Air compressor fault prediction method and system based on multi-source information fusion and machine learning

Abstract

This invention discloses a method and system for predicting air compressor faults based on multi-source information fusion and machine learning, comprising the following steps: collecting and preprocessing multi-source heterogeneous operating data during the operation of the air compressor; performing windowing processing on the standardized multi-source operating dataset and generating operating condition labels for each time window segment; extracting multi-source features from the set of multi-source time window segments and constructing structured feature vectors; performing multi-source information fusion modeling on the structured feature vectors; performing hierarchical hybrid modeling and prediction decoding processing on the fused state representation vector sequence based on the improved TSMixer model; performing time consistency processing and event aggregation processing on the fault evolution prediction result set; determining the maintenance strategy, generating predictive maintenance output results, and writing them into the storage system. This invention employs multi-source fusion and hierarchical time-series hybrid modeling to achieve air compressor fault evolution prediction, possessing the advantages of high prediction accuracy and reliable maintenance decisions.

Description

Technical Field

[0001] This invention relates to the field of air compressor fault prediction, and in particular to an air compressor fault prediction method and system based on multi-source information fusion and machine learning. Background Technology

[0002] Existing technologies for air compressor fault prediction and operational status assessment are mostly based on single sensor signals such as vibration, current, or temperature, or on simple parallel analysis of multiple signals. They often employ threshold judgment, statistical feature analysis, and traditional machine learning models to identify equipment operating status and estimate lifespan. These methods typically process collected data at a fixed time scale, rely on human experience to construct features and judgment rules, and have certain application effects in scenarios with relatively stable operating conditions or obvious fault modes.

[0003] In actual operation, air compressor operating conditions change frequently, and multi-source heterogeneous operating data exhibit significant differences in time scale, statistical characteristics, and response features. Existing technologies struggle to effectively fuse multi-source information under a unified time axis, making prediction results susceptible to abnormal fluctuations from a single information source. Furthermore, traditional time series modeling methods focus on overall sequence modeling or short-term feature analysis, failing to simultaneously characterize the evolution of local faults within a time window and long-term trends across time windows. This results in insufficient stability of prediction results, and the temporal consistency and reliability of event identification and maintenance decisions need improvement.

[0004] Therefore, how to provide a method and system for predicting air compressor failures based on multi-source information fusion and machine learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method and system for predicting air compressor failures based on multi-source information fusion and machine learning. This invention employs multi-source fusion and hierarchical time-series hybrid modeling to achieve air compressor failure evolution prediction, which has the advantages of high prediction accuracy and reliable maintenance decisions.

[0006] The air compressor fault prediction method based on multi-source information fusion and machine learning according to embodiments of the present invention includes the following steps:

[0007] Collect multi-source heterogeneous operating data during the operation of the air compressor, and preprocess it to obtain a standardized multi-source operating dataset;

[0008] Under a unified timeline, the standardized multi-source operation dataset is windowed to obtain a set of multi-source time window segments, and operating condition labels are generated for each time window segment.

[0009] For each time window segment in the set of multi-source time window segments with working condition labels, multi-source features are extracted and a structured feature vector is constructed;

[0010] Based on the consistency constraint between working condition labels and information sources, multi-source information fusion modeling is performed on the structured feature vectors to obtain a sequence of fused state representation vectors ordered by a unified time axis;

[0011] Based on the improved TSMixer model, hierarchical hybrid modeling and prediction decoding are performed on the fused state representation vector sequence to output a set of fault evolution prediction results.

[0012] Perform time consistency processing and event aggregation processing on the fault evolution prediction result set, and output the predicted event set;

[0013] The maintenance strategy is determined based on the set of predicted events, and the predictive maintenance output is generated and written to the storage system.

[0014] Optionally, the multi-source heterogeneous operating data includes vibration signal time series data, current signal time series data, and temperature time series data. The preprocessing includes performing missing value processing, outlier processing, noise reduction processing, and dimension unification processing on the multi-source heterogeneous operating data, and performing time alignment and resampling processing on the processed multi-source operating data based on the sampling timestamp.

[0015] Optionally, the generation of the working condition label specifically includes:

[0016] Under a unified timeline, windowing parameters are determined based on a standardized multi-source running dataset. Based on the windowing parameters, the standardized multi-source running dataset is truncated to obtain a set of multi-source time window segments. Time window segment indices are assigned to each time window segment in chronological order.

[0017] Generate corresponding time window segment identifiers for each time window segment in the multi-source time window segment set and establish time association relationships;

[0018] Within the time range covered by each time window segment, a set of air compressor operating parameter sequences aligned with the time window segment is obtained. The set of air compressor operating parameter sequences includes a loading state sequence, an unloading state sequence, an exhaust pressure sequence, an exhaust temperature sequence, and a motor current sequence.

[0019] Statistical calculations were performed on the exhaust pressure sequence, exhaust temperature sequence, and motor current sequence to obtain the mean exhaust pressure, mean exhaust temperature, and mean motor current.

[0020] By combining the loading state sequence and unloading state sequence within the time window segment, the loading state value and unloading state value corresponding to the time window segment are determined, and the values ​​are integrated with the average exhaust pressure, average exhaust temperature and average motor current to obtain the working condition judgment parameter group.

[0021] Based on the working condition judgment parameter group, the working condition label corresponding to the time window segment is generated, and the working condition label is bound to the time window segment identifier and written into the multi-source time window segment set.

[0022] Optionally, the construction of the structured feature vector specifically includes:

[0023] Read the vibration signal time series data, current signal time series data and temperature signal time series data corresponding to each time window segment in the multi-source time window segment set with working condition labels, and obtain the corresponding working condition labels;

[0024] Time-domain statistical calculations and frequency-domain spectrum calculations are performed on the time-series data of vibration signals to obtain the time-domain statistical feature set and the frequency-domain spectrum feature set of vibration signals.

[0025] Perform time-frequency joint feature extraction processing on the vibration signal time-series data to obtain the vibration signal time-frequency joint feature set;

[0026] Statistical calculations and fluctuation characteristic calculations are performed on the time-series data of the current signal to obtain a set of statistical characteristics and a set of fluctuation characteristics of the current signal.

[0027] Perform trend feature extraction and fluctuation feature calculation on the time series data of temperature signals to obtain a set of temperature signal trend features and a set of temperature signal fluctuation features;

[0028] Feature indices are assigned to the time-domain statistical feature set, frequency-domain spectral feature set, time-frequency joint feature set, current signal statistical feature set, current signal fluctuation feature set, temperature signal trend feature set, and temperature signal fluctuation feature set, respectively. Information source identifiers are assigned to the vibration signal features, current signal features, and temperature signal features, respectively.

[0029] The feature sets are sequentially concatenated according to the information source identifier and feature index to generate structured feature vectors corresponding to each time window segment, and the structured feature vectors are bound and stored with the corresponding working condition labels.

