A semantic alignment data generation method suitable for hydrogen compressor fault diagnosis

Abstract

This invention discloses a semantically aligned data generation method for hydrogen compressor fault diagnosis, comprising the following steps: receiving raw multidimensional time-series data of hydrogen compression and transportation processes from a hydrogen refueling station compressor; performing a first-level full-variable safety scan on a standard time-series dataset using an out-of-bounds judgment function; identifying latent compressor faults through pressure-temperature coupling verification, inter-stage cascade mutual verification, and differential pressure flow direction mutual verification; performing a third-level adaptive operating condition clustering on unlabeled data; constructing compressor fault diagnosis question-and-answer pairs; obtaining a compressor time-series command dataset; and realizing hydrogen refueling station compressor fault diagnosis. This invention effectively solves the technical problems in existing technologies, such as the lack of supervision signals in raw data of hydrogen refueling station compressors, difficulty in identifying latent faults under complex operating conditions, and difficulty in adapting heterogeneous feature space models, by constructing a three-layer semantic enhancement logic and model adaptation strategy.

Description

Technical Field

[0001] This invention relates to the field of data processing, specifically a semantically aligned data generation method suitable for fault diagnosis of hydrogen compressors. Background Technology

[0002] With the rapid development of the hydrogen energy industry, hydrogen refueling station compressors generate a large amount of multi-dimensional time-series sensor data (such as pressure, temperature, vibration, etc.) during operation. To ensure safety, the industry typically employs deep learning-based fault diagnosis methods to monitor equipment status. In recent years, with the rise of Large Language Models (LLM), time-series-based LLMs (such as ChatTs, ITFormer, etc.) have become a research hotspot. This type of technology maps time-series data into embedding vectors that can be understood by large models, and combines textual modalities for multimodal learning, aiming to leverage the powerful generalization and reasoning capabilities of large models to handle industrial fault diagnosis tasks. The existing general process typically involves: collecting time-series data and corresponding fault labels from the industrial site, formatting it into model-acceptable input, and then using a pre-trained model for fine-tuning or inference to achieve classification or prediction of equipment status.

[0003] In practical applications, when applying the aforementioned general time-series large model technology to the specific industrial scenario of hydrogen refueling station compressor fault diagnosis, at least three significant defects and difficulties exist: First, the raw data is completely semantically missing, lacking basic operating condition labels as monitoring signals. The underlying control system (PLC or SCADA) of a hydrogen refueling station typically only records raw physical sensor values, without recording system status codes or operating condition labels representing the system's logical state. This means that massive amounts of historical data are merely a pile of numbers without context, lacking both discrete labels for normal or faulty operation and operating condition markers for refueling / standby / unloading. Existing technologies often struggle when processing such data: supervised learning algorithms cannot be directly applied due to the lack of target labels; while traditional unsupervised algorithms struggle to automatically distinguish between normal standby and fault-induced shutdowns. The lack of an adaptive clustering mechanism that can automatically mine and define system modes, such as standby, refueling, and high-pressure maintenance, based on raw physical data from scratch prevents the release of data value.

[0004] Secondly, single-dimensional threshold monitoring struggles to capture latent faults based on physical coupling. Most existing monitoring methods rely on static thresholds of a single variable, such as simply determining whether pressure exceeds limits. However, many faults in hydrogen refueling station compressors do not manifest as a single numerical deviation but rather as a disruption of physical consistency. For example, according to the thermodynamic PVT relationship, a rise in pressure is usually accompanied by a rise in temperature; according to the source-load flow relationship, the source pressure should be higher than the destination pressure. Existing technologies lack this dual-modal coupling diagnostic capability, failing to identify latent faults such as false filling, inefficiency, or sensor drift—which fall within threshold ranges but violate physical laws—through logical reasoning. Furthermore, existing technologies struggle to handle concurrent faults, failing to retain composite fault information through tag splicing mechanisms, easily leading to missed detections.

