Water jet loom service life prediction method and system based on deep learning

Abstract

The invention relates to the technical field of equipment intelligent operation and maintenance, in particular to a water jet loom service life prediction method and system based on deep learning, and the method comprises the steps: obtaining the operation data of a water jet loom through a multi-modal sensor cooperation mechanism, and dynamically adjusting the sampling frequency of a sensor based on the operation state of the water jet loom; the method comprises the following steps: acquiring operation data, preprocessing the acquired operation data, performing feature extraction based on the preprocessed data to obtain a multi-dimensional feature set, constructing a life prediction model by taking extracted features as input, outputting a residual life prediction value and an uncertainty index based on the optimized life prediction model, and executing residual life grading early warning according to a prediction result. According to the method, the prediction uncertainty is quantified while the residual life prediction value is output, the early warning levels are divided according to the prediction result and the fault features in combination with a hierarchical early warning mechanism, decision blindness caused by a single prediction value is avoided, and the fault response time is shortened.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment intelligent operation and maintenance, in particular to a water-jet loom life prediction method and system based on deep learning. BACKGROUND

[0002] As the core production equipment in the textile industry, the running stability and life of the water-jet loom directly determine the production efficiency, product quality and enterprise operating cost. With the transformation of the textile industry towards automation and intelligence, the traditional post-maintenance or regular preventive maintenance mode has been difficult to meet the modern production needs. Therefore, it is urgent to realize the accurate prediction of the remaining life of the water-jet loom.

[0003] The existing related technologies attempt to realize the water-jet loom state monitoring or life prediction through data acquisition and analysis, but there are still significant defects in the practicality and adaptability, which is difficult to meet the actual production needs of enterprises. The traditional method realizes fault early warning based on feature analysis by deploying vibration and temperature sensors at key positions of the loom to collect data. However, this technology uses a fixed sampling frequency and does not combine the actual weaving load of the loom, such as high load when weaving heavy fabrics and low load when weaving light fabrics, to dynamically adjust the sampling strategy. This results in excessive redundant data, increased storage and transmission costs under light load conditions, and insufficient sampling frequency of key fault signals under high load conditions, which cannot capture early fault precursor characteristics. When the sensor drifts, such as a temperature sensor with a zero point offset of more than 2℃ or poor contact, the distorted data collected will directly lead to false alarms. In addition, new models have a sharp drop in prediction accuracy due to insufficient data accumulation, and textile enterprises update and iterate frequently. Such technology is difficult to quickly adapt to new equipment, and its practicality is greatly discounted. At present, a water-jet loom life prediction method and system based on deep learning. SUMMARY

[0004] In order to solve the problems of low prediction accuracy and low adaptability to complex scenes in traditional prediction methods, the present application provides a water-jet loom life prediction method and system based on deep learning.

[0005] In the first aspect, the present application provides a water-jet loom life prediction method based on deep learning, which adopts the following technical solution:

[0006] A water-jet loom life prediction method based on deep learning, comprising:

[0007] S1, obtaining the running data of the water-jet loom through a multi-modal sensor cooperation mechanism, and dynamically adjusting the sensor sampling frequency based on the running state of the water-jet loom;

[0008] S2, preprocessing the acquired operation data, including denoising the operation data by adaptive wavelet threshold, and introducing an attention mechanism to fill in missing values to obtain preprocessed data;

[0009] S3, feature extraction based on the preprocessed data, obtaining a multi-dimensional feature set, including extracting time domain, frequency domain and time series features, and using a fault sensitive feature enhancement method to amplify small fault features;

[0010] S4, constructing a life prediction model with the extracted features as input, including using a multi-scale LSTM network as a model framework, embedding a feature attention layer based on the model framework, and configuring an L2 regularization strategy;

[0011] S5, using a cross-model transfer learning method to optimize the life prediction model;

[0012] S6, outputting the residual life prediction value and uncertainty index based on the optimized life prediction model, and executing residual life classification warning according to the prediction result.

[0013] Further, the operation data of the water jet loom is obtained by a multi-modal sensor cooperation mechanism, including establishing a cross-sensor data association rule library based on the signal association characteristics of each key monitoring part of the water jet loom when a fault occurs, determining the associated sensor of each sensor according to the association rule library, calculating the data validity coefficient of each sensor according to the deviation degree of the real-time mean value and the standard mean value of the sensor collected data, when the validity coefficient is lower than the set threshold, the associated sensor performs the operation data acquisition task, and synchronously generates and sends the sensor maintenance alarm information, the calculation formula of the data validity coefficient is: ,

[0014] Wherein, is the number of valid data points collected by the sensor per unit time, is the total number of data points that should be collected by the sensor per unit time, is the real-time mean value of the sensor collected data, is the calibration standard mean value of the sensor when it leaves the factory.

[0015] Further, the sensor sampling frequency is dynamically adjusted based on the running state of the water jet loom, including determining the current running state of the water jet loom according to the implementation operation data, combining the historical operation data of each key part of the loom to construct a sampling frequency dynamic adjustment mechanism, and adaptively adjusting the sampling frequency of each sensor according to the change of the loom running load and the running stability of the key part, the adjustment expression of the sampling frequency is: ,

[0016] Wherein, The sampling frequency is the one after the sensor is dynamically adjusted. This is the sensor's fundamental sampling frequency. The weighting of the operating load effect is given by L, which is the real-time operating load coefficient of the loom. Weights that affect operational stability This refers to the operational stability coefficient of critical components.

[0017] Furthermore, the step of denoising the running data using an adaptive wavelet threshold includes performing multi-scale wavelet decomposition on the vibration signal in the running data according to a preset wavelet basis function to obtain low-frequency approximation coefficients and high-frequency detail coefficients. The signal power and noise power are calculated based on the low-frequency approximation coefficients and high-frequency detail coefficients, and the real-time signal-to-noise ratio is calculated. A basic threshold is calculated using a heuristic method. An adaptive threshold is generated by combining the real-time signal-to-noise ratio. Coefficients in the high-frequency detail coefficients whose absolute values ​​are less than the adaptive threshold are set to zero, while coefficients with values ​​greater than the threshold are retained to preserve valid fault features. An inverse wavelet transform is performed on the processed low-frequency approximation coefficients and high-frequency detail coefficients to reconstruct the denoised running data.

[0018] Furthermore, the introduction of an attention mechanism to impute missing values ​​and obtain preprocessed data specifically includes the following steps:

[0019] S21. Determine the associated data corresponding to the missing value, including time series data of the same type at consecutive time steps before and after the missing value, and cross-sensor associated data at the same time step as the missing value.

[0020] S22. The similarity between the associated data and the missing values ​​is calculated using the Pearson correlation coefficient, and the attention weight of each associated data point is calculated based on the similarity. Data type weights are set, and the imputation value of the missing value is calculated in combination with the attention weights.

[0021] S23. Perform standardization processing on the denoised and missing value imputation running data, and calculate the fidelity of the preprocessed data by the signal distortion after denoising, the missing value imputation deviation and the standardization data deviation.

[0022] The formula for calculating the fill value is: ,

[0023] Where K is the number of associative data types, Let i be the number of points in the i-th type of associated data. Attention weights for each associated data point, For the j-th data point of the i-th type of associated data, Weights for data types.