[0030] Optionally, obtaining the fusion state representation vector sequence specifically includes:

[0031] Read the structured feature vectors corresponding to each time window segment in the multi-source time window segment set, and obtain the working condition label, information source identifier and feature index bound to each structured feature vector. Sort each structured feature vector according to the time window segment index to form a feature input sequence.

[0032] Based on the working condition labels, conditional reweighting is performed on the feature input sequence to obtain the working condition conditional feature sequence.

[0033] Based on the information source consistency constraint, the conditional feature sequence is subjected to consistency correction processing to obtain the consistency-corrected feature sequence.

[0034] Construct a feature space mapping relationship and perform feature space mapping processing on the consistency correction feature sequence based on the feature space mapping relationship to obtain the mapped feature sequence;

[0035] The fusion generation process is performed on the mapper vectors in the unified fusion feature space to obtain the fusion state representation vector. The fusion state representation vectors corresponding to each time window segment are combined in the order of the time window segment index to obtain the fusion state representation vector sequence.

[0036] Optionally, the output of the fault evolution prediction result set specifically includes:

[0037] The fusion state representation vector sequence is fragmented and organized to form a fragmented sequence for model input;

[0038] Based on the time window segment range defined by the segment organization sequence, the fusion state representation vector sequence is input into the improved TSMixer model. In the in-window time mixing module, in-window mixing modeling is performed on the in-segment time series corresponding to each time window segment to obtain the in-window representation vector sequence corresponding to each time window segment.

[0039] The improvements to the TSMixer model are as follows: the original TSMixer model's method of uniformly performing time mixing on the entire time series is changed to a hierarchical time mixing method based on time window segments and executed separately through the intra-window time mixing module and the inter-window time mixing module; and the unconstrained feature channel mixing method in the feature dimension mixing module of the original model is changed to a feature dimension mixing method constrained by the set of multi-source information relationship constraints.

[0040] In the inter-window temporal blending module, inter-window blending modeling is performed on the intra-window representation vector sequences corresponding to multiple time window segments in the order of time window segment index to obtain the inter-window representation vector sequence across time window segments;

[0041] In the feature dimension hybrid module, a set of multi-source information relationship constraints is determined based on the information source identifier and feature index, and a feature dimension hybrid connection constraint rule is constructed based on the set of multi-source information relationship constraints.

[0042] The in-window representation vector sequence and the inter-window representation vector sequence are combined to obtain a time series hybrid representation. When performing feature dimension hybridization operation on the time series hybrid representation, the hybrid connection relationship between feature channels is limited according to the feature dimension hybrid connection constraint rule to obtain the constrained improved time series hybrid representation sequence.

[0043] The improved time series hybrid representation sequence is input into the fault evolution prediction module for prediction decoding processing, and the fault evolution prediction result set is output. The fault evolution prediction result set includes a fault risk value sequence, a performance degradation trend sequence, and a remaining service life prediction value sequence.

[0044] Optionally, the output of the predicted event set specifically includes:

[0045] Read the set of fault evolution prediction results, and align and organize the fault risk value sequence, performance degradation trend sequence and remaining service life prediction value sequence according to the time window segment index order of the multi-source time window segment set to form a set of prediction result entries that correspond one-to-one with each time window segment;

[0046] The prediction result set is subjected to adjacent time window prediction consistency constraint processing. After the consistency correction update is completed, the risk change magnitude is calculated for each time window segment based on the fault risk value sequence, and the duration of each time window segment is calculated using the start and end timestamps of the time window segment.

[0047] The event triggering criteria are determined based on the magnitude of risk change and the duration of time window segments. Threshold judgment processing is then performed on each time window segment to obtain candidate time window segments for the event.

[0048] The event candidate time window segments are subjected to continuous detection and aggregation processing according to the time window segment index order. Event candidate time window segments with consecutive time window segment indexes are merged into the same predicted event, and a predicted event identifier is generated for each predicted event.

[0049] For each predicted event, a predicted event record is generated and a predicted event set is output. The predicted event set includes a predicted event identifier, a predicted event start time, a predicted event end time, an operating condition label, a fault risk level, and a predicted remaining service life value.

[0050] Optionally, the predictive maintenance output includes a maintenance recommendation identifier, a recommended maintenance time window, a fault type identifier, and a fault risk level identifier.

[0051] An air compressor fault prediction system based on multi-source information fusion and machine learning according to an embodiment of the present invention includes:

[0052] The data preprocessing module is used to collect multi-source heterogeneous operating data during the operation of the air compressor and preprocess it to obtain a standardized multi-source operating dataset.

[0053] The window labeling module is used to perform window processing on the standardized multi-source operation dataset under a unified time axis, obtain a set of multi-source time window segments, and generate working condition labels for each time window segment.

[0054] The feature construction module is used to extract multi-source features and construct structured feature vectors for each time window segment in a set of multi-source time window segments with working condition labels.

[0055] The fusion modeling module is used to perform multi-source information fusion modeling on structured feature vectors based on the consistency constraints between working condition labels and information sources, and obtain a fusion state representation vector sequence.

[0056] The fault prediction module is used to perform hierarchical hybrid modeling and prediction decoding on the fused state representation vector sequence based on the improved TSMixer model, and output a set of fault evolution prediction results.

[0057] The event generation module is used to perform time consistency processing and event aggregation processing on the fault evolution prediction result set and output the predicted event set.

[0058] The maintenance decision module is used to determine maintenance strategies based on a set of predicted events, generate predictive maintenance output results, and write them to the storage system.

[0059] The beneficial effects of this invention are:

[0060] This invention achieves collaborative modeling and consistency constraint processing of different information source features such as vibration, current, and temperature by performing windowed modeling of multi-source heterogeneous operating data of air compressors under a unified time axis and introducing a multi-source information fusion mechanism driven by operating condition labels. This enables the fused state representation vector to simultaneously reflect the multi-dimensional operating state features of the equipment under specific operating conditions. By performing conditional reweighting, consistency correction, and feature space mapping processing on multi-source features, the interference of abnormal fluctuations from a single information source on the overall state representation is effectively reduced, improving the stability and comparability of the multi-source data fusion results and providing a unified and reliable state input foundation for subsequent fault evolution modeling.

[0061] In the fault prediction stage, this invention employs an improved TSMixer model to perform hierarchical mixing modeling of the fused state representation vector sequence. By organizing the time series into time window segments and performing intra-window and inter-window time mixing processes separately, it can simultaneously characterize the short-term state evolution features within a time window segment and the long-term change trend across time windows, effectively avoiding the smooth masking problem of local fault features in traditional overall time mixing methods. Simultaneously, by introducing a feature dimension mixing method constrained by multi-source information relationships, the information interaction between feature channels conforms to the correlation structure between physical information sources, reducing noise propagation caused by the mixing of irrelevant features, thereby improving the accuracy and consistency of fault risk assessment, performance degradation trend prediction, and remaining service life prediction results.

