotational speed of the loom may change due to different weaving process requirements in a short period of time, so the window length can be set shorter to capture its state changes more precisely. While the operation of a calender is relatively stable, the temperature and pressure fluctuate less within a certain period of time, and the window length can be set longer. Slide the set fixed-length window from left to right on the data set after time alignment, and intercept the data within one window each time it slides to form a data window unit. These data window units are continuous in time, and each data window unit covers the status data within a specific time period. For example, for the rotational speed data of a loom, a series of data window units containing rotational speed information for different time periods can be obtained through the sliding window.

[0021] Step S1222: Calculate the mean, variance, and extreme point distribution of the state data segment within each data window unit to generate basic statistical features reflecting state stability.

[0022] A detailed statistical analysis is performed on the status data segments within each data window unit. To calculate the mean, the values of all status data within that data window unit are summed and then divided by the number of data points. The result is the mean. The mean reflects the average operating status of the device within that window period. For example, the mean speed of a loom within a data window unit can reflect the average operating speed of the loom during that period. To calculate the variance, the square of the difference between each data point and the mean is calculated, and then the average of these squared values is taken. The variance measures the dispersion of the data relative to the mean. A larger variance indicates more significant fluctuations in the device status during that period. For example, a large variance in the loom speed within a data window unit indicates significant speed fluctuations and unstable operating conditions. Furthermore, extreme values are identified in the data, including maximum and minimum values. The distribution of extreme values can reflect the extreme operating conditions of the device during that period. For example, the maximum and minimum speed values of a loom within a data window unit can reveal the range of speed fluctuations within that period. These mean, variance, and extreme point distributions constitute the basic statistical characteristics that reflect the stability of the equipment state. By analyzing these characteristics, we can preliminarily determine whether the equipment is operating stably within each time window.

[0023] Step S1223: Analyze the variation of basic statistical features between adjacent data window units, and extract trend parameters of the state data segment in the continuous time window, wherein the trend parameters include an upward trend parameter, a downward trend parameter, and a trend persistence index.

[0024] Compare adjacent data window units and analyze the changes in their basic statistical characteristics. For the mean, if the mean of the subsequent data window unit is greater than that of the previous data window unit, it indicates that the device status is showing an upward trend. To calculate the upward trend parameter, subtract the mean of the previous window from the mean of the subsequent window, and then divide by the time interval to represent the magnitude and speed of the increase. Conversely, if the mean of the subsequent window is less than the mean of the previous window, it indicates a downward trend. The downward trend parameter is calculated similarly. Furthermore, by observing the persistence of trends within multiple consecutive windows, a trend persistence index is determined. If the device status maintains an upward or downward trend over multiple consecutive windows, the trend is highly persistent, and the trend persistence index is high. Conversely, if the trend changes frequently, the trend persistence index is low. For example, for loom speed data, if the mean speed continuously increases over multiple consecutive data window units, it indicates that the loom speed is on an upward trend with strong persistence, and the trend persistence index is high. These trend parameters can be used to describe the dynamic changes in device status within consecutive time windows.

[0025] Step S1224: Perform periodic pattern detection processing on the data window unit, identify the recurring fluctuation periods in the state data segment through autocorrelation analysis, and generate a period length parameter and a period stability score.

[0026] Use the autocorrelation analysis method to perform periodic pattern detection on the state data within each data window unit. Autocorrelation analysis finds the periodic patterns in the data by calculating the correlation between the data and itself at different time delays. Specifically, for the state data within a data window unit, calculate the correlation between the data and itself at different time delays to obtain a series of correlation coefficients. When the correlation coefficient at a certain time delay is large, it indicates that there are similar fluctuation situations in the data after this time interval, that is, there are periodic fluctuations. When periodic fluctuations are detected, determine its period length and generate a period length parameter, which represents the time length of the fluctuation period. At the same time, generate a period stability score according to the stability of the period in multiple data window units. If the period length is relatively stable and the fluctuation is small in different windows, it means that the period stability is high and the period stability score is also high; on the contrary, if the period length changes greatly, the period stability is low and the period stability score is also low. For example, for the rotational speed data of a loom, through autocorrelation analysis, it is found that there are similar fluctuation situations after a set time interval, determine this time interval as the period length, and then give a period stability score according to the stability of this period length in multiple data window units.

[0027] Step S1225: Integrate the basic statistical features, trend parameters, period length parameters, and period stability scores to obtain a state evolution feature including a state fluctuation pattern and a periodic anomaly identifier.

[0028] Integrate the previously calculated basic statistical features (mean, variance, extreme point distribution), trend parameters (upward trend parameter, downward trend parameter, trend persistence index), period length parameters, and period stability scores. The integration method can be to arrange and combine these features in a set order to form a multi-dimensional feature vector. These different types of features reflect the changes in the device state from different angles. Combining them together forms a comprehensive state evolution feature. By analyzing this state evolution feature, the fluctuation pattern of the device state can be identified, such as stable fluctuation, severe fluctuation, or periodic fluctuation, etc. At the same time, if it is found that the period stability score is low or the trend persistence index is abnormal, etc., these can be used as periodic anomaly identifiers, which help to timely discover potential problems in the device operation. For example, if the period stability score in the state evolution feature of a loom is low, it indicates that the operating cycle of the loom is unstable and there may be potential failure risks.

[0029] Step S123: Perform component association extraction processing on the time-aligned data set, analyze the co-variation law of the corresponding state data segments of different components based on the device physical connection relationship, and generate component association features including interaction intensity parameters and association stability indicators.

[0030] Step S1231: Determine key component pairs according to the device connection map of the target production line, and extract the time series of the corresponding state data segments of each pair of key components.

[0031] In the target production line of papermaking felt production, there are specific physical connection relationships between the components of different devices. By referring to the device connection map, determine the key component pairs that are crucial for the production process. For example, the main shaft of the loom and the weft feeding device are a pair of key components, and their coordinated work directly affects the weaving quality of the felt. From the time-aligned data set, extract the time series of the corresponding state data segments of each pair of key components. For the pair of components of the loom main shaft and the weft feeding device, extract the rotational speed time series of the main shaft and the feeding speed time series of the weft feeding device respectively. These time series record the operating states of the components at different times. When extracting the time series, ensure the integrity and accuracy of the data, and avoid data loss or errors.

[0032] Step S1232: Calculate the cross-correlation coefficient of the time series of each pair of key components, and generate an interaction intensity parameter reflecting the degree of co-variation between the components.

[0033] For the time series of each pair of key components, calculate the cross-correlation coefficient between them. The cross-correlation coefficient measures the linear correlation between two time series, and its value range is between -1 and 1. When calculating the cross-correlation coefficient, first perform centering processing on the two time series, that is, subtract their respective means, then calculate the sum of the products of the data at the corresponding moments, and then divide by the product of the standard deviations of the two time series. If the cross-correlation coefficient is close to 1, it means that the state data segments of the two components show a strong positive correlation, that is, their operating state change trends are very similar and the degree of co-variation is high; if the cross-correlation coefficient is close to -1, it means a strong negative correlation; if the cross-correlation coefficient is close to 0, it means that the correlation between the state data segments of the two components is weak and the degree of co-variation is low. By calculating the cross-correlation coefficient, generate an interaction intensity parameter reflecting the degree of co-variation between the components, and this interaction intensity parameter can characterize the mutual influence degree between different components. For example, if the cross-correlation coefficient between the rotational speed time series of the loom main shaft and the feeding speed time series of the weft feeding device is close to 1, it indicates that the operating state change trends of these two components are highly consistent and the degree of co-variation is high.