[0024] Furthermore, the method of amplifying minor fault features using fault-sensitive feature enhancement includes: statistically obtaining the average peak index, average Crest factor, and average temperature change rate under normal conditions from historical normal operation data of the loom; setting weight coefficients and extracting the peak index PI and Crest factor from the time-domain feature set; extracting the temperature change rate from the time-series feature set; calculating the fault-sensitive factor based on the weight coefficients and correlating it with time-domain, frequency-domain, and time-series features to form an enhanced feature set; calculating the correlation degree between each feature in the enhanced feature set and the remaining life of the loom; and filtering features based on the correlation degree to obtain a multi-dimensional feature set. The correlation degree calculation formula is as follows: ,

[0025] in, Let x be the mutual information entropy between feature x and lifetime Y. is the Spearman rank correlation coefficient between feature x and lifetime Y.

[0026] Furthermore, the step of constructing a lifetime prediction model using the extracted features as input includes classifying the extracted features based on the time-scale differences of different fault features of the water-jet loom, constructing an LSTM sub-network for each type of fault feature, and introducing a dynamic dropout mechanism into each sub-network. The activation function of the LSTM sub-network adopts the tanh function with an incorporation of a temperature coefficient. A feature attention layer is set between the output of the LSTM sub-network and the fully connected layer. The feature attention layer calculates the correlation weight between each feature and the remaining lifetime of the loom through a bidirectional gating mechanism, and dynamically weights the multi-scale LSTM output features based on the correlation weight. The dynamic weighting calculation formula is as follows: ,

[0027] in, Let be the attention weight for the i-th feature. Let i be the i-th feature value output by the multi-scale LSTM. and These represent different trainable parameters of the attention layer. As a fault-sensitive factor, Here, is the feedback adjustment coefficient, and N is the total number of features.

[0028] Furthermore, the step of constructing a lifetime prediction model using the extracted features as input also includes adding a hierarchical L2 regularization term to the weight parameters of the fully connected layer, and introducing dynamic adjustment of weight decay in the L2 regularization term. Different regularization coefficients are set for the weight parameters associated with fault-sensitive features and other weight parameters respectively. At the same time, a loss function that integrates the weights of the fault stage is constructed, and differentiated loss weights are set for different samples according to the stage of the remaining life of the loom. The expression of the loss function is: ,

[0029] in, Let be the fault stage weight for the i-th sample. For the i-th sample, the actual remaining lifetime is... To predict remaining lifespan, The set of weights associated with fault-sensitive features. For other weight sets, and These are regularization systems for different weight sets.

[0030] Furthermore, the optimization of the life prediction model using a cross-model transfer learning method includes: selecting multiple benchmark models with known full life cycle data based on the life prediction model architecture; training a benchmark model containing a general fault feature extraction module and a model-specific adaptation module using the multi-dimensional feature set of the benchmark models; freezing the parameters of the general fault feature extraction module to retain the feature capture and focusing capabilities obtained from the life prediction model training; inputting limited data of the target model into the benchmark model and then fine-tuning only the parameters of the model-specific adaptation module; introducing a correction coefficient based on the ratio of rated power and operating speed between the target model and the benchmark model during fine-tuning; and iteratively updating the parameters using a small learning rate gradient descent method. The iterative update expression is: ,

[0031] in, To optimize the parameters of the dedicated adaptation module for the target model, The initial parameters for the adaptation module specific to the baseline model. To fine-tune the learning rate, The gradient of the loss function with respect to the baseline parameters. The training dataset for the target device model. This is the proportionality coefficient. and These are the rated power of the target model and the reference model, respectively. and These are the operating speeds of the target model and the reference model, respectively.

[0032] Secondly, a deep learning-based life prediction system for water-jet looms includes:

[0033] The data acquisition module is configured to acquire the operating data of the water jet loom through a multimodal sensor collaborative mechanism, and dynamically adjust the sensor sampling frequency based on the operating status of the water jet loom.

[0034] The preprocessing module is configured to preprocess the acquired running data, including denoising the running data using an adaptive wavelet threshold and introducing an attention mechanism to fill in missing values ​​to obtain preprocessed data.

[0035] The conversion module is configured to: extract features based on preprocessed data to obtain a multi-dimensional feature set, including extracting time-domain, frequency-domain and time-series features, and amplify the features of minor faults using a fault-sensitive feature enhancement method;

[0036] The model module is configured to: construct a lifetime prediction model using the extracted features as input, including using a multi-scale LSTM network as the model framework, embedding a feature attention layer based on the model framework and configuring an L2 regularization strategy.

[0037] The feature extraction module is configured to optimize the lifespan prediction model using a cross-model transfer learning method.

[0038] The transformation module is configured to output the remaining lifetime prediction value and uncertainty index based on the optimized lifetime prediction model, and to perform graded early warning of remaining lifetime based on the prediction results.

[0039] In summary, the present invention has the following beneficial technical effects:

[0040] 1. This invention achieves self-diagnosis of sensor faults and backup acquisition of associated data through a multi-modal sensor collaborative mechanism, avoiding data distortion caused by the failure of a single sensor. Combined with dynamic sampling frequency adjustment, the sampling strategy is adaptively adjusted according to the loom's operating load and the stability of key parts. While ensuring the complete capture of key fault signals, redundant data is reduced, solving the dual problems of missed key signals and data redundancy in traditional fixed sampling. This provides high-quality and efficient raw data support for subsequent analysis.

[0041] 2. This invention employs adaptive wavelet threshold denoising technology, dynamically adjusting the denoising threshold based on the signal noise intensity to avoid the problems of excessive denoising resulting in the loss of fault features or incomplete denoising leaving residual noise. It effectively preserves minute fault information, introduces an attention mechanism to fill in missing data values, and calculates the filled values ​​by combining time series correlation and cross-sensor data correlation. Compared with the traditional single interpolation method, it significantly improves the filling accuracy. At the same time, fidelity verification ensures that the preprocessed data truly reflects the loom's operating status, solving the problem of insufficient preprocessing accuracy caused by data noise interference and missing data.

[0042] 3. This invention uses a fault-sensitive feature enhancement method to amplify the differences in early-stage minor faults by integrating core features such as vibration, current, and temperature, thus solving the problem of low discrimination of early-stage faults by traditional features. Based on the correlation between features and lifespan, features are dynamically selected, and redundant features are eliminated by combining multicollinearity analysis. Highly correlated features are retained to form a multi-dimensional feature set, reducing the interference of irrelevant features on the model. This makes the input features more accurately reflect the lifespan decay law and solves the problems of feature redundancy and insufficient correlation.

[0043] 4. This invention constructs a multi-scale LSTM network framework, designing dedicated sub-networks for fault characteristics at different time scales (short-term impact, mid-term performance changes, and long-term aging) to accurately capture multi-dimensional fault patterns; embedding a feature attention layer to focus on key fault features, combined with hierarchical L2 regularization and fault stage weighted loss function, which not only suppresses model overfitting but also enhances the prediction accuracy for late-stage faults, solving the problem that traditional models cannot balance multi-scale feature learning and key stage prediction accuracy, and improving the prediction reliability throughout the entire life cycle.

[0044] 5. This invention adopts a cross-machine transfer learning method. Based on the model trained on a mature machine model, the general fault feature extraction module is frozen to retain the core capabilities. Only the machine model-specific adaptation module is fine-tuned to adapt to the differences of the target machine model. Combined with the machine model parameter correction coefficient, the adaptation accuracy is improved. This allows the model to maintain high prediction performance even when the target machine model data is limited, significantly reducing the model's dependence on the amount of target machine model data and expanding the application scenarios on different models of looms.