[0062] In the prediction result output stage, this invention further performs time consistency processing and event aggregation processing on the fault evolution prediction result set. The prediction results in continuous time window segments are organized into prediction events with clear start and end times and risk levels. Based on these prediction events, structured predictive maintenance output results are generated. By using fault risk level, remaining service life prediction value, operating condition label, and fused state characteristics together for maintenance strategy determination, maintenance suggestion identifiers and maintenance time windows that match the actual operating state can be output, giving maintenance decisions clear time orientation and risk basis. Therefore, this invention realizes a complete closed-loop processing flow from multi-source data acquisition and fault evolution prediction to predictive maintenance decision-making, improving the reliability of air compressor fault prediction, the executability of maintenance decisions, and the overall engineering application value of the system. Attached Figure Description

[0063] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0064] Figure 1 The flowchart shows the air compressor fault prediction method based on multi-source information fusion and machine learning proposed in this invention.

[0065] Figure 2 This is a flowchart illustrating the generation of the fault evolution prediction result set of the air compressor fault prediction method based on multi-source information fusion and machine learning proposed in this invention.

[0066] Figure 3 This is a flowchart illustrating the process of generating the prediction event set for the air compressor fault prediction method based on multi-source information fusion and machine learning proposed in this invention. Detailed Implementation

[0067] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0068] refer to Figures 1-3 The air compressor fault prediction method based on multi-source information fusion and machine learning includes the following steps:

[0069] Collect multi-source heterogeneous operating data during the operation of the air compressor, and preprocess it to obtain a standardized multi-source operating dataset;

[0070] Under a unified timeline, the standardized multi-source operation dataset is windowed to obtain a set of multi-source time window segments, and operating condition labels are generated for each time window segment.

[0071] For each time window segment in the set of multi-source time window segments with working condition labels, multi-source features are extracted and a structured feature vector is constructed;

[0072] Based on the consistency constraint between working condition labels and information sources, multi-source information fusion modeling is performed on the structured feature vectors to obtain a sequence of fused state representation vectors ordered by a unified time axis;

[0073] Based on the improved TSMixer model, hierarchical hybrid modeling and prediction decoding are performed on the fused state representation vector sequence to output a set of fault evolution prediction results.

[0074] Perform time consistency processing and event aggregation processing on the fault evolution prediction result set, and output the predicted event set;

[0075] The maintenance strategy is determined based on the set of predicted events, and the predictive maintenance output is generated and written to the storage system.

[0076] In this embodiment, the multi-source heterogeneous operating data includes vibration signal time series data, current signal time series data, and temperature time series data. The preprocessing includes performing missing value processing, outlier processing, noise reduction processing, and dimension unification processing on the multi-source heterogeneous operating data, and performing time alignment and resampling processing on the processed multi-source operating data based on the sampling timestamp.

[0077] In this embodiment, the generation of the working condition label specifically includes:

[0078] Under a unified time axis, windowing parameters are determined based on a standardized multi-source running dataset. Based on the windowing parameters, the standardized multi-source running dataset is truncated to obtain a set of multi-source time window segments. Time window segment indices are assigned to each time window segment according to the time order. The windowing parameters are determined based on the sampling characteristics and temporal continuity of the standardized multi-source running dataset, including the time window length and time window step.

[0079] For each time window segment in the multi-source time window segment set, a corresponding time window segment identifier is generated and a time association relationship is established. The time window segment identifier is bound to the time window segment index and stored in association with the corresponding start timestamp and end timestamp, which is used to identify the time position of the time window segment in the unified time axis.

[0080] Within the time range covered by each time window segment, obtain a set of air compressor operating parameter sequences aligned with the time window segment. The set of air compressor operating parameter sequences includes loading state sequence, unloading state sequence, exhaust pressure sequence, exhaust temperature sequence, and motor current sequence.

[0081] Statistical calculations were performed on the exhaust pressure sequence, exhaust temperature sequence, and motor current sequence to obtain the mean values ​​of exhaust pressure, exhaust temperature, and motor current. For any operating parameter sequence, the parameter values ​​corresponding to the resampling times of each unified time axis within the time window segment were summarized and calculated. The number of resampling times within the time window segment was used as the normalization base to obtain the mean statistical results of the corresponding operating parameter within the time window segment.

[0082] By combining the loading state sequence and unloading state sequence within the time window segment, the loading state value and unloading state value corresponding to the time window segment are determined. The loading state value and unloading state value are then integrated with the average exhaust pressure, average exhaust temperature, and average motor current to obtain the working condition judgment parameter group.

[0083] Based on the working condition judgment parameter group, the working condition label corresponding to the time window segment is generated, and the working condition label is bound to the time window segment identifier and written into the multi-source time window segment set.

[0084] In this embodiment, the construction of the structured feature vector specifically includes:

[0085] Read the vibration signal time series data, current signal time series data and temperature signal time series data corresponding to each time window segment in the multi-source time window segment set with working condition labels, and obtain the corresponding working condition labels;

[0086] Time-domain statistical calculations and frequency-domain spectrum calculations are performed on the time-series data of vibration signals to obtain the time-domain statistical feature set and the frequency-domain spectrum feature set of vibration signals.

[0087] Time-domain statistical computation processing includes calculating the mean, maximum, minimum, root mean square, and peak-to-peak values ​​of the vibration signal sample value sequence within a time window segment;

[0088] The frequency domain spectrum calculation process includes converting the vibration signal sampling value sequence within a time window segment into a frequency component sequence, and calculating the frequency domain energy distribution based on the frequency component sequence. The frequency domain energy distribution is the result of squaring the amplitude of each frequency component and summarizing it according to the frequency interval. The frequency domain spectrum feature set includes the main frequency feature, the main frequency amplitude feature, and the frequency band energy feature. The main frequency is the frequency value corresponding to the frequency component with the largest energy in the frequency domain energy distribution, the main frequency amplitude is the amplitude corresponding to the main frequency, and the frequency band energy is the sum of the energy values ​​of each frequency component within a preset frequency interval.

[0089] Time-frequency joint feature extraction processing is performed on the vibration signal time series data to obtain the vibration signal time-frequency joint feature set. The time-frequency joint feature extraction processing includes decomposing the vibration signal sampling value sequence within the time window segment into time-frequency coefficient sequences of multiple frequency scales, and calculating time-frequency energy features based on the time-frequency coefficient sequences of each frequency scale. The time-frequency energy features are obtained by summarizing the results of squaring the time-frequency coefficient values ​​at each frequency scale according to the time dimension and the frequency scale dimension respectively.

