A fault prediction method based on adaptive weighted multi-modal joint entropy

Abstract

The present application belongs to the technical field of intelligent operation and maintenance and fault prediction of industrial equipment, and relates to a fault prediction method based on adaptive weighted multi-modal joint entropy, aiming at accurately identifying the state of the equipment and predicting the fault. The method comprises preprocessing the historical multi-modal sensing data of the industrial equipment, and constructing a condition adaptive weight mapping table for representing the corresponding relationship between different working condition parameters and the optimal basic weight of each mode. For the real-time collected multi-modal sensing data, the condition adaptive weight mapping table is queried according to the real-time working condition parameters, and the dynamic fusion weight of each mode is calculated in combination with the real-time signal quality evaluation. The real-time data of each mode is converted into a symbolic mode. A fault-tolerant fusion mechanism with similarity weighting is used to fuse the symbolic mode of each mode to generate a joint mode probability distribution, and the adaptive weighted multi-modal joint entropy value is calculated and input into a pre-trained prediction model to output the state identification or remaining useful life prediction result of the industrial equipment.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and fault prediction technology for industrial equipment, and in particular to a fault prediction method based on adaptive weighted multimodal joint entropy. Background Technology

[0002] Industrial equipment is the core foundation of modern industrial production systems, and its continuous and stable operation directly affects production safety, efficiency, and economic benefits. With the rapid development of Industrial Internet of Things (IIoT) technology, deploying multiple sensors on critical equipment for condition monitoring has become commonplace, generating massive amounts of multimodal data that provide a data foundation for predictive maintenance. However, extracting stable and reliable condition characteristics from this heterogeneous, high-dimensional data and achieving accurate fault prediction and health management remains a major challenge in the field of intelligent industrial operation and maintenance.

[0003] Currently, state monitoring methods based on information entropy have attracted widespread attention due to their ability to effectively quantify signal complexity. Existing multivariate entropy methods, such as multivariate sample entropy and multivariate permutation entropy, while capable of handling multi-channel signals, generally have significant limitations. These methods typically assume that the contribution of each modal data to state recognition is the same or fixed, employing simple feature concatenation or static weighting for fusion, failing to fully consider the dynamic changes in the quality, discriminative power, and equipment operating conditions of different modal signals in real-world industrial scenarios. When faced with complex situations such as fluctuating operating conditions, sensor performance degradation, or missing modal data, the feature representation capabilities of such methods significantly decrease, leading to insufficient accuracy and robustness in fault warnings. Summary of the Invention

[0004] To address the problems mentioned in the background art, this invention provides a fault prediction method based on adaptive weighted multimodal joint entropy, aiming to achieve efficient and robust fusion of multimodal data, thereby accurately identifying equipment status and predicting faults.

[0005] This invention adopts the following technical solution: This invention provides a fault prediction method based on adaptive weighted multimodal joint entropy, including: S1, data preprocessing and mapping relationship construction: Preprocessing historical multimodal sensing data of industrial equipment, and constructing a conditional adaptive weight mapping table based on the preprocessed historical data to characterize the correspondence between different operating parameters and the optimal basic weights of each mode. Multimodal sensing data refers to multi-channel signal data reflecting the equipment status collected by different types of sensors on industrial equipment. S2, dynamic weight calculation and symbolization: For real-time acquired multimodal sensing data, querying the conditional adaptive weight mapping table according to real-time operating parameters, and calculating the dynamic fusion weight of each mode in conjunction with real-time signal quality assessment. Simultaneously, converting the real-time data of each mode into symbolic patterns. S3, fault-tolerant fusion and joint entropy calculation: Based on the dynamic fusion weights, a similarity-weighted fault-tolerant fusion mechanism is used to fuse the symbolic patterns of each mode, generating a joint pattern probability distribution, and calculating the adaptive weighted multimodal joint entropy value accordingly. The adaptive weighted multimodal joint entropy value is an entropy characteristic index used to quantify the complexity of equipment status. S4. Equipment Status Prediction: The adaptive weighted multimodal joint entropy value is input into the pre-trained prediction model, and the output is the status identification or remaining service life prediction result of industrial equipment.