[0034] Step S1233: Analyze the fluctuation range of the cross-correlation coefficient within consecutive time windows, and generate an association stability index reflecting the persistence of the component association relationship.

[0035] Observe the variation of the cross-correlation coefficient within consecutive time windows and analyze its fluctuation range. The maximum and minimum values of the cross-correlation coefficient within multiple consecutive data window units can be calculated, and the difference between the two is the fluctuation range. If the cross-correlation coefficient is relatively stable within multiple consecutive time windows and the fluctuation range is small, it indicates that the association relationship between components is relatively stable, and the generated association stability index is high; conversely, if the cross-correlation coefficient fluctuates greatly, it indicates that the association relationship between components is unstable, and the association stability index is low. For example, for the main shaft and the weft feeding device of a loom, if their cross-correlation coefficient remains within a small fluctuation range within multiple consecutive time windows, it indicates that their cooperative working relationship is relatively stable and the association stability index is high; if the cross-correlation coefficient changes frequently and the fluctuation range is large, it indicates that their cooperative working relationship is unstable and the association stability index is low.

[0036] Step S1234: Perform causal relationship detection processing on the state data segments of the key component pairs, determine the causal influence direction between components through Granger causality test, and generate a causal relationship identifier.

[0037] Use the Granger causality test method to detect the causal relationship of the state data segments of the key component pairs. The basic idea of the Granger causality test is that if a time series (such as the state data segment of component A) can help predict another time series (such as the state data segment of component B), but not vice versa, then it can be considered that component A is the Granger cause of the change of component B. Specifically, when operating, two regression models are established. One model only contains the lag terms of the predicted time series, and the other model contains the lag terms of the predicted time series and the lag terms of the predictive time series. By comparing the prediction effects of the two models, it is judged whether there is a Granger causal relationship. Through the above test method, the causal influence direction between components is determined, and a causal relationship identifier is generated. The causal relationship identifier can clarify the influence relationship between different components and help to deeply understand the interaction mechanism between the device components. For example, for the main shaft speed and the weft feeding speed of a loom, if the change of the main shaft speed can predict the change of the weft feeding speed to a certain extent, but not vice versa, then it can be considered that the main shaft speed is the Granger cause of the change of the weft feeding speed, and the corresponding causal relationship identifier is generated.

[0038] Step S1235: Integrate the interaction intensity parameter, the association stability index, and the causal relationship identifier to obtain component association characteristics including the interaction intensity parameter and the association stability index.

[0039] In this embodiment, the integration method may be to combine the interaction intensity parameter, the association stability index, and the causal relationship identifier into a multi-dimensional feature vector according to a set rule. The interaction intensity parameter, the association stability index, and the causal relationship identifier describe the interaction relationship between device components from different aspects. Combining them together forms a comprehensive component association feature. This component association feature can comprehensively reflect the co-variation law, association stability, and causal influence relationship between different components. For example, by integrating the interaction intensity parameter, the association stability index, and the causal relationship identifier of the loom main shaft and the weft feeding device, the obtained component association feature can show the interaction situation and causal influence direction between these two components.

[0040] Step S124: Input the state evolution feature and the component association feature into a feature calibration module for dimension matching processing to eliminate the scale difference between different features and generate a calibrated feature set with a unified representation form.

[0041] Since the state evolution feature and the component association feature may have different dimensions and scales, directly using them for subsequent fault prediction may lead to inaccurate results. Therefore, these two features are input into the feature calibration module for processing. The feature calibration module first analyzes the dimensions of the state evolution feature and the component association feature. If the dimensions of the two features are different, dimension adjustment may be required, such as through feature selection or feature extraction methods, to make their dimensions match each other. At the same time, methods such as standardization or normalization are used to eliminate the scale difference between different features. Standardization can be carried out by subtracting the mean and dividing by the standard deviation, converting the feature data into data with a mean of 0 and a standard deviation of 1. Normalization can map the feature data to a set interval, such as [0, 1]. For example, for some parameters in the state evolution feature, the value range may be relatively large, while for some parameters in the component association feature, the value range is relatively small. Through the processing of the calibration module, their value ranges are unified to a similar scale. After the above processing, a calibrated feature set with a unified representation form is generated, enabling different features to play a more effective role in the subsequent fault prediction model.

[0042] Step S125: Dynamically adjust the feature weight coefficients of the state evolution feature and the component association feature in fault prediction according to the evaluation results of the influence degree of each dimension feature in the calibrated feature set on fault diagnosis.

[0043] Evaluate the influence degree of the features in each dimension of the calibration feature set. For example, some machine learning methods, such as algorithms like random forest and gradient boosting, can be used to calculate the importance of each dimension feature for fault diagnosis, so that an importance score can be assigned to each feature according to its performance during the model training process. For example, in the random forest algorithm, the importance score of a feature can be determined by calculating the information gain brought by this feature during the decision tree splitting process. According to the evaluation results, dynamically adjust the feature weight coefficients of the state evolution features and component association features in fault prediction. If some dimension features in the state evolution features have a greater impact on fault diagnosis, then correspondingly increase the weight coefficient of the state evolution features; if some dimension features in the component association features are more important, then increase the weight coefficient of the component association features. Such dynamic adjustment can make the fault prediction model pay more attention to the features that have an important impact on fault diagnosis and improve the accuracy of fault prediction. For example, if the evaluation results show that the rotational speed fluctuation feature of the loom has a greater impact on fault diagnosis, while the temperature stability feature of the shaping machine has a relatively small impact, then the weight coefficient of the dimension related to the rotational speed fluctuation of the loom in the state evolution features can be increased, and the weight coefficient of the dimension related to the temperature stability of the shaping machine can be decreased.

[0044] Step S130: Input the state evolution features and the component association features into a pre-trained fault prediction model for fault prediction, and generate diagnostic result data including the fault risk level.

[0045] Step S131: Input the state evolution features and the component association features into the feature fusion layer of the fault prediction model, and perform cross-dimensional information fusion processing in combination with the feature weight coefficients to generate a fused feature vector.

[0046] Input the processed and weight-adjusted state evolution features and component association features into the feature fusion layer of the fault prediction model. In the feature fusion layer, according to the previously determined feature weight coefficients, perform cross-dimensional information fusion processing on the state evolution features and component association features. For example, for a certain dimension feature in the state evolution features and a certain dimension feature in the component association features, perform weighted combination according to the corresponding weight coefficients. In the above way, fuse the feature information of different types and different dimensions together to generate a fused feature vector, which comprehensively combines the information of the device operation state and component interaction relationship.

[0047] Step S132: Perform time-dependence modeling processing on the fused feature vector through the time series encoder of the fault prediction model to extract the time series context features reflecting the fault development process.

[0048] Step S1321: Input the fused feature vector into the bidirectional recurrent neural network layer of the temporal encoder, and perform sequence processing on the fused feature vector respectively in the forward time order and the backward time order to generate a bidirectional hidden state containing past state information and future state information.