[0090] Statistical calculations and fluctuation characteristic calculations are performed on the time series data of the current signal to obtain a set of statistical features of the current signal and a set of fluctuation characteristics of the current signal. The set of statistical features of the current signal includes the current mean value feature, the current maximum value feature, the current minimum value feature and the current root mean square value feature. The fluctuation characteristic calculation process includes calculating the current change sequence and calculating the fluctuation intensity feature.

[0091] The temperature signal time series data is processed to extract trend features and calculate fluctuation features to obtain a set of temperature signal trend features and a set of temperature signal fluctuation features. The trend feature extraction process includes calculating the temperature change amplitude and the temperature change rate based on the temperature signal sample values ​​at the start and end of the time window segment. The temperature signal fluctuation feature set includes temperature peak-to-peak value features and temperature fluctuation intensity features.

[0092] Feature indices are assigned to the time-domain statistical feature set, frequency-domain spectral feature set, time-frequency joint feature set, current signal statistical feature set, current signal fluctuation feature set, temperature signal trend feature set, and temperature signal fluctuation feature set, respectively. Information source identifiers are assigned to the vibration signal features, current signal features, and temperature signal features, respectively.

[0093] The feature sets are sequentially concatenated according to the information source identifier and feature index to generate structured feature vectors corresponding to each time window segment, and the structured feature vectors are bound and stored with the corresponding working condition labels.

[0094] In this embodiment, obtaining the fusion state representation vector sequence specifically includes:

[0095] Read the structured feature vectors corresponding to each time window segment in the multi-source time window segment set, and obtain the working condition label, information source identifier and feature index bound to each structured feature vector. Sort each structured feature vector according to the time window segment index to form a feature input sequence.

[0096] Based on the working condition labels, conditional reweighting is performed on the feature input sequence to obtain the working condition conditional feature sequence. The conditional reweighting process includes determining the weight set corresponding to each information source identifier according to the working condition labels, and scaling and updating the feature components belonging to the same information source identifier according to the corresponding weights.

[0097] Based on the consistency constraint of information source, the conditional feature sequence of the working condition is subjected to consistency correction processing to obtain the consistency correction feature sequence. The consistency correction processing includes calculating the consistency deviation for the feature sub-vectors corresponding to different information source identifiers within the same time window segment, and performing correction and update on the deviation exceeding the threshold according to the consistency constraint threshold, so that the feature sub-vectors corresponding to different information source identifiers satisfy the consistency constraint conditions.

[0098] A feature space mapping relationship is constructed and a feature space mapping process is performed on the consistency correction feature sequence based on the feature space mapping relationship to obtain a mapped feature sequence. The feature space mapping relationship is used to map the feature sub-vectors corresponding to different information source identifiers to a unified fusion feature space. The feature space mapping process includes determining the corresponding mapping parameter set for each information source identifier and performing a mapping transformation on the feature sub-vectors corresponding to the information source identifier based on the mapping parameter set to obtain the mapped sub-vectors under the unified fusion feature space.

[0099] The mapping vectors under the unified fusion feature space are subjected to fusion generation processing to obtain fusion state representation vectors. The fusion state representation vectors corresponding to each time window segment are combined in the order of time window segment index to obtain a sequence of fusion state representation vectors. The fusion generation processing is to scale the mapping vectors corresponding to different information source identifiers according to the corresponding fusion weights and then sum them. The fusion weights are determined by the working condition labels. The adjustment degree of each mapping vector is corrected according to the consistency deviation obtained by the consistency correction processing.

[0100] In this embodiment, the output of the fault evolution prediction result set specifically includes:

[0101] The fused state representation vector sequence is fragmented and organized to form a fragmented organization sequence for model input. The fragmentation includes reading the fused state representation vector sequence and obtaining the time window segment index, start timestamp and end timestamp of the multi-source time window segment set corresponding to the fused state representation vector sequence. Based on the time window segment index, the fused state representation vector sequence is organized according to the segment order of the multi-source time window segment set.

[0102] Based on the time window segment range defined by the segment organization sequence, the fusion state representation vector sequence is input into the improved TSMixer model. In the in-window time mixing module, in-window mixing modeling is performed on the in-segment time series corresponding to each time window segment to obtain the in-window representation vector sequence corresponding to each time window segment.

[0103] The improvements to the TSMixer model are as follows: the original TSMixer model's method of uniformly performing time mixing on the entire time series is changed to a hierarchical time mixing method based on time window segments and executed separately through the intra-window time mixing module and the inter-window time mixing module; and the unconstrained feature channel mixing method in the feature dimension mixing module of the original model is changed to a feature dimension mixing method constrained by the set of multi-source information relationship constraints.

[0104] The in-window hybrid modeling includes: for each time window segment defined by the segment organization sequence, selecting the fusion state representation vectors corresponding to each moment within each time window segment in chronological order to form the in-segment time series corresponding to each time window segment; in the in-window time hybridization module, rearranging and combining the fusion state representation vectors in the in-segment time series according to their time positions to form a time rearrangement sequence containing the correspondence of different time positions; based on the time rearrangement sequence, performing feature information interaction processing on the fusion state representation vectors corresponding to different moments in the time dimension, so that the fusion state representation vectors corresponding to each moment within the segment can transmit information in the time dimension; after completing the information interaction in the time dimension, performing time dimension feature aggregation processing on the fusion state representation vectors that have undergone interaction processing to generate an in-segment representation vector sequence representing the fault evolution state within each time window segment; and outputting the in-segment representation vector sequence as the in-window representation vector sequence of the corresponding time window segment.

[0105] In the inter-window time mixing module, inter-window mixing modeling is performed on the intra-window representation vector sequences corresponding to multiple time window segments according to the time window segment index order to obtain the inter-window representation vector sequence across time window segments. The inter-window mixing modeling includes: organizing the intra-window representation vectors corresponding to multiple time window segments based on the time window segment index order, and performing cross-segment information interaction and aggregation processing at the time window segment dimension, so that the intra-window representation vectors corresponding to different time window segments form a relationship at the segment level, and generating an inter-window representation vector sequence that can characterize the fault evolution trend across time window segments;

[0106] In the feature dimension hybrid module, a set of multi-source information relationship constraints is determined based on the information source identifier and feature index, and a feature dimension hybrid connection constraint rule is constructed based on the set of multi-source information relationship constraints. The feature dimension hybrid connection constraint rule is used to limit the connection range of the feature dimension hybrid operation, so that the feature dimension hybrid operation is only executed between feature channels that satisfy the multi-source information relationship constraints.