[0045] 6. This invention outputs the remaining life prediction value while quantifying the prediction uncertainty. Combined with a graded early warning mechanism, it divides the early warning level according to the prediction results and fault characteristics, and pushes targeted operation and maintenance suggestions. This provides operation and maintenance personnel with a reference for prediction reliability and clear handling strategies, avoids the blind decision-making caused by a single prediction value, shortens the fault response time, reduces production losses caused by sudden failure downtime, and solves the problem of insufficient practicality of traditional predictions that only output numerical values ​​and lack operation and maintenance guidance. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the overall process of a deep learning-based method for predicting the lifespan of a water jet loom, according to an embodiment of the present invention. Detailed Implementation

[0047] The present invention will be further described in detail below with reference to the accompanying drawings.

[0048] Example 1

[0049] Reference Figure 1 This embodiment of a deep learning-based method for predicting the lifespan of a water-jet loom includes:

[0050] S1. Obtain the operating data of the water jet loom through a multi-modal sensor collaborative mechanism, and dynamically adjust the sensor sampling frequency based on the operating status of the water jet loom;

[0051] S2. Preprocess the acquired running data, including denoising the running data using an adaptive wavelet threshold and introducing an attention mechanism to fill in missing values ​​to obtain preprocessed data.

[0052] S3. Based on the preprocessed data, feature extraction is performed to obtain a multi-dimensional feature set, including the extraction of time domain, frequency domain and time series features, and the fault-sensitive feature enhancement method is used to amplify the features of minor faults.

[0053] S4. Construct a lifetime prediction model using the extracted features as input, including using a multi-scale LSTM network as the model framework, embedding a feature attention layer based on the model framework and configuring an L2 regularization strategy.

[0054] S5. Optimize the life prediction model using a cross-aircraft transfer learning method;

[0055] S6. Based on the optimized life prediction model, output the remaining life prediction value and uncertainty index, and execute the remaining life graded early warning according to the prediction results.

[0056] Specifically, a deep learning-based method for predicting the lifespan of a water-jet loom includes the following:

[0057] S1. Obtain the operating data of the water jet loom through a multi-modal sensor collaborative mechanism, and dynamically adjust the sensor sampling frequency based on the operating status of the water jet loom;

[0058] like Figure 1As shown, the multimodal sensor encompasses four core sensor types: vibration sensors, temperature sensors, pressure sensors, and current sensors. These are deployed in key monitoring components such as the loom motor, transmission system, and water jet system. Firstly, based on the historical fault database of the water jet loom, the correlation characteristics of multi-sensor signals during faults in each key monitoring component are analyzed. Cross-sensor data correlation rules are extracted and a rule base is constructed. Specifically, for the motor bearing component, historical data shows that before wear failure occurs, the vibration sensor exhibits impact vibration signals in a specific frequency range (e.g., 200-500Hz), accompanied by a slow increase in the temperature sensor's monitored value and an increase in the fluctuation amplitude of the current sensor's monitored value. Therefore, in the rule... A master-slave relationship is established in the database between motor bearing vibration sensors, motor bearing temperature sensors, and motor bearing current sensors. This relationship clearly identifies motor bearing wear as the corresponding fault type and labels signal association thresholds, such as vibration impact amplitude exceeding the normal average by 15%, temperature rise rate exceeding 0.03℃ / minute, and current fluctuation exceeding the normal average by 10%. For the water sprinkler system, historical data shows that impeller wear faults in the water sprinkler pump cause a sharp drop in pressure sensor readings, accompanied by a simultaneous decrease in flow sensor readings. Therefore, a master-slave relationship is established between the water sprinkler system pressure sensor and the water sprinkler system flow sensor in the rule base, labeling association thresholds (pressure drop exceeding 0.1MPa, flow drop exceeding 10%). The rule base is stored in the data processing terminal as a structured data table, containing fields such as master sensor ID, a list of associated sensor IDs, corresponding fault type, signal association threshold, and update time. It supports quarterly rule iteration updates based on new fault data to ensure the timeliness and accuracy of the association relationships.

[0059] The data processing terminal performs statistical analysis on the real-time data collected by each main sensor at fixed time intervals (e.g., 1 minute), and calculates the data validity coefficient H using a preset algorithm. The calculation formula is as follows: ,

[0060] in, The number of valid data points collected by the sensor per unit time. The total number of data points that the sensor should collect per unit time is determined by the current sampling frequency, i.e., sampling frequency × time period duration. This is the real-time average of the data collected by the sensor. The average calibration standard of the sensors at the factory is used as the benchmark. Then, an effectiveness coefficient threshold is set. When the calculated result H of a certain main sensor is lower than the threshold, it is determined that the main sensor has performance drift or hardware failure, and the collaborative acquisition mechanism is immediately triggered: On the one hand, the data processing terminal sends an acquisition command to the associated sensor corresponding to the main sensor. The command contains the acquisition parameters of the main sensor. The associated sensor replaces the main sensor to perform the data acquisition task, ensuring the continuity of the operating data of key parts; on the other hand, the data processing terminal automatically generates sensor maintenance alarm information, which includes the ID of the failed main sensor, the installation location, the failure time, and the current H value, reminding the failed sensor to be maintained.

[0061] The data processing terminal receives real-time data from sensors at key locations and determines the current operating load level and stability of key components based on preset logic. The operating load level is determined by combining the effective value of the motor current and the water spray pressure. First, the effective value of the current and the water spray pressure are compared with historical normal operating data to divide their respective load ranges (low, medium, and high). Then, the higher of the two load ranges is taken as the final operating load level, and a corresponding operating load coefficient L is assigned. Different load levels correspond to different L values, with higher loads resulting in larger L values, reflecting the influence of the load on the sampling frequency. The stability of key components is determined by the variance of the vibration signal at that component. The variance of the vibration data over the past 10 time periods is calculated and compared with the mean variance under historical normal operating conditions to divide the components into three levels: stable, generally stable, and unstable, and a corresponding operating stability coefficient S is assigned. The mean variance is obtained from historical normal operating data and reflects the vibration stability level of the loom during normal operation.

[0062] Based on the determined operating load factor L and operating stability factor S, the formula is: ,

[0063] in, The sampling frequency is the one after the sensor is dynamically adjusted. This is the sensor's fundamental sampling frequency. The weighting of the operating load effect is given by L, which is the real-time operating load coefficient of the loom. Weights that affect operational stability To ensure the operational stability coefficient of critical components, the target sampling frequency of each sensor is calculated. To avoid frequent fluctuations in sampling frequency that could lead to sensor hardware wear or data format corruption, the data processing terminal is equipped with boundary control logic: First, the amplitude of two consecutive sampling frequency adjustments must not exceed 20% of the current sampling frequency. If the difference between the calculated target sampling frequency and the current frequency exceeds this amplitude, the adjustment is made at the maximum amplitude of 20% to avoid sudden increases or decreases in frequency. Second, the target sampling frequency must not exceed the sensor's rated sampling frequency range. If the calculated result exceeds the range, the boundary value of the rated range is taken as the final sampling frequency.

[0064] S2. Preprocess the acquired running data, including denoising the running data using an adaptive wavelet threshold and introducing an attention mechanism to fill in missing values ​​to obtain preprocessed data.