[0049] In the scenario of fault prediction for papermaking wire blanket production equipment, input the fused feature vector into the bidirectional recurrent neural network (BRNN) layer. The bidirectional recurrent neural network consists of a forward recurrent neural network (RNN) and a backward recurrent neural network. The forward RNN starts from the starting point of the time series and processes the fused feature vector in sequence according to the forward time order. At each time step, the forward RNN calculates the hidden state of the current time step based on the input at the current time step (i.e., the value of the fused feature vector at this time step) and the hidden state of the previous time step. This hidden state contains information about the change in the equipment state from the starting time to the current time, that is, past state information. For example, for the operating state of a loom, the forward RNN can capture the cumulative changes in the loom speed, tension, etc. from startup to the current moment.

[0050] The backward RNN starts from the end point of the time series and processes the fused feature vector in the backward time order. Similarly, at each time step, the backward RNN calculates the hidden state of the current time step based on the input at the current time step and the hidden state of the next time step. This hidden state contains information about the possible change in the equipment state from the current time to a future moment, that is, future state information. For example, for the possible future fault trend of a loom, the backward RNN can infer the possible development situation from the current moment forward.

[0051] Through the parallel processing of the forward and backward RNNs, a bidirectional hidden state containing past state information and future state information is finally generated. This step makes full use of the front and back correlation information of the time series, laying a foundation for accurately capturing the fault development process in the follow-up.

[0052] Step S1322: Perform attention weight calculation processing on the bidirectional hidden state, and assign attention weights reflecting its importance for fault diagnosis to the hidden state of each time step through the time attention mechanism.

[0053] The purpose of the temporal attention mechanism is to highlight the time-step information that has an important impact on fault diagnosis. For each time-step hidden state in the bidirectional hidden state, it is first input into a fully connected layer for linear transformation to obtain a new feature vector. Then, this new feature vector is non-linearly transformed through an activation function (such as the tanh function) to make its value fall within a suitable range. Next, the transformed feature vector is dot-producted with a pre-trained attention weight vector to obtain a scalar value, which represents the importance score of this time-step hidden state for fault diagnosis.

[0054] Normalize the importance scores of all time steps. For example, use the softmax function to convert these scores into a probability distribution, that is, the attention weights. The value range of the attention weights is between 0 and 1, and the sum of the attention weights of all time steps is 1. In this way, each time-step hidden state is assigned an attention weight that reflects its importance for fault diagnosis. For example, in the production process of a papermaking felt, if the rotational speed of the loom suddenly shows abnormal fluctuations at a certain time step, then the attention weight corresponding to the hidden state of this time step will be relatively high because it is more important for fault diagnosis.

[0055] Step S1323: Perform a weighted sum operation on the bidirectional hidden state according to the attention weights to generate a temporal aggregation feature containing key time-step information.

[0056] After obtaining the attention weights of each time-step hidden state, multiply each time-step hidden state in the bidirectional hidden state by the corresponding attention weight. Then, sum up the weighted hidden states of all time steps to obtain a comprehensive feature vector, that is, a temporal aggregation feature containing key time-step information. Through the above weighted sum method, the time-step information that is important for fault diagnosis is highlighted, and some relatively unimportant time-step information is filtered out. For example, in the operating data of the loom, for the hidden states of the stable operating time periods that have less impact on fault diagnosis, due to their low attention weights, they contribute less to the final temporal aggregation feature; while for the hidden states of the key time steps with abnormal fluctuations, due to their large attention weights, they will occupy an important position in the temporal aggregation feature.

[0057] Step S1324: Input the temporal aggregation feature into the gated recurrent unit layer of the temporal encoder for state update processing, filter out redundant historical state information and retain key fault evolution clues, and generate a temporal context feature reflecting the fault development process.

[0058] The gated recurrent unit (GRU) layer consists of a reset gate and an update gate. When the time-series aggregated features are input into the GRU layer, the value of the reset gate is calculated first. The role of the reset gate is to determine whether to ignore the previous hidden state information. It is obtained by performing a linear transformation and sigmoid activation function processing on the current input time-series aggregated features and the hidden state of the previous time step to get a value between 0 and 1. If the value of the reset gate is close to 1, it means to retain the previous hidden state information; if it is close to 0, it means to ignore the previous hidden state information.

[0059] Next, the value of the update gate is calculated. The update gate is used to determine how much new information to add to the hidden state and how much of the previous hidden state information to retain. It is also obtained by performing a linear transformation and sigmoid activation function processing on the current input and the hidden state of the previous time step to get a value. According to the values of the reset gate and the update gate, the hidden state of the current time step is updated.

[0060] In this process, the GRU layer will automatically filter out the redundant historical state information that is not helpful for the current fault diagnosis and only retain the key fault evolution clues. For example, in the operation data of the papermaking felt production equipment, for some early equipment state information that has returned to normal, the GRU layer will filter it out through the gating mechanism; while for the key information related to the occurrence and development of faults, such as the abnormal change trend of the loom speed, the continuous instability of the tension, etc., will be retained and updated to the hidden state. Finally, after the state update processing of the GRU layer, the time-series context features reflecting the fault development process are generated.

[0061] Step S133: Use the association decoder of the fault prediction model to perform component interaction relationship parsing processing on the time-series context features to generate association context features reflecting the fault propagation path.

[0062] Step S1331: Input the time-series context features into the graph convolutional network layer of the association decoder, construct an adjacency matrix with the equipment connection relationship of the target production line as the graph structure, and extract the local interaction features between components through graph convolutional operations.

[0063] In the target production line of papermaking felt production, there are specific physical connection relationships between different equipment components. Based on these equipment connection relationships, a graph structure is constructed. The nodes in the graph represent equipment components, and the edges represent the connections between components. According to this graph structure, an adjacency matrix is constructed, and the elements in the adjacency matrix represent information such as whether there is a connection between components and the strength of the connection.

[0064] Input the temporal context features into the graph convolutional network (GCN) layer. In the graph convolution operation, first perform a linear transformation on the temporal context features of each node (component) to obtain a new feature representation. Then, according to the adjacency matrix, aggregate the feature information of the neighbor nodes of each node. Specifically, for each node, multiply the features of its neighbor nodes by the weights of the corresponding edges in the adjacency matrix, then sum up these weighted neighbor node features, add the features of the node itself after linear transformation, and finally perform a non-linear transformation through an activation function (such as the ReLU function) to obtain the updated features of the node. Through the above graph convolution operation, the local interaction features between components are extracted. For example, in the papermaking felt production equipment, the main shaft of the loom is connected to the transmission components, and through the graph convolution operation, the local interaction features between the main shaft and the transmission components can be captured, and the cooperation between them can be understood, such as how the change in the main shaft speed affects the operation of the transmission components, etc.

[0065] Step S1332: Perform global pooling on the local interaction features to generate global correlation features reflecting the interaction relationship of the components of the entire production line.

[0066] After obtaining the local interaction features between components, perform global pooling on these features. The purpose of global pooling is to integrate the local interaction features to generate a global correlation feature that can reflect the interaction relationship of the components of the entire production line. Common global pooling methods include global average pooling and global max pooling, etc.

[0067] Taking global average pooling as an example, it can calculate the average of all local interaction features in each dimension. For each dimension, add up the local interaction feature values of all nodes (components) in this dimension, and then divide by the number of nodes to obtain the average value in this dimension. Combine the average values of all dimensions to obtain a comprehensive feature vector, which represents the global interaction relationship between the components of the entire production line. For example, in the papermaking felt production process, through global pooling, the local interaction features between various equipment components such as looms, setting machines, and hot presses can be integrated to form a global correlation feature reflecting the cooperation of the equipment of the entire production line, and understand the mutual influence and cooperation degree between various equipment components in the entire production line.