[0005] Finally, there is a severe dimensionality mismatch between heterogeneous industrial short-term features and the pre-training interface of general-purpose large-scale models. The feature dimensions and time windows of fault characteristics in data collected from industrial sites are determined by specific physical equipment, exhibiting high specialization. However, current mainstream time-series large-scale models, in pursuit of generalization capabilities, typically have fixed specific input interfaces. Existing technologies lack an effective, non-intrusive method for feature alignment and adaptation. Directly using raw data will cause the model to fail to load; retraining the model from scratch is not only computationally expensive but also fails to utilize the general time-series pattern knowledge contained in the pre-training of large-scale models. This makes it difficult to migrate and deploy advanced AI models to edge computing devices at hydrogen refueling stations at low cost. Summary of the Invention

[0006] The purpose of this invention is to provide a semantically aligned data generation method suitable for fault diagnosis of hydrogen compressors, comprising the following steps:

[0007] Step 1) The server receives the raw multidimensional time series data of the hydrogen compression and transportation process of the hydrogen refueling station compressor. The dataset is then preprocessed to generate a standard time-series dataset; The time window length, Number of feature channels;

[0008] Step 2) Use the out-of-bounds detection function A first-level full-variable safety scan is performed on the standard time-series dataset to generate a set of high-risk operating condition labels for data that do not meet the safety threshold. ;

[0009] Step 3) Perform a second-level physical relationship coupling diagnosis on the unlabeled data, and identify the compressor's hidden faults through pressure-temperature coupling verification, inter-stage cascade mutual verification, and differential pressure flow direction mutual verification;

[0010] Step 4) Perform third-layer adaptive working condition clustering on the unlabeled data to map the standard time series data to discrete working condition labels. This enables the clustering and labeling of the remaining unlabeled time series data;

[0011] Step 5) Based on the labels from Steps 2) to 4), construct a semantic mapping function. With question-answering generation functions And construct a question-and-answer pair for compressor fault diagnosis;

[0012] Step 6) Obtain the data generated by the compressor during the compression and delivery of hydrogen, and preprocess it to obtain the compressor timing instruction dataset;

[0013] Step 7) Use the compressor fault diagnosis question and answer pair to identify the corresponding fault type in the compressor timing instruction dataset to realize the compressor fault diagnosis of the hydrogen refueling station.

[0014] Furthermore, in step 1), the preprocessing steps include data cleaning and missing value imputation.

[0015] Furthermore, in step 2), the out-of-bounds judgment function As shown below:

[0016] (1)

[0017] In the formula, For a moment The Data from individual sensors; ; , These are the upper and lower threshold values.

[0018] Furthermore, a set of high-risk working condition labels is generated. Then, a tag splicing mechanism is used to generate composite tags. Or string mapping functions This is to preserve the associated information of concurrent failures.

[0019] Composite label As shown below:

[0020] (2)

[0021] The string mapping function is shown below:

[0022] (3)

[0023] In the formula, For transient atomic tags.

[0024] Furthermore, in step 3), the latent compressor faults include compressor idling, thermocouple failure, interstage pressurization failure, and false charging.

[0025] When performing inter-level cascading mutual verification, if If so, it is determined that the interstage boost has failed; Minimum compression ratio; , The pressure is divided into two stages: exhaust pressure and exhaust pressure.

[0026] When performing differential pressure flow direction verification, if there exists a slope that satisfies... However, data that does not meet the physical constraints of a valid filling condition is diagnosed as false inflation and an anomaly label is generated. ;

[0027] The data diagnosed as counterfeit inflation parts are shown below:

[0028] (4)

[0029] In the formula, To measure the secondary exhaust pressure; The target hydrogen storage cylinder pressure.

[0030] Furthermore, in step 3), the steps for performing pressure-temperature coupling verification include:

[0031] Step 3.1) Define the device operating status determination function. Pressure increment With temperature increment ;

[0032] Among them, the equipment operating status determination function As shown below:

[0033] (5)

[0034] In the formula, Indicates that the program is running; , For vibration signals and thresholds;

[0035] Step 3.2) Determine the device operating status using the device operation status function. When the equipment is detected to be running, a compressor idling determination is performed. If the non-consistency constraint is met, the system indicates either compressor idling or thermocouple failure, and a fault tag is generated. ;

[0036] The non-consistency constraints are as follows:

[0037] (6)

[0038] in For pressure increment, For temperature increment, This is the temperature increment threshold; This is the pressure increment threshold.

[0039] Furthermore, in step 4), the step of performing third-layer adaptive working condition clustering on the unlabeled data includes:

[0040] An adaptive clustering algorithm based on Euclidean distance is used to partition the unlabeled sample set. The operating mode is determined, and the probability distribution of a sample belonging to a certain operating mode is calculated. ;

[0041] Among them, the unlabeled sample set is divided. When operating under certain modalities, the sum of squared errors of intra-cluster temporal characteristics is minimized. For the goal;

[0042] Squared error As shown below:

[0043] (7)

[0044] in, For the first Statistical feature vectors of time windows, For the first The center vector of each working condition cluster;

[0045] The probability distribution of a sample belonging to a certain working condition As shown below:

[0046] (8)

[0047] In the formula, For the first The center vector of each working condition cluster.