[0006] Step S1 involves constructing a conditionally adaptive weight mapping table, including: quantifying the discriminant index of each modality data. The discriminant index is a comprehensive index that integrates the statistical feature differences of the modality data under healthy and faulty states and its correlation with the equipment state label. For different operating parameter ranges, with the goal of maximizing the mode distribution differences under different equipment states, the optimal basic weights of each mode under different operating conditions are solved, and then a conditionally adaptive weight mapping table is constructed.

[0007] Step S2 calculates the dynamic fusion weights for each mode, including: retrieving the basic weights for each mode from the conditional adaptive weight mapping table based on the real-time acquired operating parameters; estimating the real-time instantaneous signal-to-noise ratio (SNR) of each mode as the real-time signal quality assessment result; and combining the basic weights, real-time instantaneous SNR, and offline-calculated discriminant indicators, performing weighted adjustments and normalization to obtain the final dynamic fusion weights.

[0008] The fault-tolerant fusion mechanism in step S3 includes: setting a target reference mode, which is determined based on the historical symbol patterns under the device's health status; calculating the similarity between the real-time symbol patterns of each modality and the target reference mode; multiplying the similarity of each modality by its corresponding dynamic fusion weight and summing the results to obtain the contribution of the joint mode; and reconstructing the joint mode probability distribution based on the contribution sequence.

[0009] Similarity is calculated using the cosine similarity function.

[0010] Step S1, which involves preprocessing historical multimodal data, includes at least data cleaning, filtering and denoising, standardization, and time-series alignment. Time-series alignment uses the device's rotational speed signal as a unified time reference.

[0011] In step S2, converting the real-time data of each modality into symbolic patterns means using a symbolization method based on data quantiles to map continuous data sequences into discrete symbolic sequences.

[0012] The prediction model in step S4 is a gradient boosting tree model.

[0013] Compared to existing technologies, the beneficial effects are as follows:

[0014] 1. This invention constructs an adaptive weight mapping table for operating conditions, dynamically associating operating condition parameters with modal weights. It also combines real-time signal quality assessment and offline discriminative indices for triple weight calibration, enabling the fusion strategy to adaptively adjust to the operating environment. This mechanism effectively overcomes the mismatch problem of fixed weights in varying operating conditions, making feature extraction more targeted and significantly improving the accuracy of condition monitoring and prediction under complex operating conditions.

[0015] 2. To address the unavoidable noise interference and sensor anomalies in industrial environments, this invention introduces a similarity-weighted fault-tolerant fusion mechanism. This mechanism calculates the similarity between symbolic patterns and health benchmarks, and then performs weighted fusion with dynamic weights, replacing traditional hard matching or direct splicing. This method enables the system to inherently tolerate data anomalies in individual modes, transient noise, or even temporary failures of some modes, ensuring that the system maintains reliable performance output even when data quality deteriorates, greatly enhancing its engineering practicality.

[0016] 3. This invention does not simply fuse multimodal data, but rather generates a new, more representative core feature—an adaptive weighted multimodal joint entropy value—through dynamic weight allocation and fault-tolerant symbol fusion. This entropy value feature deeply integrates the dynamic importance of each modality and the joint probability distribution of its symbol patterns, enabling it to more sensitively capture early degradation trends and minor fault characteristics of equipment status. This provides the backend prediction model with more discriminative and stable input, fundamentally improving the accuracy of fault identification and remaining life prediction. Attached Figure Description

[0017] Figure 1 This is a flowchart of a fault prediction method based on adaptive weighted multimodal joint entropy provided in an embodiment of this application;

[0018] Figure 2 This is a schematic diagram of experimental comparison results provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] This application provides a fault prediction method based on adaptive weighted multimodal joint entropy, for example, such as... Figure 1 As shown. The method includes:

[0021] S1. Data Preprocessing and Mapping Relationship Construction: Historical multimodal sensor data from industrial equipment is preprocessed, and based on the preprocessed historical data, a conditional adaptive weight mapping table is constructed to characterize the correspondence between different operating parameters and the optimal basic weights of each mode. Multimodal sensor data refers to multi-channel signal data reflecting the equipment status collected by different types of sensors on industrial equipment.

