Fault early warning method and system based on AI large model

Abstract

The invention discloses a fault early warning method and system based on an AI large model, and the method comprises the steps: obtaining multi-source heterogeneous data, and carrying out the denoising and standardization processing, and obtaining fusion data; inputting into a feature extraction model, and outputting a feature vector set; identifying the dynamic operation mode based on a K-means algorithm to obtain an operation mode baseline; inputting a feature sequence model, and outputting a precursor feature sequence; calculating an abnormal score according to the precursor feature sequence, marking as abnormal if the score is greater than or equal to a threshold value, otherwise, marking as normal, and obtaining an abnormal detection result; evaluating a risk level according to a detection result; inputting the risk level into a fault analysis model to obtain fault cause distribution; determining optimized operation mode parameters according to the fault cause distribution; and performing deviation analysis on the data and the optimized parameters, and if a deviation value is greater than a threshold value, triggering an early warning signal. The method can solve the problem of insufficient recognition capability in a scene with variable fault types.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent fault warning, and particularly to a fault warning method and system based on an AI large model. Background Art

[0002] With the continuous increase in the complexity of key basic systems such as industrial equipment, energy systems, and transportation facilities, the requirements for operation safety and fault response capabilities have increased significantly. Traditional manual inspections or static rule triggers can no longer meet the real-time, intelligent, and predictive needs of modern systems. Therefore, how to achieve intelligent perception of system operation status, hidden danger identification, and fault warning has become one of the key technical challenges in the current fields of intelligent operation and maintenance (AIOps), industrial Internet, and smart energy. Especially in the context of the wide deployment of multi-source heterogeneous sensors, constructing a high-precision and interpretable fault detection model based on big data and artificial intelligence methods has become an important development direction for fault warning technology.

[0003] In the prior art, a relatively mature implementation path is to construct a fault prediction model based on traditional time series analysis methods. For example, some research uses a sliding window algorithm to segment the vibration signals and temperature signals of electrical equipment, and extracts statistical features such as mean, variance, and slope to form a feature vector of a fixed length. Subsequently, a support vector machine (SVM) model is used to classify and model these feature vectors to determine whether there are abnormal patterns. In this process, the core of the technical implementation is to structure continuous signals into training samples through a sliding window and complete feature extraction with the help of statistics, thereby transforming the signal fluctuation problem into a supervised learning problem. The technical problem of this method is that the feature extraction ability directly affects the classification accuracy, and it is necessary to control sample redundancy while ensuring the coverage of training samples, resulting in the performance of the model being highly dependent on parameter settings and manual experience.

[0004] Existing methods mostly rely on manual rules to construct statistical features and use a fixed classification model as a fault identification tool. It is difficult to accurately extract potential precursor features and maintain recognition robustness when facing multi-dimensional heterogeneous time series. Since these methods mainly focus on static window interception and shallow feature analysis and lack in-depth modeling of the sequence evolution structure and feature coupling relationship, they are prone to recognition deviation when facing scenarios such as system operation state fluctuations and variable fault types, thus limiting the application breadth and generalization ability of the model in early fault detection and complex anomaly identification. In summary, the prior art has the problem of insufficient recognition ability in scenarios with variable fault types. Summary of the Invention

[0005] The present invention provides a fault warning method and system based on an AI large model to solve the problem of insufficient recognition ability in scenarios with variable fault types.

[0006] In a first aspect, to solve the above technical problems, the present invention provides a fault warning method based on an AI large model, including:

[0007] Obtain multi-source heterogeneous data; wherein, the multi-source heterogeneous data includes: real-time temperature, real-time vibration, and real-time current signals;

[0008] According to the multi-source heterogeneous data, perform data denoising and standardization processing to obtain fused data;

[0009] Input the fused data into a pre-trained feature extraction model to output a set of feature vectors;

[0010] Based on the K-means algorithm, perform dynamic operation recognition on the set of feature vectors to obtain a dynamic operation mode baseline;

[0011] Input the dynamic operation mode baseline into a pre-trained feature sequence model to output a precursor feature sequence;

[0012] According to the precursor feature sequence, calculate a feature anomaly score to obtain an anomaly score;

[0013] If the anomaly score is greater than or equal to a preset threshold, mark it as an abnormal feature; if the anomaly score is less than the preset threshold, mark it as a normal feature, thereby obtaining an anomaly detection result;

[0014] According to the anomaly detection result, perform risk assessment to obtain a risk level;

[0015] Input the risk level into a pre-trained fault analysis model to obtain a fault cause distribution;

[0016] According to the fault cause distribution, determine a set of operation regulation parameters that match it to obtain optimized operation mode parameters;

[0017] Perform deviation analysis on the multi-source heterogeneous data and the optimized operation mode parameters to obtain a deviation value. If the deviation value is greater than or equal to the preset threshold, obtain a warning signal.