[0048] Furthermore, semantic mapping functions With question-answering generation functions As shown below:

[0049] (9)

[0050] (10)

[0051] in For fault labels, For operating condition labels, For expert knowledge base; The features of input X.

[0052] Furthermore, in step 6), the preprocessing step refers to reconstructing the input sequence, and the number of time-series tokens after reconstruction. L, P, and s represent the length of the input sequence, the length of the patch, and the step size, respectively.

[0053] Further, in step 7), after identifying the fault type corresponding to the fault type in the compressor timing instruction dataset, the pre-trained large language model is supervisedly fine-tuned using the compressor timing instruction dataset. The general timing knowledge of the pre-trained large language model is combined to infer the evolution trend of the compressor in the future time window and predict the evolution trend of the compressor in the future time window.

[0054] Based on the diagnostic and prediction results, retrieve the contingency plans from the knowledge base and generate operation and maintenance suggestions.

[0055] The technical effects of this invention are undeniable. By constructing a three-layer semantic enhancement logic and model adaptation strategy, this invention effectively solves the technical problems in the prior art, such as the lack of supervision signals in the raw data of hydrogen refueling station compressors, the difficulty in identifying hidden faults under complex operating conditions, and the difficulty in adapting heterogeneous feature space models. It has the following significant beneficial effects:

[0056] This invention overcomes the bottleneck of cold start with unlabeled industrial data, achieving high-quality transformation from unsupervised data to strongly supervised instructions and significantly reducing data annotation costs. It innovatively proposes a three-layer progressive semantic enhancement mechanism: full variable scanning, physical coupling relationship diagnosis, and adaptive clustering. This mechanism automatically transforms massive, messy raw time-series data into structured samples with precise labels and operational context, turning previously unusable dormant data into high-value training resources.

[0057] This invention achieves accurate capture of latent faults based on a physical coupling mechanism, significantly reducing the false positive rate and filling the gap in latent fault detection.

[0058] This invention achieves complete semantic preservation of complex concurrent faults, overcoming the information loss problem of single-label classification. In industrial settings, equipment failures often trigger a chain reaction; the label splicing mechanism employed in this invention can completely record concurrent faults.

[0059] This invention achieves low-cost model adaptation for heterogeneous feature spaces, improving edge inference efficiency. By modifying the model input projection layer structure and employing a SmartLoading strategy, this invention automatically removes weight layers with mismatched shapes.

[0060] This invention constructs an operation and maintenance assistance system with a thought chain capability, and the output results have strong interpretability. The final product of this invention is instruction fine-tuning data in JSONL format, and the model can output a complete thought chain including perception stage 1, diagnosis stage 2, prediction stage 3, and decision stage 4. Attached Figure Description

[0061] Figure 1 is a diagram of the overall system architecture of the present invention;

[0062] Figure 2 is a flowchart of the core process of constructing the three-layer data structure of the present invention;

[0063] Figure 3 is a schematic diagram of the adaptation and inference principle of heterogeneous feature space model in an embodiment of the present invention. Detailed Implementation

[0064] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.

[0065] Example 1:

[0066] See Figures 1 to 3 A semantically aligned data generation method suitable for hydrogen compressor fault diagnosis includes the following steps:

[0067] Step 1) The server receives the raw multidimensional time series data of the hydrogen compression and transportation process of the hydrogen refueling station compressor. The dataset is then preprocessed to generate a standard time-series dataset; The time window length, Number of feature channels;

[0068] Step 2) Use the out-of-bounds detection function A first-level full-variable safety scan is performed on the standard time-series dataset to generate a set of high-risk operating condition labels for data that do not meet the safety threshold. ;

[0069] Step 3) Perform a second-level physical relationship coupling diagnosis on the unlabeled data, and identify the compressor's hidden faults through pressure-temperature coupling verification, inter-stage cascade mutual verification, and differential pressure flow direction mutual verification;

[0070] Step 4) Perform third-layer adaptive working condition clustering on the unlabeled data to map the standard time series data to discrete working condition labels. This enables the clustering and labeling of the remaining unlabeled time series data;

[0071] Step 5) Based on the labels from Steps 2) to 4), construct a semantic mapping function. With question-answering generation functions And construct a question-and-answer pair for compressor fault diagnosis;

[0072] Step 6) Obtain the data generated by the compressor during the compression and delivery of hydrogen, and preprocess it to obtain the compressor timing instruction dataset;

[0073] Step 7) Use the compressor fault diagnosis question and answer pair to identify the corresponding fault type in the compressor timing instruction dataset to realize the compressor fault diagnosis of the hydrogen refueling station.