[0068] Step S1333: Input the local interaction features and the global correlation features into the attention fusion layer of the association decoder, and assign weights to the local interaction features of different component pairs through the spatial attention mechanism to highlight the key component pairs that have a significant impact on fault propagation.

[0069] Input the local interaction features and global correlation features into the attention fusion layer of the correlation decoder. In this attention fusion layer, use the spatial attention mechanism to assign weights to the local interaction features of different component pairs. First, concatenate the local interaction features and global correlation features to obtain a new feature vector. Then, input this new feature vector into a fully connected layer for linear transformation, and then perform a non-linear transformation through an activation function (such as the tanh function). Next, perform a dot product operation between the transformed feature vector and a pre-trained attention weight vector to obtain a scalar value, which represents the importance score of the local interaction feature of this component pair for fault propagation.

[0070] Normalize the importance scores of all component pairs. For example, use the softmax function to convert these scores into a probability distribution, that is, the attention weights. According to the attention weights, assign corresponding weights to the local interaction features of different component pairs. For those component pairs that play a key role in fault propagation, assign higher weights, while for those component pairs that have less impact on fault propagation, assign lower weights. For example, in the papermaking felt production equipment, if the faults between the main shaft and the transmission components of the loom are likely to trigger faults in the entire loom system, then the local interaction features of these two component pairs will be assigned higher weights to highlight their importance in fault propagation; while for some component pairs that have less impact on fault propagation, such as the interaction features between some auxiliary components on the loom, lower weights will be assigned.

[0071] Step S1334: Concatenate the weighted local interaction features and global correlation features to generate a correlation fusion feature containing local interaction details and global correlation information.

[0072] After assigning weights to the local interaction features, concatenate the weighted local interaction features and global correlation features. The concatenation operation combines these two features in a set order to form a new feature vector, that is, the correlation fusion feature, which contains both the local interaction details between components and the global correlation information of the entire production line. For example, in papermaking felt production, the correlation fusion feature can simultaneously reflect the local interaction situation between various components of the loom (such as the cooperation details between the main shaft and the transmission components, the feeding device) and the global cooperation relationship between the equipment of the entire production line (such as the mutual influence between the loom, the shaping machine, and the hot press).

[0073] Step S1335: Input the correlation fusion feature into the fully connected layer of the correlation decoder for dimensionality compression processing to generate a correlation context feature reflecting the fault propagation path.

[0074] Input the associated fusion features into the fully connected layer of the association decoder. The role of the fully connected layer is to perform dimensionality compression processing on the associated fusion features. Since the associated fusion features may have a relatively high dimension and contain a large amount of information, the fully connected layer can compress them to a suitable dimension, extract the most crucial information, and generate associated context features reflecting the fault propagation path. The fully connected layer maps the associated fusion features to a lower-dimensional space through linear transformation. During this process, the weight parameters of the fully connected layer are learned through model training, and it can automatically select the important information related to the fault propagation path and filter out some redundant information. For example, in the fault analysis of papermaking felt production equipment, the fully connected layer processes the associated fusion features, removes some feature information that has less impact on the analysis of the fault propagation path, and only retains the key features related to the fault propagation, such as which interactions between components play a key role in the fault propagation and the possible propagation directions of the fault, so as to obtain an associated context feature that can clearly reflect the fault propagation path between equipment components.

[0075] Step S134: Input the temporal context features and the associated context features into the risk assessment layer of the fault prediction model for comprehensive evaluation processing, and generate an intermediate evaluation result containing the probability values of each fault type.

[0076] Input the temporal context features and the associated context features into the risk assessment layer of the fault prediction model. The risk assessment layer first performs a concatenation operation on these two types of features to combine them into a higher-dimensional feature vector. Then, input the concatenated feature vector into a multi-layer perceptron (MLP) for processing. The MLP consists of multiple fully connected layers and activation functions. In each fully connected layer, a linear transformation is performed on the input feature vector, and then a nonlinear factor is introduced through an activation function (such as the ReLU function) to enhance the expressive power of the model.

[0077] In the last layer of the MLP, the softmax activation function is used to convert the output into a probability distribution, which represents the probability values of different fault types occurring, that is, an intermediate evaluation result containing the probability values of each fault type is generated. For example, in papermaking felt production equipment, the risk assessment layer will calculate the probabilities of different fault types such as loom faults, calender faults, and hot press faults based on the temporal context features and the associated context features, comprehensively considering the time sequence information of the development of equipment faults and the interaction relationship between components and the fault propagation path information.

[0078] Step S135: Perform a grading process on the intermediate evaluation result, map the probability values of each fault type to the corresponding risk level according to the preset probability threshold interval, and generate diagnostic result data containing the fault risk level.

[0079] In this embodiment, a series of probability threshold intervals can be set in advance, and each probability threshold interval corresponds to a specific risk level. These probability threshold intervals are determined based on a large amount of historical failure data and actual production experience. For example, through the analysis of numerous failure cases, it is determined that the probability threshold interval corresponding to the low-risk level is a relatively low probability range, the probability threshold interval corresponding to the medium-risk level is a medium probability range, and the probability threshold interval corresponding to the high-risk level is a relatively high probability range.

[0080] For each failure type probability value in the intermediate evaluation result, compare it with the preset probability threshold interval. If the probability value of a certain failure type falls within the threshold interval corresponding to the low-risk level, then this failure type is determined to be of the low-risk level; if it falls within the threshold interval corresponding to the medium-risk level, it is determined to be of the medium-risk level; if it falls within the threshold interval corresponding to the high-risk level, it is determined to be of the high-risk level.

[0081] Integrate each failure type and its corresponding risk level information to generate diagnostic result data containing the failure risk level, so as to intuitively reflect the possible failure risk situation faced by the equipment. For example, in the diagnostic result data of the papermaking felt production equipment, it is clearly shown that the loom failure is of the high-risk level, the stenter failure is of the medium-risk level, and the hot press failure is of the low-risk level, which is convenient for the operation and maintenance personnel to take corresponding measures according to different risk levels.

[0082] Step S140: Identify the potential failure types currently existing in the target production line and the propagation characteristic information of the failure during the operation of the equipment according to the diagnostic result data.

[0083] Step S141: Analyze the failure risk level and the corresponding failure type probability values in the diagnostic result data, and screen out the top several failure types with the highest probability values as potential failure types.

[0084] Analyze the diagnostic result data, extract the failure risk level and the probability values of the corresponding failure types from it. Then, screen out the top several failure types with the highest probability values from these failure types and use them as the potential failure types currently existing in the target production line. For example, in the diagnostic result data of the papermaking felt production equipment, it may include the probability values and risk levels of various failure types such as loom failure, stenter failure, and hot press failure. Screen out several failure types with the highest probability values, such as the main shaft failure of the loom and the temperature control failure of the stenter, as potential failure types.

[0085] Step S142: Extract the temporal context features and associated context features corresponding to each potential failure type, analyze the association relationships between the failure types and the state evolution features and component association features, and generate failure trigger factor identifiers.