[0065] For the vibration signal with the most significant noise interference in the data collected in step S1 (affected by electromagnetic coupling and mechanical resonance), adaptive wavelet threshold denoising is performed. The specific process is as follows: 1. Multi-scale wavelet decomposition: The preset wavelet basis function (selected based on the non-stationary characteristics of the loom vibration signal, which has the advantage of time-frequency localization) is called to perform multi-scale wavelet decomposition on the vibration signal at a preset scale (determined according to the signal sampling frequency and the fault characteristic frequency range). After decomposition, two types of coefficients are obtained, namely low-frequency approximation coefficients. It reflects the fundamental trend and effective characteristics of vibration signals, and high-frequency detail coefficients. It mainly contains noise components and has high-frequency impact characteristics mixed with minor faults;

[0066] 2. Real-time signal-to-noise ratio calculation: Based on the decomposition coefficients, the signal power and noise power are calculated, where the signal power... noise power Then through the formula Calculate the real-time signal-to-noise ratio, quantifying the ratio of effective features to noise in the current signal, where, These are low-frequency approximation coefficients. For high-frequency detail coefficients;

[0067] 3. Adaptive threshold generation: First, a heuristic method is used to calculate the base threshold. The formula is ,in, The noise standard deviation is estimated using the median absolute deviation of the high-frequency detail coefficients, where N is the total number of data points in the vibration signal. An adaptive threshold is then constructed by combining this with the real-time signal-to-noise ratio. The two are negatively correlated, as shown in the formula. k is an adjustment coefficient that enables dynamic adaptation where the threshold is large when the noise is strong and small when the noise is weak.

[0068] Finally, the high-frequency detail coefficients The absolute value is less than The coefficients are set to zero (to remove redundant noise), and only those with absolute values ​​greater than 0 are retained. The coefficients are used to preserve minor fault characteristics, and then the processed values ​​are... and Perform inverse wavelet transform The denoised vibration signal is reconstructed; for temperature, pressure and current data, the moving average filtering method is used to denoise, ensuring that the noise interference of various data meets the requirements of subsequent processing.

[0069] Next, attention mechanism is used to fill in missing values. The location information of the missing value includes the data type and the corresponding time step t. Two types of associated data are selected according to preset association rules: First, time series data of the same type in consecutive time steps before and after the missing value. For example, if the temperature at time t is missing, the temperature data at times t−1, t−2, ..., t−n and t+1, t+2, ..., t+m are selected, where n and m are the preset number of time steps. Second, cross-sensor associated data at the same time step as the missing value. For example, if the temperature at time t is missing, the motor current and transmission system vibration data at time t are selected. This type of data has an operating condition correlation with temperature.

[0070] Next, attention weights are calculated and imputation values ​​are obtained. First, the Pearson correlation coefficient is used to calculate the similarity between associated data points and missing values, using the following formula: ,in, For the j-th data point of the i-th type of associated data, For the complete time series containing the missing values, The closer the similarity is to 1, the stronger the correlation; then, the attention weights are obtained by normalizing the similarity using the softmax function. ,in, Let the number of points be the i-th type of associated data, ensuring that the sum of the weights of all associated data points is 1; at the same time, set the data type weights. Time series data of the same type have higher weights than cross-sensor data because they have stronger temporal continuity. Finally, the imputation value formula is used:

[0071] ,

[0072] Where K is the number of associative data types, Let i be the number of points in the i-th type of associated data. Attention weights for each associated data point, For the j-th data point of the i-th type of associated data, The data type weights are used; missing values ​​are calculated, and after imputation, the time series continuity of all data is checked. If there are unfilled missing values ​​due to insufficient related data, the mean imputation method is used.

[0073] To address the dimensional differences in different data types, standardization was performed, and data fidelity was verified. The Z-score standardization method was used to map all data to a scale with a mean of 0 and a standard deviation of 1. To verify the accuracy of the pre-processed data in reflecting the loom's condition, the fidelity F was calculated using the following formula: ,

[0074] in, , and These are the weighting coefficients. This represents the mean square error of the signal before and after denoising. This is the actual data after missing values ​​have been collected. To account for the trend deviation between the standardized data and the original data, This represents the original vibration signal without past noise processing. This represents the baseline value of the trend deviation of the original data. When F is higher than the preset threshold, the preprocessed data is deemed qualified and output to step S3. If F does not meet the standard, the noise reduction threshold or the range of associated data is readjusted until the fidelity meets the requirements.

[0075] S3. Based on the preprocessed data, feature extraction is performed to obtain a multi-dimensional feature set, including the extraction of time domain, frequency domain and time series features, and the fault-sensitive feature enhancement method is used to amplify the features of minor faults.

[0076] The preprocessed data output in step S2 is used to extract time-domain, frequency-domain, and time-series features to construct a basic feature set. The specific process is as follows:

[0077] Time-domain feature extraction: For vibration and current data, extract time-domain features that reflect the statistical characteristics of the signal, including the mean. ,variance Peak indicators Crest factor ,in, For data points, N is the total number of data points. The Crest factor reflects the ratio of the signal peak value to the effective value. For pressure data, the peak value, valley value and fluctuation amplitude (peak value - valley value) are extracted to reflect the pressure stability of the water spray system.

[0078] Frequency domain feature extraction: Perform a Fast Fourier Transform (FFT) on the vibration data to convert the time-domain signal into a frequency-domain signal. The formula is as follows: ,in, For time-domain data, For frequency domain data, k represents a frequency point. Extracting frequency domain features includes characteristic frequency amplitudes (corresponding to fault characteristic frequencies of key components of the loom, such as the fault frequency of the outer ring of the motor bearing) and power spectral density. ,in, The sampling frequency reflects the distribution of signal power across various frequencies and is used to locate the frequency range of noise and fault characteristics. The centroid frequency is... ,in, The frequency point reflects the frequency location of power concentration, and the center frequency shifts during a fault.

[0079] Time series feature extraction: For continuous data such as temperature and current that change over time, extract time series features including the rate of temperature change. ,in, Let be the temperature at time t. The time interval, the mean value of the current sliding window (calculated over a preset time window to reflect the current load trend), and the trend slope are obtained by linearly fitting the time series data, using the following formula: , Timestamps reflect long-term trends in data. The slope is the trend, and N is the total number of time series data points. The actual value of the i-th current data point is used to integrate the above time-domain, frequency-domain, and time-series features to form a basic feature set.

[0080] To address the issue of low identification of minor fault features (such as subtle changes in peak indicators of early bearing wear) in the basic feature set, fault-sensitive feature enhancement is implemented. First, normal state baseline values ​​are statistically analyzed. The historical normal operation database of water-jet looms (containing data from at least three complete normal operation cycles) is accessed. The normal state mean values ​​for the peak indicator, Crest factor, and temperature change rate in the basic features are calculated separately to obtain the normal mean value of the peak indicator. Crest factor normal mean Normal average temperature change rate This serves as a benchmark for determining whether a feature is abnormal.

[0081] Next, fault sensitivity factors are calculated, and the peak index PI, crest factor CF, and temperature change rate TCR are extracted from the basic feature set. Weighting coefficients are then set for each of the three. The weight coefficients are determined through random forest feature importance evaluation; features with higher fault sensitivity have larger weights, and satisfy the following condition: Through the formula: ,

[0082] Calculate the fault sensitivity factor S, where, The peak index is extracted from the time-domain characteristics of the current vibration signal. This represents the normal average value of the peak indicator. For the current Crest factor, This represents the normal mean of the Crest factor. The current rate of temperature change. The normal average temperature change rate is used as a factor to quantify and amplify the differences in minute fault characteristics. For example, when PI rises slightly during early wear, the difference is amplified and reflected in S after being calculated by comparing it with the normal average. Then, an enhanced feature set is constructed, and the fault sensitivity factor S is added as a new feature. It is associated with the time domain, frequency domain, and time series features in the basic feature set. For example, S is bound to the characteristic frequency amplitude and current trend slope of the corresponding time step to form an enhanced feature set.