[0018] In a second aspect, the present invention provides a fault warning system based on an AI large model, including:

[0019] A data acquisition module for acquiring multi-source heterogeneous data; wherein, the multi-source heterogeneous data includes: real-time temperature, real-time vibration, and real-time current signals;

[0020] A fused data module for performing data denoising and standardization processing on the multi-source heterogeneous data to obtain fused data;

[0021] A feature vector module, configured to input the fusion data into a pre-trained feature extraction model and output a set of feature vectors;

[0022] A K-means algorithm module, configured to perform dynamic operation recognition on the set of feature vectors based on the K-means algorithm to obtain a dynamic operation mode baseline;

[0023] A precursor feature sequence module, configured to input the dynamic operation mode baseline into a pre-trained feature sequence model and output a precursor feature sequence;

[0024] An anomaly score module, configured to calculate a feature anomaly score based on the precursor feature sequence to obtain an anomaly score;

[0025] An anomaly detection result module, configured to mark as an abnormal feature if the anomaly score is greater than or equal to a preset threshold, and mark as a normal feature if the anomaly score is less than the preset threshold, thereby obtaining an anomaly detection result;

[0026] A risk level module, configured to perform risk assessment based on the anomaly detection result to obtain a risk level;

[0027] A fault cause module, configured to input the risk level into a pre-trained fault analysis model to obtain a fault cause distribution;

[0028] An operation model parameter module, configured to determine a set of operation regulation parameters matching the fault cause distribution to obtain optimized operation mode parameters;

[0029] A warning signal module, configured to perform deviation analysis on the multi-source heterogeneous data and the optimized operation mode parameters to obtain a deviation value, and obtain a warning signal if the deviation value is greater than or equal to a preset threshold.

[0030] In a third aspect, the present invention further provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, the fault warning method based on the AI large model described in any one of the above is implemented.

[0031] In a fourth aspect, the present invention further provides a computer-readable storage medium, where the computer-readable storage medium includes a stored computer program. When the computer program runs, it controls the device where the computer-readable storage medium is located to execute the fault warning method based on the AI large model described in any one of the above.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] (1) The present invention adopts a feature extraction model trained based on a CNN neural network, which can automatically extract high-dimensional fusion features from multi-source heterogeneous data. Compared with the traditional artificial rule construction method, it has stronger adaptability and feature expression ability, effectively improving the comprehensiveness and accuracy of feature acquisition.

[0034] (2) Through a dynamic operation mode recognition method based on the K-means algorithm, the present invention can cluster the device operation features into multiple state clusters and calculate the central trajectory to form an operation baseline, significantly enhancing the modeling ability of the operation state evolution trend and helping to capture and identify subsequent fault precursors.

[0035] (3) The present invention adopts a feature sequence model constructed by an LSTM neural network, which can model the cross-period feature evolution trajectory in the time series, extract precursor signals and predict the trend state, effectively making up for the deficiencies of the existing static window method in dynamic trend recognition and improving the foresight and robustness of early warning.

[0036] (4) The present invention constructs a fault analysis model based on the Transformer structure, combines the attention mechanism to achieve deep correlation extraction of multi-factor risk levels and fault causes, and can clearly reflect the contribution ratio of different causes to the current operation risk, providing an interpretable basis for subsequent parameter adjustment and control strategy formulation.

[0037] In summary, the present invention constructs a set of fault warning methods and systems for complex working conditions by integrating various artificial intelligence technologies such as neural network modeling, clustering recognition, attention mechanism reasoning, and feature mapping adaptation. It has significant advantages such as high precursor recognition accuracy, strong risk determination robustness, and high parameter regulation adaptability, and can significantly improve the intelligent operation and maintenance efficiency and system security level in complex scenarios. The method can solve the problem of insufficient recognition ability in scenarios with variable fault types. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a schematic flowchart of a fault warning method based on an AI large model provided by the first embodiment of the present invention;

[0039] Figure 2 is a schematic structural diagram of a fault warning system based on an AI large model provided by the second embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0040] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0041] Refer to Figure 1 , the first embodiment of the present invention provides a fault warning method based on an AI large model, including the following steps:

[0042] S11, Obtain multi-source heterogeneous data; wherein, the multi-source heterogeneous data includes: real-time temperature, real-time vibration and real-time current signals;

[0043] S12, According to the multi-source heterogeneous data, perform data denoising and normalization processing to obtain fusion data;

[0044] S13, Input the fusion data into a pre-trained feature extraction model, and output a set of feature vectors;

[0045] S14, Based on the K-means algorithm, perform dynamic operation recognition on the set of feature vectors to obtain a dynamic operation mode baseline;

[0046] S15, Input the dynamic operation mode baseline into a pre-trained feature sequence model, and output a precursor feature sequence;

[0047] S16, Calculate the feature anomaly score according to the precursor feature sequence to obtain an anomaly score,

[0048] S17, If the anomaly score is greater than or equal to a preset threshold, mark it as an abnormal feature; if the anomaly score is less than the preset threshold, mark it as a normal feature, thereby obtaining an anomaly detection result;

[0049] S18, Perform risk assessment according to the anomaly detection result to obtain a risk level;

[0050] S19, Input the risk level into a pre-trained fault analysis model to obtain a fault cause distribution;

[0051] S20, Determine a set of operation regulation parameters that match the fault cause distribution to obtain optimized operation mode parameters;

[0052] S21, Perform deviation analysis on the multi-source heterogeneous data and the optimized operation mode parameters to obtain a deviation value. If the deviation value is greater than or equal to a preset threshold, obtain a warning signal.

[0053] In step S11, obtain multi-source heterogeneous data; wherein, the multi-source heterogeneous data includes: real-time temperature, real-time vibration and real-time current signals;

[0054] It should be noted that in step S11, the core purpose of obtaining multi-source heterogeneous data is to provide basic support for subsequent feature extraction and model training. Multi-source heterogeneous data refers to the operating state data from different sensors with different physical meanings and sampling methods. Specifically, in the industrial field environment targeted by the present invention, the data sources mainly include three types: temperature sensors, vibration sensors, and current sensors, and each type of data has different sampling characteristics and processing requirements.

[0055] The real-time temperature data is collected by the thermistor sensors arranged on the surface or inside of the key equipment. The acquisition system is set to read the temperature at a frequency of once per second, forming a temperature signal in the form of a time series. Considering that the temperature rise of industrial equipment is slow but closely related to the fault trend, the acquisition frequency setting needs to balance data change capture and system resource occupancy. The temperature data is accessed to the edge acquisition gateway through the wired transmission module, and the preliminary timestamp marking and format unification are completed at the gateway end, forming a standardized temperature data stream.

[0056] The real-time vibration data is obtained relying on the acceleration sensors installed on the equipment bearings, motor casings and other parts. The acceleration sensor is set to the high-frequency sampling mode, and the sampling frequency is more than 100 times per second to ensure the complete recording of the tiny fluctuations during the high-speed mechanical rotation process. The vibration signal is output in the form of a three-axis vector, encoded as a digital signal through the analog-to-digital conversion circuit and then input to the acquisition terminal.

[0057] The real-time current signal is obtained by the Hall current sensor or current transformer installed at the main power input end of the equipment, and the sampling frequency is set to 10 times per second. The acquisition content is the instantaneous current intensity at the moment of equipment operation, with the unit of ampere. Since the current fluctuation is closely related to the equipment load change, the current data can effectively reflect problems such as load abnormality and overcurrent in fault identification. After the current signal is integrated with the data acquisition unit through the dedicated power monitoring module, it is uploaded to the data processing system in a unified format.

[0058] In order to ensure the timing consistency and correspondence of the data, a unified time synchronization mechanism is adopted during the data acquisition process. All sensors are uniformly controlled by the main acquisition terminal for the sampling moment, and a unified timestamp is recorded. The collected temperature, vibration, and current data of the three types are structurally encapsulated through the data aggregation module at the edge end, forming a multi-dimensional time series data structure, and uploaded to the central processing server via the industrial field bus or industrial wireless communication protocol. The finally formed multi-source heterogeneous data set not only retains the original change information of each physical quantity but also has good timing alignment characteristics, providing real, continuous, and available data support for subsequent feature fusion and model input.

[0059] For example, in an example of monitoring the operation of a certain motor system, the system collects temperature data once per second, current data ten times per second, and vibration data one hundred times per second. The collection results show that after continuous operation for 30 minutes, the temperature rises from 45°C to 58°C, the peak value in the X-axis direction of vibration increases from 0.4g to 0.9g, and the current fluctuates from 15A to 18A.