[0086] For each potential fault type, extract its corresponding temporal context features and associated context features. Through the analysis of these features, study the correlation between the fault type and the previously extracted state evolution features and component association features. For example, if the temporal context features corresponding to a certain potential fault type show abnormal fluctuations in the device state within a certain period of time, and the associated context features show a change in the interaction relationship between a certain pair of key components, further analyze in combination with the state evolution features and component association features to find out the possible causes leading to the occurrence of this fault type, and generate a fault trigger factor identifier. This fault trigger factor identifier can clearly indicate which factors trigger the potential fault. For example, the loom main shaft fault may be caused by factors such as long-term high-load operation (reflected by the state evolution features) and loose connection between the main shaft and the transmission components (reflected by the component association features).

[0087] Step S143: Perform a retrospective analysis on the state data segment related to the potential fault type in the device monitoring data stream to determine the time point when the fault first occurred and the corresponding initial affected component.

[0088] Perform a retrospective analysis on the state data segment related to the potential fault type in the device monitoring data stream. Starting from the current moment, trace the data forward to find the first sign of the fault occurrence. By analyzing the changes in the state data segment, determine the time point when the fault first occurred. At the same time, in combination with the component association features, find out the components affected at this time point and use them as the initial affected components. For example, in a papermaking felt production device, if it is found that the loom main shaft fault is a potential fault type, through retrospective analysis of the state data segments such as the rotation speed and vibration of the loom main shaft in the device monitoring data stream, determine the time point when the main shaft fault first occurred, and judge according to the component association features that the transmission components connected to the main shaft may be the initially affected components.

[0089] Step S144: Track the changes in the state evolution features of the initial affected component within the subsequent time window, and in combination with the interaction intensity parameter and causal relationship identifier in the component association features, deduce the propagation order and propagation speed of the fault between components.

[0090] Step S1441: Taking the initial affected component as the starting point, determine the set of downstream components directly affected by it according to the causal relationship identifier in the component association features.

[0091] Starting from a determined initial affected component, based on the causal relationship identification in the component association features, find the set of downstream components directly affected by this initial affected component. The causal relationship identification clarifies the direction of influence between components. According to this causal relationship identification, it can be determined which components will be directly affected by the initial affected component. For example, in a papermaking felt production device, if the main shaft of a loom is the initial affected component, according to the causal relationship identification, it can be determined that the transmission parts, feeding devices, etc. connected to the main shaft are the downstream components directly affected by it, and these components form a set of downstream components.

[0092] Step S1442: Extract the change amounts of the state evolution characteristics of each downstream component in each time window after the initial affected component fails, and calculate the delay time of the fault influence transmission in combination with the interaction intensity parameter.

[0093] For each component in the set of downstream components, extract the change amounts of the state evolution characteristics in each time window after the initial affected component fails. The change amount of the state evolution characteristic reflects the degree of change in the state of the component after the fault occurs. At the same time, in combination with the interaction intensity parameter in the component association features, calculate the delay time for the fault influence to be transmitted from the initial affected component to the downstream component. The interaction intensity parameter reflects the degree of co-variation between components. The stronger the interaction intensity, the shorter the delay time for the fault influence to be transmitted may be. For example, in a loom, if the interaction intensity between the main shaft and the transmission parts is relatively high, then the influence of the main shaft fault on the transmission parts may be reflected more quickly, and the delay time is shorter.

[0094] Step S1443: Determine the order in which the fault propagates from the initial affected component to the downstream components according to the delay time, and generate a propagation order identifier including a list of component orders.

[0095] According to the calculated delay time of the fault influence transmission, determine the order in which the fault propagates from the initial affected component to the downstream components. The shorter the delay time, the earlier the fault propagates to this downstream component. According to this order, arrange the downstream components into a list, and generate a propagation order identifier including a list of component orders. For example, in a papermaking felt production device, if the influence delay time of the main shaft fault on the transmission parts is shorter and the influence delay time on the feeding device is longer, then the list of component orders in the propagation order identifier will rank the transmission parts in front of the feeding device, indicating that the fault first propagates to the transmission parts and then to the feeding device.

[0096] Step S1444: Calculate the time interval for the fault to propagate between adjacent components, and generate a propagation speed index reflecting the fault diffusion rate in combination with the physical distance parameter between components.

[0097] Calculate the time interval for a fault to propagate between adjacent components, that is, the time it takes for a fault to spread from one component to its adjacent component. At the same time, combined with the physical distance parameter between components, generate a propagation speed index reflecting the fault diffusion rate. The propagation speed index can be obtained by dividing the physical distance by the propagation time interval, which can intuitively reflect the diffusion speed of the fault between components. For example, in a papermaking felt production equipment, if the physical distance between the main shaft and the transmission components of a loom is a certain value, and the time interval for a fault to spread from the main shaft to the transmission components is a certain value, the propagation speed index of the fault between these two components can be obtained through the calculation of the two.

[0098] Step S1445: Recursively analyze and process the downstream components of the downstream components until all components affected by the fault are covered, and generate a complete fault propagation order list and propagation speed distribution.

[0099] Recursively analyze the downstream components of the downstream components. For example, repeat the above steps of determining downstream components, calculating delay time, determining propagation order and propagation speed until all components affected by the fault are covered. Through the above recursive analysis, generate a complete fault propagation order list to clarify the propagation order of the fault between all affected components, and at the same time generate a propagation speed distribution to show the propagation speed of the fault between different components. For example, in a papermaking felt production equipment, starting from the initial affected component, analyze its downstream components, downstream components of the downstream components, etc. in turn, and finally obtain a complete propagation order list and propagation speed distribution covering all components affected by the fault.

[0100] Step S1446: Integrate the propagation order list and the propagation speed distribution to obtain the propagation order and propagation speed of the fault between components.

[0101] Integrate the generated propagation order list and propagation speed distribution to obtain the propagation order and propagation speed of the fault between components, so as to comprehensively reflect the propagation process of the fault between the equipment components. For example, in a papermaking felt production equipment, the integrated propagation order and propagation speed can clearly show how the fault gradually spreads from the initial affected component to other components and the propagation speed between different components.

[0102] Step S145: Integrate the fault trigger factor identification, initial affected component, propagation order and propagation speed to generate propagation characteristic information reflecting the propagation characteristics of the fault during the operation of the equipment.

[0103] Integrate the fault trigger factors identification, initial affected components, propagation order, and propagation speed to generate propagation characteristic information that reflects the propagation characteristics of faults during the operation of the device. This propagation characteristic information includes key contents such as the cause of the fault, the starting component, the propagation path, and the propagation speed. For example, in a papermaking felt production device, the propagation characteristic information can clearly indicate that the loom main shaft fault is triggered by high-load operation and loose connections. Starting from the main shaft, it affects other components according to the set propagation order, and also gives the propagation speed of the fault between different components, which helps the operation and maintenance personnel quickly locate the fault and formulate corresponding maintenance strategies.

[0104] Step S150: Generate maintenance guidance data containing fault location identification based on the potential fault type and the propagation characteristic information, and transmit the maintenance guidance data to the factory operation and maintenance system to trigger fault intervention operations.

[0105] Step S151: Match the preset maintenance strategy library according to the potential fault type, and extract the maintenance operation steps and required tool identification corresponding to each potential fault type.