[0074] Example 2:

[0075] A semantically aligned data generation method for hydrogen compressor fault diagnosis is technically the same as in Embodiment 1. Further, in step 1), the preprocessing steps include data cleaning and missing value imputation, as detailed below:

[0076] First, perform data cleaning to remove outlier data points that are physically impossible or statistically significantly different (e.g., more than three standard deviations). Second, interpolate missing time points to complete the data, preferably using linear interpolation or time series decomposition-based methods.

[0077] Example 3:

[0078] A semantically aligned data generation method for hydrogen compressor fault diagnosis is provided, with the technical content identical to any one of embodiments 1-2. Further, in step 2), the out-of-bounds judgment function... As shown below:

[0079] (1)

[0080] In the formula, For a moment The Data from individual sensors; ; , These are the upper and lower threshold values.

[0081] Example 4:

[0082] A semantic alignment data generation method for hydrogen compressor fault diagnosis is provided, with the same technical content as any one of embodiments 1-3, further comprising generating a set of high-risk operating condition labels. Then, a tag splicing mechanism is used to generate composite tags. Or string mapping functions This is to preserve the associated information of concurrent failures.

[0083] Composite label As shown below:

[0084] (2)

[0085] The string mapping function is shown below:

[0086] (3)

[0087] In the formula, For transient atomic tags.

[0088] Example 5:

[0089] A semantically aligned data generation method for hydrogen compressor fault diagnosis has the same technical content as any one of embodiments 1-4. Further, in step 3), the latent compressor faults include compressor idling, thermocouple failure, interstage pressurization failure, and false charging.

[0090] When performing inter-level cascading mutual verification, if If so, it is determined that the interstage boost has failed; Minimum compression ratio; , The pressure is divided into two stages: exhaust pressure and exhaust pressure.

[0091] When performing differential pressure flow direction verification, if there exists a slope that satisfies... However, data that does not meet the physical constraints of a valid filling condition is diagnosed as false inflation and an anomaly label is generated. ;

[0092] The data diagnosed as counterfeit inflation parts are shown below:

[0093] (4)

[0094] In the formula, To measure the secondary exhaust pressure; The target hydrogen storage cylinder pressure.

[0095] Example 6:

[0096] A semantic alignment data generation method for hydrogen compressor fault diagnosis has the same technical content as any one of embodiments 1-5. Further, in step 3), the step of performing pressure-temperature coupling verification includes:

[0097] Step 3.1) Define the device operating status determination function. Pressure increment With temperature increment ;

[0098] Among them, the equipment operating status determination function As shown below:

[0099] (5)

[0100] In the formula, Indicates that the program is running; , For vibration signals and thresholds;

[0101] Step 3.2) Determine the device operating status using the device operation status function. When the equipment is detected to be running, a compressor idling determination is performed. If the non-consistency constraint is met, the system indicates either compressor idling or thermocouple failure, and a fault tag is generated. ;

[0102] The non-consistency constraints are as follows:

[0103] (6)

[0104] in For pressure increment, For temperature increment, This is the temperature increment threshold; This is the pressure increment threshold.

[0105] Example 7:

[0106] A semantically aligned data generation method for hydrogen compressor fault diagnosis has the same technical content as any one of embodiments 1-6. Further, in step 4), the step of performing third-layer adaptive operating condition clustering on the unlabeled data includes:

[0107] An adaptive clustering algorithm based on Euclidean distance is used to partition the unlabeled sample set. The operating mode is determined, and the probability distribution of a sample belonging to a certain operating mode is calculated. ;

[0108] Among them, the unlabeled sample set is divided. When operating under certain modalities, the sum of squared errors of intra-cluster temporal characteristics is minimized. For the goal;

[0109] Squared error As shown below:

[0110] (7)

[0111] in, For the first Statistical feature vectors of time windows, For the first The center vector of each working condition cluster;

[0112] The probability distribution of a sample belonging to a certain working condition As shown below:

[0113] (8)

[0114] In the formula, For the first The center vector of each working condition cluster.