[0022] As one possible implementation, the preprocessing of historical multimodal data in step S1 includes at least data cleaning, filtering and denoising, standardization, and time-series alignment. The time-series alignment uses the device's rotational speed signal as a unified time reference.

[0023] Data cleaning and filtering directly remove impulse interference and background noise, preventing outliers from generating incorrect symbol patterns during symbolization and avoiding noise interference with real-time signal-to-noise ratio estimation, thus ensuring the reliability of dynamic fusion weight calculation. Standardization eliminates the impact of differences in dimensions and magnitudes among modal data, ensuring all modalities are on a fair comparison benchmark when calculating discriminant indicators, similarity, and fusion. Timing alignment using the rotational speed signal as a unified time benchmark ensures that data points from different modalities have a true physical causal relationship on the time axis. This step eliminates feature confusion and fusion distortion caused by data asynchrony, laying a solid foundation for subsequent calculations of adaptive weighted multimodal joint entropy that accurately reflects the physical state of the equipment, thereby directly improving the accuracy of state recognition and prediction.

[0024] Step S1 involves constructing a conditionally adaptive weight mapping table, including: quantifying the discriminant index of each modality data. The discriminant index is a comprehensive index that integrates the statistical feature differences of the modality data under healthy and faulty states and its correlation with the equipment state label. For different operating parameter ranges, with the goal of maximizing the mode distribution differences under different equipment states, the optimal basic weights of each mode under different operating conditions are solved, and then a conditionally adaptive weight mapping table is constructed.

[0025] For example, discriminant indicators The calculation formula is:

[0026]

[0027] in, and These represent the mean values ​​of the modal data under healthy and faulty states, respectively. and Standard deviation, This is the modal data. with status labels Mutual information between them, where λ is the balance coefficient.

[0028] For example, before solving for the optimal basic weights of each mode under this operating condition by maximizing the KL divergence, the method also includes calculating the average signal-to-noise ratio (SNR) of each mode. m This is a sub-step, namely, performing signal quality assessment.

[0029] For example, the optimal basic weights of each mode under this operating condition are solved by maximizing the KL divergence. The calculation formula is as follows:

[0030]

[0031] Where RPM is the speed operating condition parameter, p(c|RPM) represents the prior probability of state c at a specific speed RPM, and KL(p(Π|c)∥p(Π)) represents the KL divergence between the joint pattern distribution and the overall distribution in state c.

[0032] S2. Calculate dynamic weights and symbolization: For real-time acquired multimodal sensor data, the adaptive weight mapping table is queried based on real-time operating parameters, and combined with real-time signal quality assessment, the dynamic fusion weights of each mode are calculated. At the same time, the real-time data of each mode is converted into symbolic mode.

[0033] As one possible implementation, step S2 calculates the dynamic fusion weights for each mode, including: retrieving the basic weights for each mode from a conditional adaptive weight mapping table based on real-time acquired operating parameters, such as rotational speed RPM(t); and estimating the real-time instantaneous signal-to-noise ratio (SNR) for each mode. m (t) is used as the real-time signal quality assessment result. This is combined with the base weights and the real-time instantaneous signal-to-noise ratio (SNR). m (t) and the discriminant index calculated offline. The weighted adjustments and normalization processes are then performed to obtain the final dynamic fusion weights. .

[0034] In step S2, the real-time data of each mode is converted into symbolic mode. This refers to using a symbolization method based on data quantiles to map a continuous data sequence into a discrete symbol sequence.

[0035] S3. Fault-Tolerant Fusion and Joint Entropy Calculation: Based on dynamic fusion weights, a similarity-weighted fault-tolerant fusion mechanism is used to fuse the symbol patterns of each modality, generating a joint pattern probability distribution, and calculating the adaptive weighted multimodal joint entropy value AWMJE accordingly. The adaptive weighted multimodal joint entropy value is an entropy characteristic index used to quantify the state complexity of a device.