[0106] Match according to the previously determined potential fault type in the preset maintenance strategy library. The maintenance strategy library stores the maintenance operation steps and required tool identification corresponding to various fault types. For each potential fault type, extract the corresponding maintenance operation steps and required tool identification from the maintenance strategy library. For example, in a papermaking felt production device, if the potential fault type is loom main shaft fault, extract the maintenance operation steps for the main shaft fault from the maintenance strategy library, such as checking the connection of the main shaft and replacing worn components, and at the same time extract the required tool identification, such as wrenches, screwdrivers, etc.

[0107] Step S152: Analyze the initial affected components and propagation order list in the propagation characteristic information to determine the set of key components that need to be checked first.

[0108] Analyze the initial affected components and propagation order list in the propagation characteristic information. Based on the fact that the initial affected component is the starting point of the fault and the propagation order list reflects the fault propagation path, determine the set of key components that need to be checked first. These key components are the components that are most easily affected by the fault. Checking them can quickly locate the scope and degree of the fault. For example, in a papermaking felt production device, if the propagation characteristic information shows that the fault starts from the loom main shaft and affects the transmission components, feeding device, etc. in a set order, then components such as the main shaft, transmission components, and feeding device will be determined as the set of key components that need to be checked first.

[0109] Step S153: Extract the status data segments corresponding to the key components from the device monitoring data stream, and generate an abnormal status description of the key components by combining the fluctuation pattern and periodic anomaly identification in the status evolution characteristics.

[0110] Extract the status data segments corresponding to the key components from the device monitoring data stream. Analyze the status data segments of the key components by combining the fluctuation pattern and periodic anomaly identification in the previously extracted status evolution characteristics, and generate an abnormal status description of the key components. For example, for the main shaft of a loom, extract the status data segments such as its rotation speed and vibration. If the status evolution characteristics show that there are periodic abnormal fluctuations in the main shaft rotation speed, then an abnormal status description of the main shaft can be generated as "There are periodic abnormal fluctuations in the main shaft rotation speed".

[0111] Step S154: Perform an association mapping process on the physical location information of the key components and the abnormal status description to generate a fault location identifier containing component position coordinates and abnormal characteristics.

[0112] Perform an association mapping process on the physical location information of the key components and the abnormal status description. Each key component has its specific physical location coordinates on the production line. Associate these coordinates with the corresponding abnormal status description to generate a fault location identifier containing component position coordinates and abnormal characteristics. For example, in a papermaking felt production device, associate the physical location coordinates of the loom main shaft with the abnormal status description of "There are periodic abnormal fluctuations in the main shaft rotation speed" to form a clear fault location identifier, which is convenient for maintenance personnel to quickly find the faulty component.

[0113] Step S155: Integrate the maintenance operation steps, required tool identifiers, key component set, and fault location identifier to generate a maintenance task list including the time sequence.

[0114] Integrate the extracted maintenance operation steps, required tool identifiers, key component set, and fault location identifier. Generate a maintenance task list according to the logic and time sequence of fault handling. This maintenance task list clarifies the content of each maintenance task, the required tools, the key components involved, and the position information of the components, and arranges them in the order of priority. For example, in a papermaking felt production device, the maintenance task list may first be to check the connection of the loom main shaft using a wrench (corresponding to the fault location identifier of the main shaft), and then to replace the worn parts of the transmission components using a screwdriver (corresponding to the fault location identifier of the transmission components), etc.

[0115] Step S156: Perform a binding process on the maintenance task list and the fault location identifier to generate maintenance guidance data for guiding on-site maintenance personnel.

[0116] Bind the maintenance task list to the fault location identifier. Through the above binding, each maintenance task is associated with the corresponding fault location identifier, forming a complete maintenance guidance data. This maintenance guidance data can intuitively guide on-site operation and maintenance personnel to handle faults. They can quickly and accurately reach the location of the faulty component according to the task list and fault location identifier in the maintenance guidance data, and take corresponding maintenance operations. For example, in the papermaking felt production equipment, on-site operation and maintenance personnel can, according to the maintenance guidance data, first find the position of the loom main shaft, check and repair according to the steps in the maintenance task list, and then find the position of the transmission components according to the guidance for corresponding treatment. Finally, the generated maintenance guidance data is transmitted to the factory operation and maintenance system to trigger a fault intervention operation, and the factory operation and maintenance system will arrange corresponding personnel and resources for fault handling according to the maintenance guidance data.

[0117] During the whole process, the data collection involved may include some privacy-sensitive data, such as specific operating parameters of the equipment. To protect this privacy-sensitive data, a variety of privacy protection and anti-disclosure technical means are adopted. For example, during data transmission, encryption algorithms are used to encrypt the data to ensure that the data is not stolen or tampered with during transmission. In terms of data storage, access control technology is adopted, and only authorized personnel can access this data, and detailed logs of data access are recorded for tracking and auditing. At the same time, the data is anonymized to remove the sensitive information that may be contained in the data to further enhance the security of the data. And the data storage system is regularly subjected to security detection and vulnerability repair to prevent the leakage of privacy-sensitive data due to system vulnerabilities.

[0118] In the construction and training of the artificial intelligence model, for the fault prediction model, it mainly consists of modules such as a feature fusion layer, a time series encoder, an association decoder, and a risk assessment layer. The feature fusion layer is responsible for cross-dimensional information fusion of the state evolution feature and the component association feature, and generates a fusion feature vector in combination with the previously dynamically adjusted feature weight coefficient. During the fusion process, it is necessary to ensure the dimension matching and dimension consistency of different features to avoid unreasonable calculations. For example, for the weighted addition operation of different features, it is necessary to ensure that these features are consistent in dimension, otherwise the calculation result may lose its practical meaning.

[0119] The sequential encoder includes a bidirectional recurrent neural network layer, a temporal attention mechanism, and a gated recurrent unit layer. The bidirectional recurrent neural network layer processes the fused feature vectors in sequence from the forward time order and the backward time order respectively, generating bidirectional hidden states containing past state information and future state information. The purpose of this step is to fully capture the temporal correlation information of the device state before and after. The temporal attention mechanism assigns attention weights to the hidden states at each time step, highlighting the time step information important for fault diagnosis. The gated recurrent unit layer then updates the state of the sequential aggregated features, filters out redundant information, and retains the key fault evolution clues, finally generating sequential context features reflecting the fault development process.

[0120] The correlation decoder consists of a graph convolutional network layer, a global pooling layer, an attention fusion layer, and a fully connected layer. The graph convolutional network layer constructs an adjacency matrix with the device connection relationship of the target production line as the graph structure, and extracts the local interaction features between components through graph convolutional operations. The global pooling layer processes the local interaction features, generating global correlation features reflecting the interaction relationship of the components of the entire production line. The attention fusion layer uses a spatial attention mechanism to assign weights to the local interaction features of different component pairs, highlighting the key component pairs. The fully connected layer performs dimensionality compression processing on the correlation fusion features, generating correlation context features reflecting the fault propagation path.

[0121] The risk assessment layer comprehensively evaluates the sequential context features and the correlation context features, generating an intermediate evaluation result containing the probability values of each fault type, and then through a grading process, generating diagnostic result data containing the fault risk level.