[0107] The determination of the multi-source information relationship constraint set includes: grouping each feature index based on the information source identifier in the structured feature vector to form feature index subsets corresponding to vibration signal features, current signal features, and temperature signal features, respectively; establishing feature channel relationships that allow mutual connection for feature index subsets corresponding to the same information source identifier, which is used to limit the mixed connection range between feature channels within the same information source; determining cross-information source feature index pairs that allow connection based on the alignment relationship of different information source features within the same time window segment for feature index subsets corresponding to different information source identifiers, and forming cross-information source connection constraint relationships, which are used to limit the mixed connection range between feature channels of different information sources; summarizing the connection constraint relationships within the same information source and the cross-information source connection constraint relationships to form a multi-source information relationship constraint set used to describe the allowed connection relationships between feature channels;

[0108] The in-window representation vector sequence and the inter-window representation vector sequence are combined to obtain a time series hybrid representation. When performing feature dimension mixing operation on the time series hybrid representation, the hybrid connection relationship between feature channels is limited according to the feature dimension mixing connection constraint rules to obtain a constrained improved time series hybrid representation sequence. The combination includes: selecting the inter-window representation vector corresponding to each time window segment index from the inter-window representation vector sequence according to the time window segment index order, aligning and associating the inter-window representation vector with the in-window representation vector sequence of the corresponding time window segment, and summarizing the aligned representation sequences according to a unified time axis order.

[0109] The improved time series mixed representation sequence is obtained by: taking the mixed representation vector corresponding to each time position in the time series mixed representation as the processing object, and traversing the mixed representation vector corresponding to each time position one by one; during the traversal, according to the feature dimension mixing connection constraint rules, determining the set of feature channels allowed to participate in feature dimension mixing in the current mixed representation vector, and excluding feature channels that do not meet the constraint rules; for the set of feature channels allowed to participate in mixing, performing feature value combination and update processing on the feature values ​​of the corresponding feature channels to generate updated feature channel values; for feature channels not selected to participate in mixing, keeping the feature values ​​of the corresponding feature channels unchanged; merging the updated feature channel values ​​with the unchanged feature channel values ​​to form the constrained mixed representation vector corresponding to the current time position; arranging the constrained mixed representation vectors corresponding to each time position according to a unified time axis order to generate the improved time series mixed representation sequence.

[0110] The improved time series hybrid representation sequence is input into the fault evolution prediction module for prediction and decoding processing, and the fault evolution prediction result set is output. The fault evolution prediction result set includes the fault risk value sequence, the performance degradation trend sequence, and the remaining service life prediction value sequence.

[0111] The predictive decoding process includes: using the hybrid representation vectors corresponding to each time position in the improved time series hybrid representation sequence as input data, and sequentially performing fault risk calculation, performance degradation calculation, and remaining service life calculation on the hybrid representation vectors corresponding to each time position according to the time axis order; in the fault risk calculation process, for the hybrid representation vector corresponding to each time position, feature components representing fault-related state changes are extracted, and weighted combination calculation is performed on the feature components to obtain the fault risk value for the corresponding time position, and arranged in the time axis order to form a fault risk value sequence; in the performance degradation calculation process, for the hybrid representation vector corresponding to each time position, feature components representing performance state changes are extracted, and the change amplitude of the feature components is calculated in combination with the hybrid representation vectors corresponding to adjacent time positions to obtain the performance degradation trend value for the corresponding time position, and arranged in the time axis order to form a performance degradation trend sequence; in the remaining service life calculation process, for the hybrid representation vector corresponding to each time position, the persistence of the state change trend of the hybrid representation vector is estimated in combination with the chronological relationship of the time position in the time axis, the length of time that the equipment can sustainably operate in the current operating state is calculated, the predicted value of the remaining service life for the corresponding time position is obtained, and arranged in the time axis order to form a remaining service life prediction value sequence.

[0112] In this embodiment, the output of the predicted event set specifically includes:

[0113] Read the set of fault evolution prediction results, and align and organize the fault risk value sequence, performance degradation trend sequence and remaining service life prediction value sequence according to the time window segment index order of the multi-source time window segment set to form a set of prediction result entries that correspond one-to-one with each time window segment;

[0114] The prediction result set is subjected to adjacent time window prediction consistency constraint processing. After the consistency correction update is completed, the risk change magnitude is calculated for each time window segment based on the fault risk value sequence, and the duration of each time window segment is calculated using the start and end timestamps of the time window segment.

[0115] The adjacent time window prediction consistency constraint processing includes: calculating the risk consistency deviation and the degradation consistency deviation based on the difference between the fault risk values ​​and the difference between the performance degradation trends corresponding to two adjacent time window segments, respectively, and comparing the risk consistency deviation and the degradation consistency deviation with the corresponding consistency constraint thresholds. For prediction result entries that exceed the consistency constraint thresholds, consistency correction and updates are performed to ensure that the change magnitude of the fault risk values ​​and the change magnitude of the performance degradation trends corresponding to adjacent time window segments meet the consistency constraint conditions.

[0116] The risk change magnitude is the absolute value of the difference between the fault risk value corresponding to the current time window segment and the fault risk value corresponding to the previous time window segment. The duration of the time window segment is the value of the time corresponding to the end timestamp minus the value corresponding to the start timestamp.

[0117] The event triggering criteria are determined based on the magnitude of risk change and the duration of time window segments. Threshold determination processing is performed on each time window segment to obtain candidate time window segments for events. The threshold determination process compares the magnitude of risk change with the risk change threshold and the duration of time window segments with the duration threshold. Time window segments that simultaneously meet the risk change threshold and duration threshold conditions are marked as candidate time window segments for events.

[0118] The event candidate time window segments are subjected to continuous detection and aggregation processing according to the time window segment index order. Event candidate time window segments with consecutive time window segment indexes are merged into the same predicted event, and a predicted event identifier is generated for each predicted event.

[0119] For each predicted event, a predicted event record is generated and a predicted event set is output. The predicted event set includes a predicted event identifier, a predicted event start time, a predicted event end time, an operating condition label, a fault risk level, and a predicted remaining service life value. The generation of the predicted event record includes: determining the predicted event start time based on the start timestamp of the first event candidate time window segment included in the predicted event, and determining the predicted event end time based on the end timestamp of the last event candidate time window segment included in the predicted event; determining the fault risk level based on the fault risk values ​​covered by the predicted event; extracting the operating condition label corresponding to the predicted event and the corresponding predicted remaining service life value within the coverage area of ​​the predicted event; encapsulating the predicted event identifier, predicted event start time, predicted event end time, operating condition label, fault risk level, and predicted remaining service life value into a predicted event record set, and outputting the predicted event record set as the predicted event set.