[0036] As one possible implementation, the fault-tolerant fusion mechanism in step S3 includes: setting a target reference mode, which is determined based on the historical symbol patterns under the device's health state; calculating the similarity between the real-time symbol patterns of each modality and the target reference mode; multiplying the similarity of each modality by its corresponding dynamic fusion weight and summing the results to obtain the contribution of the joint mode; and reconstructing the joint mode probability distribution based on the contribution sequence.

[0037] For example, the contribution calculation formula for the joint mode is as follows:

[0038]

[0039] Where sim(·) is the similarity function, and π_m(t) is the sign pattern of mode m at time t. It is the target reference mode.

[0040] For example, similarity can be calculated using a cosine similarity function.

[0041] S4. Equipment Status Prediction: The adaptive weighted multimodal joint entropy value is input into the pre-trained prediction model, and the output is the status identification or remaining service life prediction result of industrial equipment.

[0042] As one possible implementation, the prediction model in step S4 is a gradient boosting tree model.

[0043] This application also provides an embodiment, using NASA's C-MAPSS aero-engine degradation dataset as an example, to detail the specific implementation of the method provided in this application for industrial equipment condition monitoring and fault prediction. The method includes the following steps:

[0044] 1. Data Acquisition: Acquire multimodal sensor data during the operation of the aero-engine. The multimodal sensor data includes, but is not limited to, vibration signals, temperature signals, pressure signals, and speed signals, which are key operating parameters.

[0045] For example, the vibration signal uses an ICP accelerometer with a range of ±500g and a sampling frequency of 51.2kHz.

[0046] Temperature signal: K-type thermocouple is used, with a measurement range of 0-1200℃ and a sampling frequency of 1Hz;

[0047] Pressure signal: A piezoresistive pressure sensor is used, with a range of 0-5MPa and a sampling frequency of 10kHz;

[0048] Speed ​​signal: photoelectric encoder with a resolution of 1024 pulses / revolution and a sampling frequency of 100Hz;

[0049] 2. Data preprocessing:

[0050] Data cleaning: Outliers in the data are identified and removed using statistical distribution-based criteria (such as the 3σ criterion). A sliding window of 256 sampling points is set, and the mean and standard deviation of the data within each window are calculated. Data points exceeding the μ±3σ range are marked as outliers and removed.

[0051] Filtering and denoising: Daubechies 5 wavelet decomposition is performed at 5 levels. The wavelet coefficients are denoised using the soft thresholding method. The signal is reconstructed, retaining the effective frequency components of 0-10kHz. Moving average filtering is used for the temperature signal with a window length of 10.

[0052] Standardization: Z-Score standardization method is used.

[0053]

[0054] Where μ is the mean of the training set data and σ is the standard deviation. Data for each modality is processed separately to eliminate the influence of dimensions.

[0055] Time alignment: Using the device's rotational speed signal as a unified time reference, a time series alignment algorithm (such as Dynamic Time Warping (DTW)) is employed to synchronize modal data with different sampling rates. The maximum bending window is set to 10 sampling points, and Euclidean distance is used as the similarity metric.

[0056] Resampling: All modal data are resampled to a uniform sampling frequency to ensure consistency in data length and timestamps. For example, all modal data are uniformly resampled to 256Hz using a linear interpolation method for sampling rate conversion.

[0057] 3. Based on the preprocessed historical data, construct a conditional adaptive weight mapping table:

[0058] Samples of equipment in healthy and faulty states are extracted from historical data, and discriminant indices for each modality are calculated. This index is a comprehensive indicator that integrates the statistical differences in the data of that modality under healthy and faulty states, as well as the correlation between the data of that modality and the equipment status label. 1000 healthy state samples and 1000 faulty state samples are selected from the training set, and the Fisher Score components for each modality are calculated.

[0059]

[0060] Calculate the mutual information I(S_m; Y) between each mode and the state label, and set the balance coefficient λ=0.5.