[0115] Example 8:

[0116] A semantically aligned data generation method for hydrogen compressor fault diagnosis is provided, with the technical content identical to any one of embodiments 1-7, further comprising a semantic mapping function. With question-answering generation functions As shown below:

[0117] (9)

[0118] (10)

[0119] in For fault labels, For operating condition labels, For expert knowledge base; The features of input X.

[0120] Example 9:

[0121] A semantic alignment data generation method for hydrogen compressor fault diagnosis has the same technical content as any one of embodiments 1-8. Further, in step 6), the preprocessing step refers to reconstructing the input sequence, and the number of reconstructed time-series tokens... L, P, and s represent the length of the input sequence, the length of the patch, and the step size, respectively.

[0122] Example 10:

[0123] A semantic alignment data generation method for hydrogen compressor fault diagnosis is provided. The technical content is the same as any one of embodiments 1-9. Further, in step 7), after identifying the fault type corresponding to the fault type in the compressor timing instruction dataset, the pre-trained large language model is supervisedly fine-tuned using the compressor timing instruction dataset. This allows the model to learn the unique physical coupling and fault logic of the hydrogen refueling station compressor (i.e., the hydrogen refueling station domain knowledge injected into the fine-tuning). Combined with the general timing knowledge possessed by the pre-trained large language model (the parameters and capabilities inherent in the pre-trained large language model itself, such as the ability to identify waveform rise, fall, and periodicity), the evolution trend of the compressor in the future time window is deduced, and the evolution trend of the compressor in the future time window is predicted.

[0124] Based on the diagnostic and prediction results, retrieve the contingency plans from the knowledge base and generate operation and maintenance suggestions.

[0125] Example 11:

[0126] To address the issues of raw hydrogen refueling station data lacking discrete fault labels, ambiguous operating condition semantics, and mismatch between feature space and pre-trained models, this invention constructs a three-layer semantic enhancement logic consisting of a threshold perception layer, a physical coupling diagnosis layer, and an operating condition clustering layer. This layer maps unsupervised time-series data into thought chain instruction data with causal structure step by step.

[0127] like Figure 1 This invention provides a schematic diagram of the overall architecture of a hydrogen refueling station fault diagnosis data generation and model adaptation system. The system covers the entire process from the physical site to decision output. First, the high-pressure hydrogen cylinder cluster trailer, compressor, hydrogen storage tank, and refueling mechanism constitute the multi-stage physical process flow of a hydrogen refueling station. Second, sensor acquisition units are deployed at key nodes of the aforementioned equipment to collect 23-dimensional time-series data containing physical quantities such as pressure, temperature, and flow rate in real time. It should be noted that this raw data does not contain fault labels or operating condition markers during the acquisition phase. Subsequently, the data enters the hydrogen refueling station time-series question-and-answer dataset construction module, where it undergoes semantic enhancement processing to be transformed into structured instruction data. Finally, the data is input into the model adaptation and training module, which uses a large model to generate diagnostic results in natural language form to assist maintenance personnel in decision-making.

[0128] like Figure 2 The diagram shown is a flowchart of a data generation and model adaptation method in an embodiment of the present invention, including the following steps:

[0129] Step 201, Data Acquisition and Preprocessing.

[0130] The server receives raw multidimensional time series data uploaded by the hydrogen refueling station control system. ,in The time window length, This represents the number of feature channels. At this point, the data is in an unlabeled state.

[0131] Step 202: Perform the first-level full-variable safety scan. Using the preset "Full-Variable Safety Threshold Configuration Table," perform a bidirectional scan of all sensor channels and define the time intervals. The The sensor data is Set its corresponding lower threshold as and upper limit threshold are The system uses the following out-of-bounds detection function. Encode the data status:

[0132]

[0133] Based on this determination, a set of instantaneous atomic tags is generated. Regarding the time window Multiple anomalies appearing within the code are used to generate composite tags using a tag concatenation mechanism.

[0134]

[0135] Or it can be represented as a string mapping function:

[0136]

[0137] For example, generating Critical_UnderLimit_p_in|Critical_OverLimit_t_water completely preserves the correlation information of concurrent failures.