[0120] In this embodiment, the predictive maintenance output includes a maintenance suggestion identifier, a suggested maintenance time window, a fault type identifier, and a fault risk level identifier. The maintenance suggestion identifier is determined based on the fault risk level corresponding to the predicted event, the predicted remaining useful life, and the operating condition label, and the determined maintenance suggestion category is encoded. The suggested maintenance time window is determined based on the predicted event end time and the corresponding predicted remaining useful life, wherein the predicted event end time is used as a reference time point, and the start and end times for maintenance execution are determined in combination with the predicted remaining useful life. The fault type identifier is determined based on the abnormal distribution characteristics of each feature component in the fusion state representation vector subsequence corresponding to the predicted event coverage time range, combined with the operating condition label, and the determined fault type is encoded. The corresponding fault risk level identifier is determined based on the fault risk value sequence corresponding to the predicted event, wherein the risk level is determined based on the fault risk values ​​within the predicted event coverage time range and encoded.

[0121] An air compressor fault prediction system based on multi-source information fusion and machine learning includes:

[0122] The data preprocessing module is used to collect multi-source heterogeneous operating data during the operation of the air compressor and preprocess it to obtain a standardized multi-source operating dataset.

[0123] The window labeling module is used to perform window processing on the standardized multi-source operation dataset under a unified time axis, obtain a set of multi-source time window segments, and generate working condition labels for each time window segment.

[0124] The feature construction module is used to extract multi-source features and construct structured feature vectors for each time window segment in a set of multi-source time window segments with working condition labels.

[0125] The fusion modeling module is used to perform multi-source information fusion modeling on structured feature vectors based on the consistency constraints between working condition labels and information sources, and obtain a fusion state representation vector sequence.

[0126] The fault prediction module is used to perform hierarchical hybrid modeling and prediction decoding on the fused state representation vector sequence based on the improved TSMixer model, and output a set of fault evolution prediction results.

[0127] The event generation module is used to perform time consistency processing and event aggregation processing on the fault evolution prediction result set and output the predicted event set.

[0128] The maintenance decision module is used to determine maintenance strategies based on a set of predicted events, generate predictive maintenance output results, and write them to the storage system.

[0129] Example 1: To verify the feasibility of this invention in practice, it was applied to the scenario of industrial air compressor operation monitoring and maintenance decision-making in a manufacturing park. The air compressors in this park operate continuously for extended periods, with frequent changes in load conditions. The large number of machines and long operating times make it difficult to identify potential fault evolution processes in a timely manner using traditional methods relying on manual inspections and single-threshold alarms. This often results in delayed alarms, frequent false alarms, and a lack of basis for maintenance decisions, posing risks to equipment safety and production continuity.

[0130] In this scenario, the air compressor continuously collects multi-source heterogeneous operational data, including vibration, current, and temperature, during operation, and uploads this data in real time to the data processing unit via the field control system. The collected data first undergoes missing data processing, anomaly handling, noise reduction, and dimension unification, and is then aligned and resampled along a unified time axis to form a standardized multi-source operational dataset. Subsequently, based on data sampling characteristics and operational continuity, the data is windowed, and operating condition labels are generated for each time window, combining loading and unloading states with statistical results of operating parameters, thus accurately reflecting the differences in equipment status at different operational stages. Building upon this, multi-dimensional features related to vibration, current, and temperature are extracted from the multi-source signals within each time window, and structured feature vectors are constructed. By introducing operating condition labels and information source consistency constraints, multi-source information fusion modeling is completed, resulting in a fused state representation vector sequence that characterizes the overall operational status changes of the equipment.

[0131] The fused state representation vector sequence is further input into the improved time series hybrid model. Through hierarchical time hybrid modeling, it simultaneously characterizes short-term evolutionary features within a time window and long-term evolutionary trends across time windows. Multi-source information relationship constraints are introduced into the feature dimension to avoid interference from irrelevant features, thereby outputting continuous fault risk evolution information, performance degradation trends, and remaining service life prediction results. For these prediction results, the system performs consistency constraint processing on the time axis and combines the risk change magnitude and duration conditions to perform event-level aggregation of the prediction results, automatically generating a structured set of predicted events, clearly defining the occurrence time range, risk level, and corresponding operating conditions of each potential fault event.

[0132] In actual operation, the predicted event set is further used to generate predictive maintenance output results. Based on the risk level, remaining service life, and operating condition information corresponding to the predicted events, the system automatically generates maintenance suggestion identifiers and suggested maintenance time windows, and outputs them to the maintenance decision interface, providing on-site maintenance personnel with intuitive and actionable maintenance references. Through practical application within the continuous operation cycle of this park, it can be seen that this invention can continuously and stably monitor and analyze the operating status of air compressors without increasing manual intervention, identify fault evolution trends in advance, reduce the risk of sudden downtime, and avoid resource waste caused by over-maintenance, significantly improving equipment operating safety and the scientific nature of maintenance decisions.

[0133] Table 1. Performance Comparison of the Invention and Traditional Air Compressor Fault Prediction Methods

[0134]

[0135] As can be clearly seen from Table 1, the method of the present invention is superior to the traditional method in many indicators.

[0136] In terms of fault identification accuracy, the traditional method achieves an accuracy of 88.4%, while the method of this invention improves it to 92.7%. This improvement mainly stems from the multi-source information fusion modeling introduced at the feature level in this invention. Through operating condition label constraints and information source consistency constraints, vibration, current, and temperature features are collaboratively modeled under a unified time axis and a unified fusion space, reducing the interference of single signal fluctuations on the judgment results, thereby improving the overall identification stability.

[0137] Regarding the false alarm rate, the traditional method has a false alarm rate of 15.3%, while the method of this invention reduces it to 8.2%. This improvement is directly related to the time consistency constraint and event-level aggregation processing. By imposing consistency constraints on the prediction results of adjacent time windows and determining events based on the magnitude and duration of risk changes, short-term misjudgments caused by transient anomalies can be effectively filtered out, avoiding the misidentification of short-period fluctuations as fault events.

[0138] The false negative rate dropped from 9.6% to 7.5%, a decrease of nearly half, mainly due to the improved time series hybrid modeling approach. The hierarchical time hybrid structure can simultaneously capture subtle evolutionary features within a time window as well as long-term degradation trends across time windows, allowing latent, slowly evolving fault modes to appear earlier in the model output, thereby reducing false negatives.

[0139] The early warning time for faults was improved from 6.2 hours to 8.4 hours, significantly enhancing the early warning capability. This result is closely related to the reduction in the remaining useful life prediction error. This invention reduces the remaining useful life prediction error from 18.7% to 14.5%, making the prediction results more stable and reliable over time, thus enabling effective early warnings before fault risks become apparent.

[0140] Regarding the consistency of risk level determination, the traditional method achieves a consistency rate of 82.5%, while this invention improves it to 89.2%. This improvement is mainly due to the event-level risk aggregation mechanism introduced into the predicted event set. By uniformly determining the sequence of fault risk values ​​within the event coverage time range, the problem of frequent jumps in risk levels caused by point-in-time determination is avoided.