[0122] For the training of the fault prediction model, a large amount of historical device monitoring data and corresponding fault label data are required. The training steps are as follows: Collect the historical monitoring data of the papermaking felt production equipment, including the state data collected by various sensors, and at the same time label the fault types and fault occurrence times and other information corresponding to these data to form a training data set. Preprocess the data, including data cleaning to remove noise data and outliers; perform data normalization or standardization processing to ensure the dimensional consistency of different features, facilitating model learning.

[0123] Initialize the various modules and parameters of the fault prediction model, including the weights of the feature fusion layer, the neuron parameters of the bidirectional recurrent neural network layer, the convolution kernel parameters of the graph convolutional network layer, etc. The initial values of these parameters can be initialized randomly, but it is necessary to ensure that the value range of the parameters is reasonable to avoid numerical instability.

[0124] Input the preprocessed training data into the fault prediction model and train it according to the forward propagation and backward propagation processes. During forward propagation, the data sequentially passes through the feature fusion layer, the temporal encoder, the association decoder, and the risk assessment layer to generate the predicted fault risk level and the probability values of fault types. During backward propagation, calculate the loss function, such as the cross-entropy loss function, based on the prediction results and the true fault label data, which can measure the difference between the model prediction results and the true labels. Update the model parameters through optimization algorithms, such as the stochastic gradient descent algorithm, to gradually reduce the value of the loss function and continuously improve the prediction performance of the model.

[0125] Use a part of the test data that did not participate in the training to evaluate the trained model. Calculate the evaluation metrics of the model, such as accuracy, recall rate, F1 value, etc., to evaluate the performance of the model. If the evaluation metrics do not meet the requirements, adjust the model parameters or increase the training data and retrain.

[0126] When the evaluation metrics of the model reach a satisfactory level, deploy the trained model to the actual intelligent factory environment for real-time fault prediction and diagnosis.

[0127] During the implementation process of the entire fault diagnosis method, when the device monitoring data stream shows abnormal changes in some status data of the loom, unify the data with different acquisition frequencies to the standard time axis through data alignment processing, and then perform state evolution extraction and component association extraction. The state evolution features may show that the rotation speed fluctuation pattern of the loom has changed and the periodic anomaly identification has become abnormal; the component association features may indicate that the interaction intensity parameter and the association stability index between the main shaft and the transmission components of the loom have changed. Input these features into the pre-trained fault prediction model, and the model predicts the possible fault types and fault risk levels according to the knowledge and rules obtained from training.

[0128] Next, identify the potential fault types and fault propagation feature information based on the diagnostic result data. If it is predicted that the main shaft of the loom may fail, determine the time point when the fault first occurred and the initial affected components through backtracking analysis, and track the propagation order and speed of the fault among the components. For example, the fault may start from the main shaft and affect the transmission components, the feeding device, etc. in a set order, and the speed of fault diffusion can be predicted according to the propagation speed index.

[0129] Finally, maintenance guidance data is generated based on potential fault types and propagation characteristic information. Match the preset maintenance strategy library according to the potential fault types, and extract the maintenance operation steps and the required tool identifiers. Determine the set of key components that need to be checked first, such as the main shaft, transmission components, etc., extract the abnormal state descriptions of these key components, and map them associatively with their physical location information to generate fault location identifiers. Bind the maintenance task list with the fault location identifiers to form maintenance guidance data and transmit it to the factory operation and maintenance system. The factory operation and maintenance system arranges maintenance personnel to perform fault intervention operations according to the maintenance guidance data, such as checking the connection of the main shaft, replacing worn transmission components, etc., so as to solve equipment faults in time and ensure the normal operation of the papermaking wire blanket production.

[0130] Figure 2 FIG. shows a schematic diagram of exemplary hardware and software components of an intelligent factory fault diagnosis system 100 based on an AI prediction model that can implement the idea of the present application provided by some embodiments of the present application. For example, the processor 120 can be used on the intelligent factory fault diagnosis system 100 based on the AI prediction model and is used to execute the functions in the present application.

[0131] The intelligent factory fault diagnosis system 100 based on the AI prediction model can be a general-purpose server or a special-purpose server, both of which can be used to implement the intelligent factory fault diagnosis method based on the AI prediction model of the present application. Although only one server is shown in the present application, for convenience, the functions described in the present application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0132] For example, the intelligent factory fault diagnosis system 100 based on the AI prediction model can include a network port 110 connected to the network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as disks, ROM, or RAM, or any combination thereof. Exemplarily, the intelligent factory fault diagnosis system 100 based on the AI prediction model can also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present application can be implemented according to these program instructions. The intelligent factory fault diagnosis system 100 based on the AI prediction model also includes an I / O interface 150 between the computer and other input and output devices.

[0133] For ease of explanation, only one processor is described in the intelligent factory fault diagnosis system 100 based on the AI prediction model. However, it should be noted that the intelligent factory fault diagnosis system 100 based on the AI prediction model in this application may also include multiple processors. Therefore, the steps performed by one processor described in this application can also be jointly performed or separately performed by multiple processors. For example, if the processor of the intelligent factory fault diagnosis system 100 based on the AI prediction model performs step A and step B, it should be understood that step A and step B can also be jointly performed by two different processors or separately performed in one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0134] In addition, an embodiment of the present invention also provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the above intelligent factory fault diagnosis method based on the AI prediction model is implemented.

[0135] It should be noted that, in order to simplify the description of the present invention disclosure and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, sometimes multiple features are merged into one embodiment, drawing, or description thereof.

Claims

1. An intelligent factory fault diagnosis method based on an AI prediction model, characterized in that The method includes: Obtaining the device monitoring data stream of the target production line in the intelligent factory, where the device monitoring data stream includes multiple groups of status data segments with timestamp identifiers collected continuously; Performing diagnostic feature construction processing on the device monitoring data stream to generate a state evolution feature reflecting the device operation state and a component association feature reflecting the interaction relationship between device components; Inputting the state evolution feature and the component association feature into a pre-trained fault prediction model for fault prediction to generate diagnostic result data including a fault risk level; Identifying the potential fault types currently existing in the target production line and the propagation feature information of the faults during the device operation according to the diagnostic result data; Generating maintenance guidance data including fault location identifiers based on the potential fault types and the propagation feature information, and transmitting the maintenance guidance data to the factory operation and maintenance system to trigger a fault intervention operation.

2. The intelligent factory fault diagnosis method based on the AI prediction model according to claim 1, characterized in that The performing diagnostic feature construction processing on the device monitoring data stream to generate a state evolution feature reflecting the device operation state and a component association feature reflecting the interaction relationship between device components includes: Performing data alignment processing on the device monitoring data stream, mapping status data segments with different collection frequencies to a standard time axis uniformly to generate a time-aligned data set; Performing state evolution extraction processing on the time-aligned data set, analyzing the change trend of continuous status data segments through a sliding time window to generate a state evolution feature including a state fluctuation pattern and a periodic anomaly identifier; Performing component association extraction processing on the time-aligned data set, analyzing the co-variation law of status data segments corresponding to different components based on the device physical connection relationship to generate a component association feature including an interaction intensity parameter and an association stability index; Inputting the state evolution feature and the component association feature into a feature calibration module for dimension matching processing to eliminate the scale difference between different features and generate a calibrated feature set with a unified representation form; Dynamically adjusting the feature weight coefficients of the state evolution feature and the component association feature in fault prediction according to the evaluation result of the influence degree of each dimension feature in the calibrated feature set on fault diagnosis.