[0138] Step 203: Perform the second-layer dual-modal coupling diagnosis. Utilizing the inherent physical coupling between the hydrogen refueling station compressor, hydrogen storage cylinder, and cooling system, multivariate logical cross-verification is performed to identify latent faults that meet a single threshold range but violate physical laws. Specifically, this includes the following four sub-steps:

[0139] Perform pressure-temperature coupled verification: First, define the device operating status determination function. When vibration signal Exceeding the threshold The time is determined to be running:

[0140]

[0141] Define pressure increment With temperature increment .based on Equations are used to construct the logic for detecting hidden faults:

[0142] True operation judgment: If the exhaust pressure rises, the exhaust pre-cooling temperature must show a synchronous upward trend;

[0143] Sensor fault diagnosis: If the equipment is running and the pressure rises but the temperature does not rise, that is:

[0144]

[0145] At this point, the Fault_Sensor_Error label is generated.

[0146] Motor idling detection: If the vibration signal is high, but the exhaust pressure and temperature do not rise, it is determined that the belt is broken or the intake valve is not open, generating a Fault_Inefficiency_Idle (motor idling) tag. The above detection rule can be further formalized into the following mathematical constraints, based on the differential form of the ideal gas law, and the exhaust pressure change rate... With exhaust temperature change rate Positive correlation coupling should be satisfied:

[0147]

[0148] When the device is detected to be in operation (vibration signal) When the following inconsistency constraints are satisfied:

[0149]

[0150] in For pressure increment, For temperature increment, This is the noise threshold. At this point, it is determined that the system is experiencing compressor idling or thermocouple failure, and a fault tag is generated. .

[0151] 2. Perform inter-stage cascade verification: Based on the sequential nature of gas flow, verify the first-stage exhaust pressure. With secondary exhaust pressure The cascading relationship.

[0152] Interstage pressure ratio constraint: If (in If the minimum compression ratio is reached, then the interstage supercharging failure is determined.

[0153] Waveform consistency verification: Calculate the Pearson correlation coefficient between the two pressure levels. .like If the waveforms are out of sync, an Inter_Stage_Fault label will be generated.

[0154] 3. Perform differential pressure flow verification: Based on fluid mechanics principles, fluid flows from high pressure to low pressure, verifying the physical validity of the filling process. Monitor the differential pressure relationship between the secondary exhaust pressure and the terminal pressure:

[0155] Valid filling determination: The necessary physical condition for confirming Mode_Filling is: secondary exhaust pressure > target hydrogen storage cylinder current pressure + pipeline resistance constant;

[0156] False filling identification: If the pressure slope of the time-series data is greater than 0, but the measured secondary exhaust pressure is lower than the hydrogen storage tank pressure, it is judged as a physically impossible false filling state, which may be due to sensor drift or filling of non-monitored tank groups. This sample is marked as Ambiguous and is discarded or downweighted. The above judgment rule can be further formalized into the following mathematical constraints, defining the secondary exhaust pressure as... The target hydrogen storage cylinder pressure is The pipeline resistance coefficient is The system determines the current physical constraints for effective filling (Mode_Filling) as follows:

[0157]

[0158] If within the time window Within, the data satisfies the slope. (i.e., nominally the pressure increases), but this violates the above inequality:

[0159]

[0160] The sample is then determined to be either "false inflation" or "sensor drift" fault, and an anomaly label is generated. .

[0161] 4. Perform cooling efficiency verification: Based on the principle of heat exchange, verify the performance status of the cooling system. Compare the pre-cooling and post-cooling temperatures of the first / secondary exhaust gases, and combine this with the cooling water inlet temperature for joint diagnosis to define cooling efficiency indicators. :

[0162]

[0163] like If the efficiency threshold is used, it is determined that the cooler is scaled or the water flow is insufficient, and a Cooling_Inefficiency tag is generated.

[0164] Step 204: Perform the third-layer adaptive working condition clustering.

[0165] For unlabeled sample sets An adaptive clustering algorithm based on Euclidean distance is used to divide the operating conditions into modes. The set of operating condition clusters is defined as follows: ,in The preset number of operating conditions (e.g., standby, refueling, unloading) is used. The algorithm aims to minimize the sum of squared errors of intra-cluster temporal characteristics. :

[0166]

[0167] in, For the first The statistical feature vector (including mean, variance, and kurtosis) of each time window. For the first The center vector of each working condition cluster can be calculated. Simultaneously, the probability distribution of a sample belonging to a specific working condition can be calculated.

[0168]

[0169] Finally, the original time-series data is mapped to discrete operating condition labels. This enables semantic completion of unsupervised data.