[0141] The maintenance recommendation matching rate increased from 79.1% to 87.3%, reflecting a higher degree of matching between the maintenance recommendations generated by this invention and the actual equipment operating status. This improvement is directly related to the combined use of predicted events, operating condition labels, and remaining service life in determining maintenance strategies, enabling maintenance recommendations to be based not only on a single risk value but also on a comprehensive consideration of the operating context and evolutionary trends.

[0142] Regarding maintenance decision response time, this invention reduces the average response time from 18.5 minutes to 15.3 minutes. This is because the predictive maintenance output directly outputs maintenance suggestion identifiers and suggested maintenance time windows in a structured form, reducing manual analysis and secondary judgment processes.

[0143] Finally, the number of unplanned downtimes decreased from 2.6 times / month to 2.1 times / month, indicating that the present invention effectively reduced the risk of production interruption caused by sudden failures in actual operation. The fundamental reason is the continuous modeling and event-level prediction of the failure evolution process, which enables maintenance decisions to be made in advance to a stage where the risk is controllable.

[0144] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting air compressor faults based on multi-source information fusion and machine learning, characterized in that, Includes the following steps: Collect multi-source heterogeneous operating data during the operation of the air compressor, and preprocess it to obtain a standardized multi-source operating dataset; Under a unified timeline, the standardized multi-source operation dataset is windowed to obtain a set of multi-source time window segments, and operating condition labels are generated for each time window segment. For each time window segment in the set of multi-source time window segments with working condition labels, multi-source features are extracted and a structured feature vector is constructed; Based on the consistency constraint between working condition labels and information sources, multi-source information fusion modeling is performed on the structured feature vectors to obtain a sequence of fused state representation vectors ordered by a unified time axis; Based on the improved TSMixer model, hierarchical hybrid modeling and prediction decoding are performed on the fused state representation vector sequence to output a set of fault evolution prediction results. Perform time consistency processing and event aggregation processing on the fault evolution prediction result set, and output the predicted event set; The maintenance strategy is determined based on the set of predicted events, the predictive maintenance output is generated and written to the storage system; The output of the fault evolution prediction result set specifically includes: The fusion state representation vector sequence is fragmented and organized to form a fragmented sequence for model input; Based on the time window segment range defined by the segment organization sequence, the fusion state representation vector sequence is input into the improved TSMixer model. In the in-window time mixing module, in-window mixing modeling is performed on the in-segment time series corresponding to each time window segment to obtain the in-window representation vector sequence corresponding to each time window segment. The improvements to the TSMixer model are as follows: the original TSMixer model's method of uniformly performing time mixing on the entire time series is changed to a hierarchical time mixing method based on time window segments and executed separately through the intra-window time mixing module and the inter-window time mixing module; and the unconstrained feature channel mixing method in the feature dimension mixing module of the original model is changed to a feature dimension mixing method constrained by the set of multi-source information relationship constraints. In the inter-window temporal blending module, inter-window blending modeling is performed on the intra-window representation vector sequences corresponding to multiple time window segments in the order of time window segment index to obtain the inter-window representation vector sequence across time window segments; In the feature dimension hybrid module, a set of multi-source information relationship constraints is determined based on the information source identifier and feature index, and a feature dimension hybrid connection constraint rule is constructed based on the set of multi-source information relationship constraints. The in-window representation vector sequence and the inter-window representation vector sequence are combined to obtain a time series hybrid representation. When performing feature dimension hybridization operation on the time series hybrid representation, the hybrid connection relationship between feature channels is limited according to the feature dimension hybrid connection constraint rule to obtain the constrained improved time series hybrid representation sequence. The improved time series hybrid representation sequence is input into the fault evolution prediction module for prediction decoding processing, and the fault evolution prediction result set is output. The fault evolution prediction result set includes a fault risk value sequence, a performance degradation trend sequence, and a remaining service life prediction value sequence. The output of the predicted event set specifically includes: Read the set of fault evolution prediction results, and align and organize the fault risk value sequence, performance degradation trend sequence and remaining service life prediction value sequence according to the time window segment index order of the multi-source time window segment set to form a set of prediction result entries that correspond one-to-one with each time window segment; The prediction result set is subjected to adjacent time window prediction consistency constraint processing. After the consistency correction update is completed, the risk change magnitude is calculated for each time window segment based on the fault risk value sequence, and the duration of each time window segment is calculated using the start and end timestamps of the time window segment. The event triggering criteria are determined based on the magnitude of risk change and the duration of time window segments. Threshold judgment processing is then performed on each time window segment to obtain candidate time window segments for the event. The event candidate time window segments are subjected to continuous detection and aggregation processing according to the time window segment index order. Event candidate time window segments with consecutive time window segment indexes are merged into the same predicted event, and a predicted event identifier is generated for each predicted event. For each predicted event, a predicted event record is generated and a predicted event set is output. The predicted event set includes a predicted event identifier, a predicted event start time, a predicted event end time, an operating condition label, a fault risk level, and a predicted remaining service life value.

2. The air compressor fault prediction method based on multi-source information fusion and machine learning according to claim 1, characterized in that, The multi-source heterogeneous operating data includes vibration signal time series data, current signal time series data, and temperature time series data. The preprocessing includes performing missing value processing, outlier processing, noise reduction processing, and dimension unification processing on the multi-source heterogeneous operating data, and performing time alignment and resampling processing on the processed multi-source operating data based on the sampling timestamp.

3. The air compressor fault prediction method based on multi-source information fusion and machine learning according to claim 1, characterized in that, The generation of the operating condition label specifically includes: Under a unified timeline, windowing parameters are determined based on a standardized multi-source running dataset. Based on the windowing parameters, the standardized multi-source running dataset is truncated to obtain a set of multi-source time window segments. Time window segment indices are assigned to each time window segment in chronological order. Generate corresponding time window segment identifiers for each time window segment in the multi-source time window segment set and establish time association relationships; Within the time range covered by each time window segment, a set of air compressor operating parameter sequences aligned with the time window segment is obtained. The set of air compressor operating parameter sequences includes a loading state sequence, an unloading state sequence, an exhaust pressure sequence, an exhaust temperature sequence, and a motor current sequence. Statistical calculations were performed on the exhaust pressure sequence, exhaust temperature sequence, and motor current sequence to obtain the mean exhaust pressure, mean exhaust temperature, and mean motor current. By combining the loading state sequence and unloading state sequence within the time window segment, the loading state value and unloading state value corresponding to the time window segment are determined, and the values ​​are integrated with the average exhaust pressure, average exhaust temperature and average motor current to obtain the working condition judgment parameter group. Based on the working condition judgment parameter group, the working condition label corresponding to the time window segment is generated, and the working condition label is bound to the time window segment identifier and written into the multi-source time window segment set.