[0061] Calculate the comprehensive discriminant index:

[0062]

[0063] The sliding window method was used to calculate the average signal-to-noise ratio (SNR) of each mode across the entire historical dataset, serving as the offline evaluation result of its signal quality. The window length was set to 256 sampling points, and the signal power was calculated as follows:

[0064]

[0065] Calculate noise power:

[0066]

[0067] The SNR curves are smoothed using a moving average filter with a window length of 10, and the average SNR of each mode is calculated over the entire training set. m value.

[0068] Key operating parameters (such as rotational speed) are divided into several continuous intervals. For each operating parameter interval, with the goal of maximizing the difference between the joint mode distribution and the overall distribution under different equipment health states, an optimization algorithm is used to solve for the optimal basic weights of each mode under that operating condition. Finally, a mapping relationship between the operating parameter interval index and the optimal basic weights of each mode is established, namely, a conditional adaptive weight mapping table.

[0069] The rotational speed range was divided into 10 equally spaced intervals, and 100 sample data points were collected for each interval. The optimal weights were then determined by maximizing the KL divergence.

[0070]

[0071] The gradient descent method was used for optimization, with a learning rate of 0.01 and 1000 iterations. A mapping dictionary from the RPM interval to the optimal weights was established, and the weights were normalized to ensure that the sum of the weights in each RPM interval was 1.

[0072] 4. Online Dynamic Weight Calculation and Symbolization: Real-time signal processing is performed first. A sliding window is used to process the real-time acquired multimodal sensor data stream. The sliding window length is set to 256 sampling points, and the step size is set to 1 sampling point to achieve real-time processing. The signal-to-noise ratio is calculated in real-time for the data within each window, and a recursive calculation method is used to improve computational efficiency.

[0073] Then, the dynamic fusion weights are calculated, and the current operating parameters of the equipment are obtained in real time, such as the current rotational speed value RPM(t). Based on the current operating parameters, the conditional adaptive weight mapping table is queried to obtain the basic weights corresponding to each mode, such as obtaining the basic weights of the current rotational speed range. .

[0074] Calculate the weight adjustment factor:

[0075] The instantaneous signal-to-noise ratio (SNR) of each mode is estimated in real time and used as the real-time signal quality assessment result. Combined with the basic weights obtained from the query, the real-time estimated instantaneous SNR, and the discriminant index calculated offline, the final dynamic fusion weights are calculated through weighted adjustment and normalization.

[0076]

[0077] Ensure the weights satisfy ∑w m (t)=1.

[0078] Symbolization converts the real-time data sequence of each modality into a symbolic pattern. Continuous data values ​​are mapped to a predefined discrete symbol set, forming a fixed-length symbol sequence, i.e., the symbolic pattern for each modality. A quartile-based symbolization method is used to calculate the Q1, Q2, and Q3 quartiles for each modality's data, converting the data into four types of symbols:

[0079] 0:x <Q1

[0080] 1: Q1 ≤ x <Q2

[0081] 2: Q2 ≤ x <Q3

[0082] 3: x ≥ Q3

[0083] The symbol pattern length is set to 6, and the symbol sequence π is generated using a sliding window method. m (t).

[0084] 5. A similarity-weighted soft fusion mechanism is adopted, and the specific implementation steps are as follows:

[0085] (1) Target mode setting: The target mode is a typical mode determined based on the symbolic pattern of historical data of the device in a healthy state.

[0086] Typical patterns of each modality under healthy conditions are selected as target reference patterns. For each modality, 100 typical patterns are extracted from the healthy state data, and the mean of these patterns is calculated as the final target reference pattern π. target .

[0087] (2) Similarity calculation: Calculate the similarity between the real-time symbolic pattern of each modality and the target reference pattern.

[0088] Cosine similarity is used to measure symbolic pattern similarity:

[0089]

[0090] Where L is the symbol pattern length, with a value of 6, and the similarity is normalized to ensure that the value range is [0,1].

[0091] (3) Joint mode fusion: The similarity of each mode is multiplied by its corresponding dynamic fusion weight and then summed to obtain the joint mode at the current time.