[0170] Step 205: Generate multimodal thought chain instruction data. Based on the rich semantic tags generated in steps 202-204, construct a semantic mapping function. With question-answering generation functions :

[0171]

[0172]

[0173] in For fault labels, For operating condition labels, This is an expert knowledge base (Action Map). Based on this, question-answer pairs with four stages are generated, as shown at the end of the document:

[0174] Stage 1 perception: Based on computational time-series statistical features such as slope, linearity, and fluctuation variance, generate natural language descriptions of the physical trends of signals, such as detecting a non-linear and rapid decrease in secondary exhaust pressure;

[0175] Stage 2 Diagnosis: The cleaned aggregated labels are transformed into a multiple-choice question-and-answer task for fault diagnosis, requiring the model to identify the correct fault type from the distracting options, such as identifying interstage leakage or sensor drift.

[0176] Stage 3 prediction: Based on the currently diagnosed fault status and historical time sequence characteristics, perform time sequence deduction to predict the evolution trend of the system in the future time window, such as predicting whether the leak will worsen or quantitatively assessing the safety risk level of the current working condition.

[0177] Stage 4 Decision Making: Based on the diagnostic and prediction results, retrieve contingency plans from the knowledge base and generate specific operational and maintenance recommendations, such as immediately triggering an ESD emergency shutdown and checking the cooling water valve opening. The final output includes training data in JSONL format containing time-series inputs, text commands, and text responses.

[0178] Step 206: Model adaptation and training in heterogeneous feature spaces.

[0179] To adapt to the short-time characteristics of hydrogen refueling stations, the input sequence is reconstructed. The length of the input sequence is defined as... (Value 128), Patch length is (Value 16), step size is (Value 16). This determines the number of time-series tokens generated. The calculation is as follows:

[0180]

[0181] The constant 2 includes special markers (CLS) for the beginning and end of the sequence. Based on this formula, the present invention dynamically adjusts the number of placeholders in the text modality, ensuring accurate alignment of heterogeneous feature spaces.

[0182] Construct the target model architecture, setting the input layer dimension to 23. Load the pre-trained model weight set. At that time, intelligent weight filtering is performed, retaining only the weight subset that matches the shape. :

[0183]

[0184] At the same time, define the Time Token mapping function for the input data. Patch length The value was adjusted to 16 to accommodate the short window. Finally, the input layer was randomly initialized and supervised fine-tuning was performed.

[0185] Figure 3The inference process mainly consists of three parts: a multimodal input layer, a core encoding and alignment layer, and an LLM cognitive inference layer. In the input layer, the system receives 23-dimensional multi-source sensor data from the hydrogen refueling station site, as well as textual inquiries from maintenance personnel, such as "Why is the current pressure fluctuating?". In the core encoding and alignment layer, the short-time-series data is first segmented into fine-grained slices (Patch_len=16) using the Time Patching module, and then mapped to Time Tokens using the FeatureEncoder. Subsequently, utilizing the semantic alignment mechanism proposed in this invention, mismatched input projection layers are automatically filtered when loading pre-trained weights, enabling the model to directly adapt to the 23-dimensional heterogeneous feature space. The time-series tokens and the tokenized text InstructionTokens are then deeply fused in the latent space. Finally, in the LLM cognitive reasoning layer, the large language model receives the fused multimodal vectors and performs joint reasoning based on pre-trained general temporal knowledge and fine-tuned injected hydrogen refueling station domain knowledge. The final output includes natural language text containing root cause analysis of faults, such as secondary piston ring wear and treatment suggestions, realizing end-to-end diagnosis from low-level signals to high-level semantics.

Claims

1. A semantic aligned data generation method suitable for hydrogen compressor fault diagnosis, characterized in that, Includes the following steps: Step 1) The server receives the original multi-dimensional time series data in the process of hydrogen compression and transportation by the hydrogen station compressor , and carries out preprocessing to generate a standard time series data set; wherein is the length of the time window, is the number of feature channels; Step 2) using the out-of-bound determination function performing a first tier full variable security scan on the standard timing data set to generate a set of high risk condition tags for data that does not meet a security threshold ; Step 3) Perform a second-level physical relationship coupling diagnosis on the unlabeled data, and identify the compressor's hidden faults through pressure-temperature coupling verification, inter-stage cascade mutual verification, and differential pressure flow direction mutual verification; Step 4) performing third layer adaptive working condition clustering on the unlabeled data, mapping the standard time series data into discrete working condition labels , so as to realize clustering labeling of the remaining unlabeled time series data; Step 5) Constructing a semantic mapping function based on the labels of steps 2) - step 4) With the question and answer generation function And constructing a compressor fault diagnosis question and answer pair; Step 6) Obtain the data generated by the compressor during the compression and delivery of hydrogen, and preprocess it to obtain the compressor timing instruction dataset; Step 7) Use the compressor fault diagnosis question and answer pair to identify the corresponding fault type in the compressor timing instruction dataset to realize the compressor fault diagnosis of the hydrogen refueling station. 2.The semantic alignment data generation method for hydrogen compressor fault diagnosis according to claim 1, wherein, In step 1), the preprocessing steps include data cleaning and missing value imputation.