[0083] Finally, feature selection and multi-dimensional feature set determination are performed. The correlation coefficient Corr between each feature in the enhanced feature set and the remaining lifespan Y of the loom is calculated. The correlation coefficient calculation formula is as follows: ,

[0084] in, Let x be the mutual information entropy between feature x and lifetime Y. The Spearman rank correlation coefficient between feature x and lifetime Y. To quantify the nonlinear correlation, the calculation formula is as follows: , For joint probability distribution, , To determine the probability distribution of the relationship, a correlation threshold is set. Preserve the enhanced feature set By extracting the features and removing features with low correlation, a multi-dimensional feature set with no redundancy and high lifespan correlation is finally obtained. Through the above multi-dimensional basic feature extraction and fault-sensitive feature enhancement, the operating status features of the loom in different dimensions can be effectively captured, the identification of minor fault features can be amplified, and redundant features can be eliminated.

[0085] S4. Construct a lifetime prediction model using the extracted features as input, including using a multi-scale LSTM network as the model framework, embedding a feature attention layer based on the model framework and configuring an L2 regularization strategy.

[0086] Based on the physical process of the failure of key components in water jet looms from occurrence to deterioration, the multi-dimensional feature set is divided into three categories according to time scale: short-term fault feature subset, medium-term fault feature subset, and long-term fault feature subset. The short-term fault feature subset corresponds to sudden faults, such as transmission jamming caused by yarn entanglement or instantaneous gear tooth breakage during meshing. Its feature change cycle is short, and in this embodiment, it is set within 10 sampling periods, including features such as vibration impact peak value and sudden current rise amplitude. The medium-term fault feature subset corresponds to gradual faults, such as initial wear of motor bearings and slight wear of the water jet pump impeller. Its feature change cycle is moderate, and in this embodiment, it is set within 30 sampling periods, including features such as slow temperature rise rate and pressure fluctuation standard deviation. The long-term fault feature subset corresponds to chronic aging faults, such as aging of motor winding insulation and metal fatigue of the transmission system. Its feature change cycle is long, and in this embodiment, it is set within 60 sampling periods, including features such as the long-term trend slope of the effective current value and the centroid frequency shift of the vibration power spectral density.

[0087] For each of the three feature subsets, an LSTM subnetwork is constructed, with the parameters of each subnetwork adapted to the corresponding feature scale. The short-term LSTM subnetwork takes the number of features in the short-term feature subset as its input dimension. Time step Matching short-term feature cycles, number of neurons in the hidden layer The weights of the forget gate, input gate, and output gate are initialized using a He normal distribution, and the bias term is initialized to 0.1.

[0088] Mid-term LSTM sub-network: The input dimension is the number of features in the mid-term feature subset, denoted as . Time step Number of neurons in the hidden layer The weights and biases are initialized the same as in the short-term subnetwork. For the long-term LSTM subnetwork: the input dimension is the number of features in the long-term feature subset, set to... Time step Number of neurons in the hidden layer The weights and biases are initialized the same as in the short-term subnetwork, and the cell state update formulas for each subnetwork follow the basic principles of LSTM.

[0089] To address the issue of uneven generalization ability of traditional fixed-probability dropout on features with varying sparsity, a dynamic dropout mechanism is introduced between the hidden and output layers of each LSTM sub-network. The dropout probability of this mechanism is dynamically adjusted according to the sparsity of the input features. The formula for calculating the dynamic dropout probability is as follows: ,

[0090] in, Based on dropout probability, Feature sparsity is defined as the proportion of zero-value features in the feature set to the total number of features. , The zero-valued characteristic number, The total number of features is given. The dynamic dropout mechanism works by the following principle: when a short-term feature subset has few data samples due to a sudden failure, resulting in high feature sparsity, then... The sparsity is lower, reducing the loss of effective features; the long-term feature subset has low feature sparsity due to sufficient samples of chronic fault data, resulting in correspondingly lower sparsity. The relatively high performance enhances the ability to suppress overfitting and ensures that each sub-network maintains good generalization performance under different feature sparsity conditions.

[0091] Next, activation function optimization was performed. The output layer activation function of each LSTM sub-network adopted the tanh function with an introduced temperature coefficient. This optimized the traditional tanh function, which is prone to gradient saturation when the feature values ​​vary greatly. The expression of the tanh function with the introduced temperature coefficient is as follows: ,

[0092] Where T is the temperature coefficient. The linear computation result of the output layer neurons of the LSTM subnetwork is dynamically set according to the feature value range. For features with a large value range (such as vibration peak index, in μm, with a value range usually greater than 10), T is set to 1.2 to widen the linear interval of the activation function, avoid gradient saturation due to excessively large feature values, and ensure that the model can effectively learn changes in such features. For features with a small value range (such as temperature change rate, in °C / min, with a value range usually less than 1), T is set to 0.8 to enhance the sensitivity of the activation function to small feature changes, ensure that the model can capture early signs of failure reflected by such features, and ultimately achieve effective response of each subnetwork to features with different value ranges, thereby improving the accuracy of feature learning.

[0093] To address the issue of insufficient attention to key features (such as fault sensitivity factor correlation features) in the output features of multi-scale LSTM sub-networks, a feature attention layer is set between the output of each LSTM sub-network and the fully connected layer. A bidirectional gating mechanism is used to calculate the correlation weights between features and remaining lifetime, and the features are dynamically weighted based on these weights. The feature attention layer comprises three core modules: a linear transformation module, a bidirectional gating calculation module, and a weight normalization module. The linear transformation module is used to transform the hidden states of each LSTM sub-network output and the short-term sub-network output... Mid-term sub-network output Long-term subnetwork output And they are all unified to the same dimension, which is set to 64 in this embodiment to avoid weight calculation deviations caused by differences in feature dimensions. The linear transformation formula is: , For dimension The weight matrix, This corresponds to the number of neurons in the hidden layer of the sub-network. As a bias term of dimension 64, the bidirectional gating calculation module is used to fuse the fault sensitivity factor of step S3, calculate the correlation between each feature and the remaining life of the loom, and ensure that the correlation calculation takes into account both the feature's own attributes and the fault risk.

[0094] Features after linear transformation The input is a bidirectional gated computation module, using trainable parameters from the attention layer. and Perform a linear mapping to obtain the linearly mapped values ​​of the features. This value initially reflects the correlation potential between the feature itself and the remaining lifetime; subsequently, the fault sensitivity factor S calculated in step S3 is introduced, and the feedback adjustment coefficient is used. Construct a two-way gating factor, calculated using the following formula: This factor, expressed as a feature linear mapping value multiplied by a fault risk adjustment term, bidirectionally adjusts the correlation between a feature and its remaining lifetime. When the feature linear mapping value is significant and the S-value is large, the bidirectional gating factor increases, indicating a high correlation between the feature and its remaining lifetime. Conversely, when the feature linear mapping value is weak or the S-value is small, the bidirectional gating factor decreases, indicating a low correlation. Next, the bidirectional gating factor is input into a weight normalization module, and the softmax function is used to normalize the bidirectional gating factors of all features, yielding the attention weights for each feature. The normalization calculation formula strictly follows the technical solution requirements, namely: ,

[0095] in, Let be the attention weight for the i-th feature. This is the i-th feature value output by the multi-scale LSTM, which here corresponds to the linearly transformed... , and These represent different trainable parameters of the attention layer. As a fault-sensitive factor, Here, N is the feedback adjustment coefficient, and N is the total number of features. This formula amplifies the difference in bidirectional gating factors between highly correlated and low-correlation features through an exponential function. Then, it uses summation and normalization to ensure that the sum of attention weights for all features is 1, conforming to the basic logic of weight allocation and avoiding feature attention imbalance caused by excessively high or low weights for any particular feature. Finally, the features with unified dimensions are... Used for corresponding attention weights Perform element-wise multiplication to obtain the weighted features. Weighted feature set As input to the fully connected layer, this enhances key features while suppressing non-key features. Experiments have shown that the average attention weight of the fault-sensitive factor associated features after weighting is approximately 0.2, significantly higher than the average of 0.05 for ordinary features, ensuring that key features are given sufficient attention in subsequent model calculations.