[0092] Calculate the joint model contribution at each time point:

[0093]

[0094] The contribution series was smoothed using an exponentially weighted moving average:

[0095]

[0096] 6. Calculate the adaptive weighted multimodal joint entropy value based on the fusion results. The specific implementation steps are as follows:

[0097] (1) Probability distribution construction: Based on the contribution sequence over a period of time, the joint pattern probability distribution is reconstructed.

[0098] The joint model contribution values ​​were collected at 1000 consecutive time points. The range of contribution values ​​was divided into 10 equal intervals, and the frequency of occurrence of contribution values ​​within each interval was counted.

[0099] Calculate the probability distribution:

[0100]

[0101] in Let N be the frequency of the i-th interval, and N=1000 be the total number of samples.

[0102] (2) Entropy calculation

[0103] Calculate Shannon entropy:

[0104]

[0105] Normalize the entropy value:

[0106]

[0107] Set the update frequency to update the entropy value once every 10 time points.

[0108] (3) Trend Analysis

[0109] The first-order difference of the entropy value is used to calculate the trend of change.

[0110]

[0111] The trend curve is smoothed using the moving average method, and an anomaly detection threshold is set: an anomaly warning is triggered when |ΔAWMJE|>0.1.

[0112] 7. Status Identification and Fault Prediction

[0113] The specific implementation steps for equipment status identification and fault prediction based on AWMJE values ​​are as follows:

[0114] (1) Feature Engineering

[0115] Time-domain characteristics: mean, variance, skewness, and kurtosis of AWMJE values.

[0116] Frequency domain features: Perform FFT transform on the AWMJE sequence to extract the main frequency components.

[0117] Time series characteristics: Statistical properties of first-order and second-order differences

[0118] The feature vector is constructed with a dimension of 20.

[0119] (2) Model training

[0120] Gradient Boosting Tree (GBDT) model

[0121] The learning rate was set to 0.05; the maximum tree depth was set to 6; the number of trees was set to 100; the minimum number of leaf node samples was set to 10; the model was trained using 5-fold cross-validation and the hyperparameters were optimized using grid search.

[0122] (3) Online prediction

[0123] A status prediction is performed every 60 seconds. The input is the feature vectors of the last 100 time points, and the output is a device health status score, ranging from [0,1]. An early warning threshold is set: when the score is below 0.3, an early warning is issued, predicting the remaining useful life (RUL). A particle filter algorithm is used to update the RUL prediction results.

[0124] (4) Results visualization

[0125] It displays the AWMJE value change curve in real time, shows the dynamic weight distribution of each modality, displays the health status score and RUL prediction results, and provides historical data query and trend analysis functions.

[0126] 8: System Performance Evaluation and Verification

[0127] A comprehensive evaluation of system performance is conducted, and the specific implementation steps are as follows:

[0128] (1) Evaluation indicators

[0129] Accuracy: The consistency between the predicted result and the true label.

[0130] Noise gain coefficient: The ability to maintain performance under different noise levels.

[0131] Sensitivity index: robustness to parameter changes.

[0132] Computational efficiency: The time required for a single processing run.

[0133] Memory usage: Memory usage during runtime.

[0134] (2) Test Plan

[0135] Validation was performed using the C-MAPSS dataset, with 70% of the data used for training and 30% for testing. Robustness was tested by adding Gaussian white noise of different intensities (η=0.3, 0.7), and performance stability was tested under different parameter settings.

[0136] For example, to verify the overall performance of the AWMJE method proposed in this invention, it was compared with several mainstream multivariate entropy algorithms on a simulated dataset, including Multivariate Dispersion Entropy (MDDPE), Multivariate Sample Entropy (MVSE), Slope Fuzzy Entropy (SFE), and Multivariate Permutation Entropy (MVPE). The performance comparison results are as follows: Figure 2 As shown.

[0137] Reference Figure 2 It can be seen that the method (AWMJE) provided in this application achieves an accuracy of 92.58%, which is superior to all comparative algorithms, indicating that it has the strongest ability to identify equipment status. Under different intensities of noise interference, the noise gain coefficient of this method is the highest (G=0.98 when η=0.3, G=0.82 when η=0.7), significantly higher than the comparative methods, proving that its dynamic weight and fault-tolerant fusion mechanism can effectively suppress the influence of noise. The sensitivity index S is only 0.12, far lower than other comparative algorithms, indicating that the method of this invention has the lowest sensitivity to parameter changes and has the best parameter robustness and operational stability.