3. The semantically aligned data generation method for hydrogen compressor fault diagnosis according to claim 1, characterized in that, In step 2), the out-of-bounds judgment function As shown below: ；（1） In the formula, For a moment The Data from individual sensors; ; , These are the upper and lower threshold values.

4. The semantically aligned data generation method for hydrogen compressor fault diagnosis according to claim 1, characterized in that, Generate a set of high-risk working condition tags Then, a tag splicing mechanism is used to generate composite tags. Or string mapping functions This is to preserve the associated information of concurrent failures; Composite label As shown below: ；（2） The string mapping function is shown below: ；（3） In the formula, For transient atomic tags.

5. The semantically aligned data generation method for hydrogen compressor fault diagnosis according to claim 1, characterized in that, In step 3), the latent compressor faults include compressor idling, thermocouple failure, interstage pressurization failure, and false charging. When performing inter-level cascading mutual verification, if If so, it is determined that the interstage boost has failed; Minimum compression ratio; , The pressure is divided into two stages: exhaust pressure and exhaust pressure. When performing differential pressure flow direction verification, if there exists a slope that satisfies... However, data that does not meet the physical constraints of a valid filling condition is diagnosed as false inflation and an anomaly label is generated. ; The data diagnosed as counterfeit inflation parts are shown below: ；（4） In the formula, To measure the secondary exhaust pressure; The target hydrogen storage cylinder pressure.

6. The semantically aligned data generation method for hydrogen compressor fault diagnosis according to claim 5, characterized in that, Step 3) includes the following steps for performing pressure-temperature coupling verification: Step 3.1) Define the device operating status determination function. Pressure increment With temperature increment ; Among them, the equipment operating status determination function As shown below: ；（5） In the formula, Indicates that the program is running; , For vibration signals and thresholds; Step 3.2) Determine the device operating status using the device operation status function. When the equipment is detected to be running, a compressor idling determination is performed. If the non-consistency constraint is met, the system indicates either compressor idling or thermocouple failure, and a fault tag is generated. ; The non-consistency constraints are as follows: ；（6） in For pressure increment, For temperature increment, This is the temperature increment threshold; This is the pressure increment threshold.

7. The semantically aligned data generation method for hydrogen compressor fault diagnosis according to claim 1, characterized in that, Step 4) involves performing third-layer adaptive working condition clustering on the unlabeled data, including: An adaptive clustering algorithm based on Euclidean distance is used to partition the unlabeled sample set. The operating mode is determined, and the probability distribution of a sample belonging to a certain operating mode is calculated. ; Among them, the unlabeled sample set is divided. When operating under certain modalities, the sum of squared errors of intra-cluster temporal characteristics is minimized. For the goal; Squared error As shown below: ；（7） in, For the first Statistical feature vectors of time windows, For the first The center vector of each working condition cluster; The probability distribution of a sample belonging to a certain working condition As shown below: ；（8） In the formula, For the first The center vector of each working condition cluster.

8. The semantically aligned data generation method for hydrogen compressor fault diagnosis according to claim 1, characterized in that, semantic mapping function With question-answering generation functions As shown below: ；（9） ；（10） in For fault labels, For operating condition labels, For expert knowledge base; The features of input X.

9. A semantically aligned data generation method for hydrogen compressor fault diagnosis according to claim 1, characterized in that, In step 6), the preprocessing step refers to reconstructing the input sequence, and the number of time-series tokens after reconstruction. L, P, and s represent the length of the input sequence, the length of the patch, and the step size, respectively.

10. A semantically aligned data generation method for hydrogen compressor fault diagnosis according to claim 1, characterized in that, In step 7), after identifying the fault type corresponding to the fault type in the compressor timing instruction dataset, the pre-trained large language model is supervisedly fine-tuned using the compressor timing instruction dataset. The general timing knowledge of the pre-trained large language model is combined to infer the evolution trend of the compressor in the future time window and predict the evolution trend of the compressor in the future time window. Based on the diagnostic and prediction results, retrieve the contingency plans from the knowledge base and generate operation and maintenance suggestions.

Images