[0096] To prevent overfitting due to excessive parameters in the fully connected layer during model training, and to avoid the problem of inconsistent penalty for key and non-key features by traditional single L2 regularization, a layered L2 regularization term is added to the weight parameters of the fully connected layer, and a dynamic adjustment mechanism for weight decay is introduced. The fully connected layer adopts a two-layer structure design. The first layer has an input dimension of the feature dimension (64) output by the feature attention layer, and an output dimension of 32 (used to further map the attention-weighted features to low-dimensional features, reducing the computational complexity of the model); the second layer has an input dimension of 32 and an output dimension of 1, directly outputting the predicted value of the remaining life of the loom, realizing end-to-end prediction. Based on the correlation between the weight parameters of the fully connected layer and the fault-sensitive features, the weight parameters are divided into two sets: (The set of weights associated with fault-sensitive features) and (Other weight sets), where, This includes the weights of the features after linear transformation of the corresponding fault-sensitive factors in the first layer, and the weights of the connections between the output nodes of the fault-sensitive features in the second layer and the first layer. These weights directly affect the model's judgment of fault risk and are crucial to the accuracy of life prediction. Therefore, their learning ability needs to be protected. This includes the weights in the first layer corresponding to ordinary time domain, frequency domain, and time series features, and the weights in the second layer connected to the output nodes of ordinary features in the first layer. The feature redundancy associated with these weights is relatively high, and overfitting can be suppressed through strong regularization.

[0097] The differential L2 regularization coefficients of the two weight sets were determined using a 5-fold cross-validation method. ( The corresponding regularization coefficient is set to 0.001. ( The corresponding regularization coefficient is set to 0.005. Less than The reason is that, The associated fault-sensitive features have a significant impact on prediction accuracy; a smaller regularization coefficient can reduce the penalty to its parameters and avoid the loss of key learning capabilities; while The associated common features have high redundancy, and a large regularization coefficient can effectively suppress parameter overfitting, ensuring a balance between fitting ability and generalization ability. Subsequently, a dynamic adjustment mechanism of weight decay is introduced during model training. The regularization coefficient is adjusted according to the loss trend of the model on the validation set. The adjustment rule is as follows: if the validation set loss decreases for three consecutive rounds, it indicates that the model is still effectively learning new feature patterns. At this time, the regularization coefficient is updated to 0.9 times the original coefficient (reducing the regularization strength, allowing the model to further fit effective features); if the validation set loss increases for two consecutive rounds, it indicates that the model is overfitting. At this time, the regularization coefficient is updated to 1.1 times the original coefficient (increasing the regularization strength, suppressing excessive parameter growth); in other cases, the regularization coefficient remains unchanged. This dynamic adjustment mechanism can ensure that the fully connected layer maintains a good parameter state at different stages of training, without losing effective learning results or overfitting.

[0098] Based on statistical analysis of the full lifecycle operation data of water jet looms, the remaining lifespan Y of the looms is divided into three stages, and corresponding weights are set for each sample failure stage. The early stage is the remaining lifespan. During this stage, the fault characteristics are not obvious, and even if the prediction has some deviation, the impact on operation and maintenance work is relatively small. Therefore, it is set... The intermediate stage is During this stage, fault characteristics gradually emerge, and prediction errors may lead to the maintenance plan being implemented ahead of schedule or behind schedule. The late stage is During this stage, the fault characteristics are significant, and prediction errors can easily lead to sudden machine shutdowns, causing serious losses such as production interruptions and order delays. Therefore, a system is set... The weight values ​​are determined based on actual operation and maintenance loss statistics. That is, by analyzing data such as downtime, maintenance costs, and production losses caused by prediction deviations at different fault stages, it is determined that the comprehensive loss of prediction deviations in the late stage is about 2.5 times that in the early stage, ensuring that the weight settings meet the actual application requirements.

[0099] Finally, the total loss function is calculated. The total loss function consists of two parts: a weighted mean squared error loss, reflecting the deviation between the predicted and actual values ​​and implementing a differentiation stage penalty; and a hierarchical L2 regularization loss, reflecting the parameter complexity of the fully connected layer and suppressing overfitting. ,

[0100] in, Let be the fault stage weight for the i-th sample. For the i-th sample, the actual remaining lifetime is... To predict remaining lifespan, The set of weights associated with fault-sensitive features. For other weight sets, and They are respectively and The corresponding regularization system, this function through To achieve differentiated penalties for sample bias at different fault stages, an end-to-end lifetime prediction model is constructed by building the multi-scale LSTM framework, embedding feature attention layers, configuring hierarchical L2 regularization, and constructing a weighted loss function for each fault stage.

[0101] S5. Optimize the life prediction model using a cross-aircraft transfer learning method;

[0102] Select 3-5 aircraft models from the same series as the target model that have complete lifecycle data as benchmark models. Ensure that the benchmark models and the target model are identical in core component structure and principle, and that the fault evolution patterns are transferable. Collect multi-dimensional feature sets and corresponding real remaining lifecycle data for each benchmark model. The multi-dimensional feature sets are consistent with the feature types extracted in step S3, including short-term, medium-term, and long-term fault features and fault sensitivity factors. Construct a benchmark training dataset. Based on the multi-scale LSTM lifetime prediction model in step S4, the benchmark training dataset is used. The input model is jointly trained. The training process uses the loss function of the fusion fault stage weights defined in step S4. The Adam algorithm is selected as the transformer. The iteration stops when the validation set loss does not decrease for 10 consecutive rounds, and a converged benchmark model is obtained. This benchmark model has the ability to extract and predict the general fault characteristics of water jet looms by learning the common fault patterns of multiple benchmark looms.

[0103] The baseline model is split into two core modules: a general fault feature extraction module and a model-specific adaptation module. The general fault feature extraction module includes the multi-scale LSTM sub-networks (short-term, medium-term, and long-term LSTM sub-networks) and feature attention layers from step S4. This module is responsible for extracting cross-model common fault evolution features (such as vibration and shock, temperature rise, current changes, and other common features directly related to fault types) from the input features. Its parameters have reusable value across different models. The model-specific adaptation module includes the fully connected layer and layered L2 regularization parameters from step S4. This module is responsible for mapping general features to the remaining life prediction values ​​for specific models. Its parameters are strongly correlated with the physical attributes of the model (such as rated power and operating speed) and need to be adapted separately for different models. The module split is achieved by freezing the parameter interfaces of the general module and the specific module, ensuring that the two modules are both independent and can work collaboratively.