[0138] The above description, in conjunction with preferred technical solutions, provides a further detailed explanation of the present invention and should not be construed as limiting the specific implementation of the invention to these descriptions. For those skilled in the art, simple deductions and substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A fault prediction method based on adaptive weighted multimodal joint entropy, characterized in that, The process includes: S1, Data Preprocessing and Mapping Relationship Construction: Historical multimodal sensing data from industrial equipment is preprocessed, and based on the preprocessed historical data, a conditional adaptive weight mapping table is constructed to characterize the correspondence between different operating parameters and the optimal basic weights of each mode. The multimodal sensing data refers to multi-channel signal data reflecting the equipment status collected by different types of sensors on the industrial equipment. S2, Dynamic Weight Calculation and Symbolization: For real-time acquired multimodal sensing data, the conditional adaptive weight mapping table is queried according to real-time operating parameters, and combined with real-time signal quality assessment, the dynamic fusion weights of each mode are calculated. Simultaneously, the real-time data of each mode is converted into symbolic patterns. S3, Fault-Tolerant Fusion and Joint Entropy Calculation: Based on the dynamic fusion weights, a similarity-weighted fault-tolerant fusion mechanism is used to fuse the symbolic patterns of each mode, generating a joint pattern probability distribution, and calculating the adaptive weighted multimodal joint entropy value accordingly. The adaptive weighted multimodal joint entropy value is an entropy characteristic index used to quantify the complexity of the equipment status. S4. Equipment Status Prediction: Input the adaptive weighted multimodal joint entropy value into the pre-trained prediction model and output the status identification or remaining service life prediction result of the industrial equipment. Step S1 involves constructing a conditionally adaptive weight mapping table, including: quantifying the discriminative index of each modality data, wherein the discriminative index is a comprehensive index that integrates the statistical feature differences of the modality data under healthy and faulty states and its correlation with the equipment state label; and for different operating condition parameter ranges, with the goal of maximizing the mode distribution differences under different equipment states, solving for the optimal basic weight of each modality under different operating conditions, and then constructing the conditionally adaptive weight mapping table. Step S2 calculates the dynamic fusion weights for each mode, including: querying the conditional adaptive weight mapping table to obtain the basic weights for each mode based on the real-time acquired operating parameters; estimating the real-time instantaneous signal-to-noise ratio of each mode as the real-time signal quality assessment result; and combining the basic weights, the real-time instantaneous signal-to-noise ratio, and the discriminant index calculated offline to perform weighted adjustment and normalization processing to obtain the final dynamic fusion weights. The fault-tolerant fusion mechanism in step S3 includes: setting a target reference mode, which is determined based on the historical symbol mode under the device health state; calculating the similarity between the real-time symbol mode of each modality and the target reference mode; multiplying the similarity of each modality by its corresponding dynamic fusion weight and summing the results to obtain the contribution of the joint mode, and reconstructing the probability distribution of the joint mode based on the contribution sequence.

2. The fault prediction method based on adaptive weighted multimodal joint entropy according to claim 1, characterized in that, Similarity is calculated using the cosine similarity function.

3. The fault prediction method based on adaptive weighted multimodal joint entropy according to claim 1, characterized in that, Step S1 involves preprocessing historical multimodal data, including at least data cleaning, filtering and denoising, standardization, and time alignment. Time alignment uses the device's rotation speed signal as a unified time reference.

4. The fault prediction method based on adaptive weighted multimodal joint entropy according to claim 1, characterized in that, In step S2, converting the real-time data of each modality into symbolic patterns means using a symbolization method based on data quantiles to map continuous data sequences into discrete symbolic sequences.

5. The fault prediction method based on adaptive weighted multimodal joint entropy according to claim 1, characterized in that, The prediction model in step S4 is a gradient boosting tree model.

Images