3. The intelligent factory fault diagnosis method based on the AI prediction model according to claim 2, wherein, The performing state evolution extraction processing on the time-aligned data set, analyzing the change trend of continuous status data segments through a sliding time window to generate a state evolution feature including a state fluctuation pattern and a periodic anomaly identifier includes: Setting a sliding time window with a fixed length to segment the time-aligned data set to obtain multiple data window units with time continuity; Calculating the mean, variance and extreme point distribution of the status data segments in each data window unit to generate basic statistical features reflecting state stability; Analyzing the change amount of the basic statistical features between adjacent data window units, and extracting the trend parameters of the status data segments in continuous time windows, where the trend parameters include an upward trend parameter, a downward trend parameter and a trend persistence index; Perform periodic pattern detection processing on the data window unit, identify the recurring fluctuation periods in the state data segment through autocorrelation analysis, and generate a period length parameter and a period stability score; Integrate the basic statistical features, trend parameters, period length parameters, and period stability scores to obtain state evolution features including state fluctuation patterns and periodic anomaly identifiers.

4. The intelligent factory fault diagnosis method based on the AI prediction model according to claim 2, wherein Perform component association extraction processing on the time-aligned data set, analyze the co-variation law of the corresponding state data segments of different components based on the device physical connection relationship, and generate component association features including interaction intensity parameters and association stability indicators, including: Determine key component pairs according to the device connection map of the target production line, and extract the time series of the corresponding state data segments of each pair of key components; Calculate the cross-correlation coefficient of the time series of each pair of key components to generate an interaction intensity parameter reflecting the degree of co-variation between components; Analyze the fluctuation range of the cross-correlation coefficient within consecutive time windows to generate an association stability indicator reflecting the persistence of the component association relationship; Perform causality detection processing on the state data segments of the key component pairs, determine the direction of causal influence between components through Granger causality test, and generate a causality identifier; Integrate the interaction intensity parameters, association stability indicators, and causality identifiers to obtain component association features including interaction intensity parameters and association stability indicators.

5. The intelligent factory fault diagnosis method based on the AI prediction model according to claim 2, wherein, Input the state evolution features and the component association features into a pre-trained fault prediction model for fault prediction, and generate diagnostic result data including fault risk levels, including: Input the state evolution features and the component association features into the feature fusion layer of the fault prediction model, perform cross-dimensional information fusion processing in combination with the feature weight coefficients, and generate a fused feature vector; Perform time-dependency modeling processing on the fused feature vector through the time series encoder of the fault prediction model to extract time series context features reflecting the fault development process; Use the association decoder of the fault prediction model to perform component interaction relationship parsing processing on the time series context features to generate association context features reflecting the fault propagation path; Input the time series context features and the association context features into the risk assessment layer of the fault prediction model for comprehensive evaluation processing to generate an intermediate evaluation result including the probability values of each fault type; Perform level division processing on the intermediate evaluation result, map the probability values of each fault type to the corresponding risk levels according to the preset probability threshold interval, and generate diagnostic result data including fault risk levels.

6. The intelligent factory fault diagnosis method based on the AI prediction model according to claim 5, characterized in that Perform time-dependency modeling processing on the fused feature vector through the time series encoder of the fault prediction model to extract time series context features reflecting the fault development process, including: Input the fused feature vector into the bidirectional recurrent neural network layer of the time series encoder, perform sequence processing on the fused feature vector from both the forward time order and the reverse time order, and generate bidirectional hidden states including past state information and future state information; Perform attention weight calculation processing on the bidirectional hidden state, and assign attention weights reflecting their importance for fault diagnosis to the hidden states at each time step through the temporal attention mechanism; Perform weighted summation processing on the bidirectional hidden state according to the attention weights to generate a temporal aggregation feature containing key time step information; Input the temporal aggregation feature into the gated recurrent unit layer of the temporal encoder for state update processing, filter redundant historical state information, and retain key fault evolution clues to generate a temporal context feature reflecting the fault development process.

7. The intelligent factory fault diagnosis method based on the AI prediction model according to claim 5, characterized in that, Use the association decoder of the fault prediction model to perform component interaction relationship parsing processing on the temporal context feature to generate an association context feature reflecting the fault propagation path, including: Input the temporal context feature into the graph convolutional network layer of the association decoder, construct an adjacency matrix with the device connection relationship of the target production line as the graph structure, and extract local interaction features between components through graph convolutional operations; Perform global pooling processing on the local interaction features to generate a global association feature reflecting the component interaction relationship of the entire production line; Input the local interaction feature and the global association feature into the attention fusion layer of the association decoder, and assign weights to the local interaction features of different component pairs through the spatial attention mechanism to highlight the key component pairs that have a significant impact on fault propagation; Perform concatenation processing on the weighted local interaction feature and the global association feature to generate an association fusion feature containing local interaction details and global association information; Input the association fusion feature into the fully connected layer of the association decoder for dimension compression processing to generate an association context feature reflecting the fault propagation path.

8. The intelligent factory fault diagnosis method based on the AI prediction model according to claim 1, characterized in that, Identify the potential fault types currently existing in the target production line and the propagation feature information of the faults during the device operation according to the diagnostic result data, including: Parse the fault risk level and the corresponding fault type probability values in the diagnostic result data, and screen the top N fault types with the highest probability values as potential fault types; Extract the temporal context feature and the association context feature corresponding to each potential fault type, analyze the association relationship between the fault type and the state evolution feature and the component association feature, and generate a fault trigger factor identifier; Perform retrospective analysis processing on the state data segment related to the potential fault type in the device monitoring data stream to determine the time point when the fault first appears and the corresponding initial affected component; Track the change of the state evolution feature of the initial affected component in the subsequent time window, and combine the interaction intensity parameter and the causal relationship identifier in the component association feature to deduce the propagation order and propagation speed of the fault between components; Integrate the fault trigger factor identifier, the initial affected component, the propagation order, and the propagation speed to generate propagation feature information reflecting the propagation characteristics of the fault during the device operation.

9. The intelligent factory fault diagnosis method based on the AI prediction model according to claim 8, characterized in that, Track the change of the state evolution feature of the initial affected component in the subsequent time window, and combine the interaction intensity parameter and the causal relationship identifier in the component association feature to deduce the propagation order and propagation speed of the fault between components, including: Starting from the initial impact component, determine the set of downstream components directly affected by it according to the causal relationship identification in the component association features; Extract the change amount of the state evolution characteristics of each downstream component in each time window after the initial impact component fails, and calculate the delay time of the fault impact transmission in combination with the interaction intensity parameter; Determine the order in which the fault propagates from the initial impact component to the downstream components according to the delay time, and generate a propagation order identifier including a component order list; Calculate the time interval for the fault to propagate between adjacent components, and generate a propagation speed index reflecting the fault diffusion rate in combination with the physical distance parameter between the components; Perform recursive analysis and processing on the downstream components of the downstream components until all components affected by the fault are covered, and generate a complete fault propagation order list and propagation speed distribution; Integrate the propagation order list and the propagation speed distribution to obtain the fault propagation order and propagation speed between the components.

10. An intelligent factory fault diagnosis system based on an AI prediction model, characterized in that, It includes a processor and a memory. The memory is connected to the processor. The memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the intelligent factory fault diagnosis method based on the AI prediction model according to any one of claims 1-9 above.

Images