[0104] To preserve the general fault feature capture and focusing capabilities learned by the baseline model, the parameters of the general fault feature extraction module are frozen. Limited data from the target model are input into the baseline model. Based on the differences in core physical parameters between the target and baseline models, correction coefficients are calculated to quantify the impact of model attributes on remaining life prediction. The calculation basis for the correction coefficients is the ratio of the rated power of the target and baseline models. Ratio of operating speed The device-specific adaptation module parameters are iteratively updated using a small learning rate gradient descent method, with an initial learning rate of... Set to 0.0001, which is 1 / 10 of the training learning rate of the baseline model, to avoid drastic fluctuations in parameters due to insufficient target data and to ensure the stability of fine-tuning.

[0105] The iterative update process takes the common fault feature vector of the target aircraft model as input and the remaining life label of the target aircraft model as the supervision signal, and calculates the loss function. Current parameters of the dedicated adaptation module gradient By dynamically adjusting the parameter update step size in conjunction with the correction coefficient, the iterative update expression strictly adheres to the technical solution requirements, namely... ,

[0106] in, To optimize the parameters of the dedicated adaptation module for the target model, The initial parameters for the adaptation module specific to the baseline model. To fine-tune the learning rate, The gradient of the loss function with respect to the baseline parameters. The training dataset for the target device model. This is the proportionality coefficient. and These are the rated power of the target model and the reference model, respectively. and The operating speeds of the target and baseline models are respectively. During the iterative update process, the loss value of the target model's validation set is calculated after each round of fine-tuning. Fine-tuning stops when the validation set loss decreases by less than 1e-4 for five consecutive rounds. The parameters are for a specific adaptation module for the target device model after optimization.

[0107] The general fault feature extraction module of the frozen parameters is reintegrated with the model-specific adaptation module after fine-tuning to form an optimized life prediction model suitable for the target model. The integrated model retains the general fault feature learning results of the benchmark model, while adapting to the physical attributes and limited data patterns of the target model. It can be directly used for the remaining life prediction of the target model, solving the model accuracy problem caused by the scarcity of data for the target model.

[0108] S6. Based on the optimized life prediction model, output the remaining life prediction value and uncertainty index, and execute the remaining life graded early warning according to the prediction results.

[0109] The multi-dimensional feature set of the target aircraft model (consistent with the feature types extracted in step S3, including short-term, medium-term, and long-term fault features and fault sensitivity factors) is input into the optimized lifetime prediction model. The model captures common fault patterns through a general fault feature extraction module, and maps them to the remaining lifetime prediction value of the target aircraft model through a model-specific adaptation module. To avoid decision-making blindness caused by a single prediction value, an uncertainty index is calculated based on the probability distribution characteristics of the model output layer. The "prediction variance" is used to quantify the degree of uncertainty, and the calculation formula is as follows: Where M is the number of predicted samples generated by the model through the Dropout ensemble method. Let m be the remaining lifetime value predicted in the m-th prediction. The mean of M predicted values; the uncertainty index ranges from... A smaller value indicates a more stable and reliable prediction result, while a larger value indicates greater uncertainty in the model's judgment of the current loom condition. This is based on the remaining life prediction value. With uncertainty indicators The warning levels are divided into three levels:

[0110] Level 1 Warning (Low Risk): When triggered, it indicates that the remaining lifespan of the loom is sufficient and the prediction is reliable. At this time, routine inspection suggestions are pushed out, and maintenance personnel will inspect the key components of the loom every 7 days as originally planned.

[0111] Level 2 Warning (Medium Risk): When triggered, it indicates that the remaining lifespan of the loom has entered the planned maintenance cycle, and the predicted reliability is moderate. At this time, a planned maintenance suggestion is pushed out, and the maintenance personnel need to formulate a maintenance plan within 1 week.

[0112] Level 3 Warning (High Risk): When When triggered, it indicates that the remaining lifespan of the loom is running low or the forecast is highly uncertain, posing a risk of sudden failure. At this time, an emergency shutdown and maintenance suggestion is pushed out. By clarifying the maintenance priority under different scenarios through graded early warning, the probability of sudden failure of the loom is effectively reduced, and the stability of equipment operation and production continuity are improved.

[0113] Example 2

[0114] The difference between this embodiment and Embodiment 1 is that this embodiment provides a deep learning-based water jet loom life prediction system, including:

[0115] The data acquisition module is configured to acquire the operating data of the water jet loom through a multimodal sensor collaborative mechanism, and dynamically adjust the sensor sampling frequency based on the operating status of the water jet loom.

[0116] The preprocessing module is configured to preprocess the acquired running data, including denoising the running data using an adaptive wavelet threshold and introducing an attention mechanism to fill in missing values ​​to obtain preprocessed data.

[0117] The conversion module is configured to: extract features based on preprocessed data to obtain a multi-dimensional feature set, including extracting time-domain, frequency-domain and time-series features, and amplify the features of minor faults using a fault-sensitive feature enhancement method;

[0118] The model module is configured to: construct a lifetime prediction model using the extracted features as input, including using a multi-scale LSTM network as the model framework, embedding a feature attention layer based on the model framework and configuring an L2 regularization strategy.

[0119] The feature extraction module is configured to optimize the lifespan prediction model using a cross-model transfer learning method.

[0120] The transformation module is configured to output the remaining lifetime prediction value and uncertainty index based on the optimized lifetime prediction model, and to perform graded early warning of remaining lifetime based on the prediction results.

[0121] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for predicting the lifespan of a water-jet loom, characterized in that, include: S1. Obtain the operating data of the water jet loom through a multi-modal sensor collaborative mechanism, and dynamically adjust the sensor sampling frequency based on the operating status of the water jet loom; S2. Preprocess the acquired running data, including denoising the running data using an adaptive wavelet threshold and introducing an attention mechanism to fill in missing values ​​to obtain preprocessed data. S3. Based on the preprocessed data, feature extraction is performed to obtain a multi-dimensional feature set, including the extraction of time domain, frequency domain and time series features, and the fault-sensitive feature enhancement method is used to amplify the features of minor faults. S4. Construct a lifetime prediction model using the extracted features as input, including using a multi-scale LSTM network as the model framework, embedding a feature attention layer based on the model framework and configuring an L2 regularization strategy. S5. Optimize the life prediction model using a cross-aircraft transfer learning method; S6. Based on the optimized life prediction model, output the remaining life prediction value and uncertainty index, and execute the remaining life graded early warning according to the prediction results.

2. The method for predicting the lifespan of a water-jet loom based on deep learning according to claim 1, characterized in that, The process of acquiring operational data of the water-jet loom through a multimodal sensor collaborative mechanism includes: establishing a cross-sensor data association rule base based on the signal correlation characteristics of faults occurring at key monitoring parts of the water-jet loom; determining the associated sensors for each sensor based on the association rule base; calculating the data validity coefficient of each sensor based on the deviation between the real-time average and the standard average of the sensor-collected data; and when the validity coefficient is lower than a set threshold, the associated sensor performs the operational data acquisition task and simultaneously generates and sends sensor maintenance alarm information. The formula for calculating the data validity coefficient is as follows: ， in, The number of valid data points collected by the sensor per unit time. This represents the total number of data points that the sensor should collect per unit of time. This is the real-time average of the data collected by the sensor. This is the average value of the sensor's calibration standard at the time of manufacture.