4. The air compressor fault prediction method based on multi-source information fusion and machine learning according to claim 1, characterized in that, The construction of the structured feature vector specifically includes: Read the vibration signal time series data, current signal time series data and temperature signal time series data corresponding to each time window segment in the multi-source time window segment set with working condition labels, and obtain the corresponding working condition labels; Time-domain statistical calculations and frequency-domain spectrum calculations are performed on the time-series data of vibration signals to obtain the time-domain statistical feature set and the frequency-domain spectrum feature set of vibration signals. Perform time-frequency joint feature extraction processing on the vibration signal time-series data to obtain the vibration signal time-frequency joint feature set; Statistical calculations and fluctuation characteristic calculations are performed on the time-series data of the current signal to obtain a set of statistical characteristics and a set of fluctuation characteristics of the current signal. Perform trend feature extraction and fluctuation feature calculation on the time series data of temperature signals to obtain a set of temperature signal trend features and a set of temperature signal fluctuation features; Feature indices are assigned to the time-domain statistical feature set, frequency-domain spectral feature set, time-frequency joint feature set, current signal statistical feature set, current signal fluctuation feature set, temperature signal trend feature set, and temperature signal fluctuation feature set, respectively. Information source identifiers are assigned to the vibration signal features, current signal features, and temperature signal features, respectively. The feature sets are sequentially concatenated according to the information source identifier and feature index to generate structured feature vectors corresponding to each time window segment, and the structured feature vectors are bound and stored with the corresponding working condition labels.

5. The air compressor fault prediction method based on multi-source information fusion and machine learning according to claim 1, characterized in that, The specific steps involved in obtaining the fusion state representation vector sequence are as follows: Read the structured feature vectors corresponding to each time window segment in the multi-source time window segment set, and obtain the working condition label, information source identifier and feature index bound to each structured feature vector. Sort each structured feature vector according to the time window segment index to form a feature input sequence. Based on the working condition labels, conditional reweighting is performed on the feature input sequence to obtain the working condition conditional feature sequence. Based on the information source consistency constraint, the conditional feature sequence is subjected to consistency correction processing to obtain the consistency-corrected feature sequence. A feature space mapping relationship is constructed, and feature space mapping processing is performed on the consistency correction feature sequence based on the feature space mapping relationship to obtain a mapped feature sequence, wherein the mapped feature sequence includes multiple mapping sub-vectors under a unified fusion feature space; The fusion generation process is performed on the mapper vectors in the unified fusion feature space to obtain the fusion state representation vector. The fusion state representation vectors corresponding to each time window segment are combined in the order of the time window segment index to obtain the fusion state representation vector sequence.

6. The air compressor fault prediction method based on multi-source information fusion and machine learning according to claim 1, characterized in that, The predictive maintenance output includes maintenance suggestion identifiers, suggested maintenance time windows, fault type identifiers, and fault risk level identifiers.

7. An air compressor fault prediction system based on multi-source information fusion and machine learning, comprising executing the air compressor fault prediction method based on multi-source information fusion and machine learning as described in any one of claims 1 to 6, characterized in that, include: The data preprocessing module is used to collect multi-source heterogeneous operating data during the operation of the air compressor and preprocess it to obtain a standardized multi-source operating dataset. The window labeling module is used to perform window processing on the standardized multi-source operation dataset under a unified time axis, obtain a set of multi-source time window segments, and generate working condition labels for each time window segment. The feature construction module is used to extract multi-source features and construct structured feature vectors for each time window segment in a set of multi-source time window segments with working condition labels. The fusion modeling module is used to perform multi-source information fusion modeling on structured feature vectors based on the consistency constraints between working condition labels and information sources, and obtain a fusion state representation vector sequence. The fault prediction module is used to perform hierarchical hybrid modeling and prediction decoding on the fused state representation vector sequence based on the improved TSMixer model, and output a set of fault evolution prediction results. The event generation module is used to perform time consistency processing and event aggregation processing on the fault evolution prediction result set and output the predicted event set. The maintenance decision module is used to determine maintenance strategies based on the set of predicted events, generate predictive maintenance output results, and write them to the storage system. The output of the fault evolution prediction result set specifically includes: The fusion state representation vector sequence is fragmented and organized to form a fragmented sequence for model input; Based on the time window segment range defined by the segment organization sequence, the fusion state representation vector sequence is input into the improved TSMixer model. In the in-window time mixing module, in-window mixing modeling is performed on the in-segment time series corresponding to each time window segment to obtain the in-window representation vector sequence corresponding to each time window segment. The improvements to the TSMixer model are as follows: the original TSMixer model's method of uniformly performing time mixing on the entire time series is changed to a hierarchical time mixing method based on time window segments and executed separately through the intra-window time mixing module and the inter-window time mixing module; and the unconstrained feature channel mixing method in the feature dimension mixing module of the original model is changed to a feature dimension mixing method constrained by the set of multi-source information relationship constraints. In the inter-window temporal blending module, inter-window blending modeling is performed on the intra-window representation vector sequences corresponding to multiple time window segments in the order of time window segment index to obtain the inter-window representation vector sequence across time window segments; In the feature dimension hybrid module, a set of multi-source information relationship constraints is determined based on the information source identifier and feature index, and a feature dimension hybrid connection constraint rule is constructed based on the set of multi-source information relationship constraints. The in-window representation vector sequence and the inter-window representation vector sequence are combined to obtain a time series hybrid representation. When performing feature dimension hybridization operation on the time series hybrid representation, the hybrid connection relationship between feature channels is limited according to the feature dimension hybrid connection constraint rule to obtain the constrained improved time series hybrid representation sequence. The improved time series hybrid representation sequence is input into the fault evolution prediction module for prediction decoding processing, and the fault evolution prediction result set is output. The fault evolution prediction result set includes a fault risk value sequence, a performance degradation trend sequence, and a remaining service life prediction value sequence. The output of the predicted event set specifically includes: Read the set of fault evolution prediction results, and align and organize the fault risk value sequence, performance degradation trend sequence and remaining service life prediction value sequence according to the time window segment index order of the multi-source time window segment set to form a set of prediction result entries that correspond one-to-one with each time window segment; The prediction result set is subjected to adjacent time window prediction consistency constraint processing. After the consistency correction update is completed, the risk change magnitude is calculated for each time window segment based on the fault risk value sequence, and the duration of each time window segment is calculated using the start and end timestamps of the time window segment. The event triggering criteria are determined based on the magnitude of risk change and the duration of time window segments. Threshold judgment processing is then performed on each time window segment to obtain candidate time window segments for the event. The event candidate time window segments are subjected to continuous detection and aggregation processing according to the time window segment index order. Event candidate time window segments with consecutive time window segment indexes are merged into the same predicted event, and a predicted event identifier is generated for each predicted event. For each predicted event, a predicted event record is generated and a predicted event set is output. The predicted event set includes a predicted event identifier, a predicted event start time, a predicted event end time, an operating condition label, a fault risk level, and a predicted remaining service life value.

Images