3. The method for predicting the lifespan of a water-jet loom based on deep learning according to claim 1, characterized in that, The method of dynamically adjusting the sensor sampling frequency based on the operating status of the water jet loom includes determining the current operating status of the water jet loom based on the actual operating data, constructing a dynamic sampling frequency adjustment mechanism by combining historical operating data of key parts of the loom, and adaptively adjusting the sampling frequency of each sensor according to changes in the loom's operating load and the operational stability of key parts. The expression for adjusting the frequency is as follows: ， in, The sampling frequency is the one after the sensor is dynamically adjusted. This is the sensor's fundamental sampling frequency. The weighting of the operating load effect is given by L, which is the real-time operating load coefficient of the loom. Weights that affect operational stability This refers to the operational stability coefficient of critical components.

4. The method for predicting the lifespan of a water-jet loom based on deep learning according to claim 1, characterized in that, The step of denoising the running data using an adaptive wavelet threshold includes performing multi-scale wavelet decomposition on the vibration signal in the running data according to a preset wavelet basis function to obtain low-frequency approximation coefficients and high-frequency detail coefficients. The signal power and noise power are calculated based on the low-frequency approximation coefficients and high-frequency detail coefficients, and the real-time signal-to-noise ratio is calculated. A basic threshold is calculated using a heuristic method, and an adaptive threshold is generated by combining the real-time signal-to-noise ratio. Coefficients in the high-frequency detail coefficients whose absolute values ​​are less than the adaptive threshold are set to zero, while coefficients with values ​​greater than the threshold are retained to preserve valid fault features. An inverse wavelet transform is performed on the processed low-frequency approximation coefficients and high-frequency detail coefficients to reconstruct the denoised running data.

5. The method for predicting the lifespan of a water-jet loom based on deep learning according to claim 4, characterized in that, The introduction of an attention mechanism to impute missing values ​​and obtain preprocessed data specifically includes the following steps: S21. Determine the associated data corresponding to the missing value, including time series data of the same type at consecutive time steps before and after the missing value, and cross-sensor associated data at the same time step as the missing value. S22. The similarity between the associated data and the missing values ​​is calculated using the Pearson correlation coefficient, and the attention weight of each associated data point is calculated based on the similarity. Data type weights are set, and the imputation value of the missing value is calculated in combination with the attention weights. S23. Perform standardization processing on the denoised and missing value imputation running data, and calculate the fidelity of the preprocessed data by the signal distortion after denoising, the missing value imputation deviation and the standardization data deviation. The formula for calculating the fill value is: ， Where K is the number of associative data types, Let i be the number of points in the i-th type of associated data. Attention weights for each associated data point, For the j-th data point of the i-th type of associated data, Weights for data types.

6. The method for predicting the lifespan of a water-jet loom based on deep learning according to claim 1, characterized in that, The method of amplifying minor fault features using fault-sensitive feature enhancement includes: statistically analyzing historical normal operation data of the loom to obtain the average peak index, average Crest factor, and average temperature change rate under normal conditions; setting weight coefficients and extracting the peak index PI and Crest factor from the time-domain feature set; extracting the temperature change rate from the time-series feature set; calculating the fault-sensitive factor based on the weight coefficients and correlating it with time-domain, frequency-domain, and time-series features to form an enhanced feature set; calculating the correlation degree between each feature in the enhanced feature set and the remaining life of the loom; and filtering features based on the correlation degree to obtain a multi-dimensional feature set. The correlation degree calculation formula is as follows: ， in, Let x be the mutual information entropy between feature x and lifetime Y. is the Spearman rank correlation coefficient between feature x and lifetime Y.

7. The method for predicting the lifespan of a water-jet loom based on deep learning according to claim 1, characterized in that, The process of constructing a lifespan prediction model using extracted features as input includes classifying the extracted features based on the time-scale differences of different fault features of the water-jet loom, constructing an LSTM sub-network for each type of fault feature, and introducing a dynamic dropout mechanism into each sub-network. The activation function of the LSTM sub-network is a tanh function with an added temperature coefficient. A feature attention layer is set between the output of the LSTM sub-network and the fully connected layer. The feature attention layer calculates the correlation weight between each feature and the remaining lifespan of the loom through a bidirectional gating mechanism, and dynamically weights the multi-scale LSTM output features based on the correlation weight. The dynamic weighting calculation formula is as follows: ， in, Let be the attention weight for the i-th feature. Let i be the i-th feature value output by the multi-scale LSTM. and These represent different trainable parameters of the attention layer. As a fault-sensitive factor, Here, is the feedback adjustment coefficient, and N is the total number of features.

8. The method for predicting the lifespan of a water-jet loom based on deep learning according to claim 1, characterized in that, The method of constructing a lifespan prediction model using the extracted features as input also includes adding a hierarchical L2 regularization term to the weight parameters of the fully connected layer, and introducing dynamic adjustment of weight decay in the L2 regularization term. Different regularization coefficients are set for the weight parameters associated with fault-sensitive features and other weight parameters respectively. At the same time, a loss function that integrates the weights of the fault stage is constructed, and differentiated loss weights are set for different samples according to the stage of the remaining lifespan of the loom. The expression of the loss function is: ， in, Let be the fault stage weight for the i-th sample. For the i-th sample, the actual remaining lifetime is... To predict remaining lifespan, The set of weights associated with fault-sensitive features. For other weight sets, and These are regularization systems for different weight sets.

9. The method for predicting the lifespan of a water-jet loom based on deep learning according to claim 1, characterized in that, The optimization of the life prediction model using a cross-model transfer learning method includes: selecting multiple benchmark models with known full life cycle data based on the life prediction model architecture; training a benchmark model containing a general fault feature extraction module and a model-specific adaptation module using the multi-dimensional feature set of the benchmark models; freezing the parameters of the general fault feature extraction module; inputting limited data of the target model into the benchmark model and then fine-tuning only the parameters of the model-specific adaptation module; introducing a correction coefficient based on the ratio of rated power and operating speed between the target model and the benchmark model during fine-tuning; and using a small learning rate gradient descent method to iteratively update the parameters. The iterative update expression is as follows: , in, To optimize the parameters of the dedicated adaptation module for the target model, The initial parameters for the adaptation module specific to the baseline model. To fine-tune the learning rate, The gradient of the loss function with respect to the baseline parameters. The training dataset for the target device model. This is the proportionality coefficient. and These are the rated power of the target model and the reference model, respectively. and These are the operating speeds of the target model and the reference model, respectively.

10. A deep learning-based life prediction system for water-jet looms, executing the method of claim 1, characterized in that, include: The data acquisition module is configured to acquire the operating data of the water jet loom through a multimodal sensor collaborative mechanism, and dynamically adjust the sensor sampling frequency based on the operating status of the water jet loom. The preprocessing module is configured to preprocess the acquired running data, including denoising the running data using an adaptive wavelet threshold and introducing an attention mechanism to fill in missing values ​​to obtain preprocessed data. The conversion module is configured to: extract features based on preprocessed data to obtain a multi-dimensional feature set, including extracting time-domain, frequency-domain and time-series features, and amplify the features of minor faults using a fault-sensitive feature enhancement method; The model module is configured to: construct a lifetime prediction model using the extracted features as input, including using a multi-scale LSTM network as the model framework, embedding a feature attention layer based on the model framework and configuring an L2 regularization strategy. The feature extraction module is configured to optimize the lifespan prediction model using a cross-model transfer learning method. The transformation module is configured to output the remaining lifetime prediction value and uncertainty index based on the optimized lifetime prediction model, and to perform graded early warning of remaining lifetime based on the prediction results.

Images