[0060] In step S12, based on the multi-source heterogeneous data, data denoising and normalization processing are performed to obtain fused data.

[0061] It should be noted that in step S12, performing data denoising and normalization processing on the collected multi-source heterogeneous data is a key link to ensure the input quality of the subsequent feature extraction model. Since there are significant differences in the sampling methods, physical units, and noise interference levels of the temperature, vibration, and current sensor signals, it is necessary to systematically clean and transform the original data to ensure that the finally formed data has a unified structure, comparability, and stability.

[0062] First, data denoising processing is carried out. For the temperature signal, the moving average algorithm is used to smooth the data. The specific method is to set a fixed window, for example, five sampling points, calculate the average value within this window to replace the current point, and eliminate the small fluctuations caused by equipment jitter or signal jitter. In the processing of current data, the median filtering method is used to suppress the spike interference in the power sensor signal. Median filtering can effectively remove the abnormal mutation values caused by instantaneous electromagnetic interference and retain the true trend of current fluctuations. For the vibration signal, a band-pass filter is used to process the signal, retaining the signal components within the main vibration frequency band and removing the low-frequency drift and high-frequency background noise. The upper and lower limit frequencies of the filter are set according to the actual frequency of equipment operation. For example, if the motor speed is 1500 revolutions per minute, the band-pass range is set from 10 Hz to 100 Hz.

[0063] After denoising is completed, the normalization processing stage is entered. Since the temperature is in the dimension of degrees Celsius, the vibration is in acceleration units such as g or meters per square second, and the current is in amperes, the existence of different dimensions and value ranges will cause problems of dimensional imbalance in the subsequent calculation of the neural network model. To solve this problem, this step uses the Z-score normalization method to process all signals. The specific method is to first calculate the historical average value and standard deviation of each type of data, and then subtract the mean value from each data point and divide by the standard deviation to obtain the normalized data with a mean value of zero and a variance of one. After such processing, the numerical values of different physical quantities are mapped to a unified scale space, enhancing the model's ability to fuse and process different signals.

[0064] In addition, considering that there may be sampling time alignment deviations in various sensors, all signals are re-synchronized in time during the processing. Using the temperature sampling frequency as the main clock, the vibration and current data are interpolated or downsampled according to timestamps to ensure that the data of different sensors are aligned at the same time node, facilitating the subsequent construction of a time series tensor.

[0065] Taking a specific example, during a certain operation, the temperature data collected showed periodic noise jitter, the vibration signal had high-frequency spikes under external mechanical interference, and the current signal contained multiple short-term abnormal points. Through the above denoising process, the fluctuation amplitude of the temperature signal was compressed within a stable range, the abnormal spikes of the vibration signal were effectively weakened, and the mutation points in the current signal were replaced with the median of adjacent values. After normalization, these three types of signals were respectively converted into floating-point number sequences with zero mean and unit variance, and were uniformly synchronized to one time point per second, jointly constituting a fusion data tensor for subsequent model input.

[0066] In step S13, the fusion data is input into a pre-trained feature extraction model, and a set of feature vectors is output, including:

[0067] The feature extraction model is obtained by training a CNN neural network model;

[0068] Through the input layer of the feature extraction model, the dimensions of the fusion data are matched to obtain an input tensor;

[0069] Through the hidden layer of the feature extraction model, multi-scale convolution operations and channel information extraction are performed on the input tensor to obtain an intermediate feature map;

[0070] Through the output layer of the feature extraction model, full connection mapping and dimensionality reduction processing are performed on the intermediate feature map to obtain a set of feature vectors.

[0071] It is worth noting that in step S13, inputting the fusion data after denoising and normalization into a pre-trained feature extraction model and outputting a set of feature vectors is a key information extraction link before fault recognition in the entire system. This step relies on the structural advantages of the deep convolutional neural network model, which can automatically learn discriminative feature representations from high-dimensional, multi-source, and heterogeneous original signals, avoiding the subjectivity and limitations of manually constructing features in traditional methods. The feature extraction model adopted in the present invention is constructed based on the CNN convolutional neural network and is pre-trained with a large amount of industrial scenario data, making it have stronger feature recognition ability and transfer adaptability when facing actual complex working conditions.

[0072] The model training process is completed based on a preset training dataset. The dataset is sourced from a large number of temperature, vibration, and current signals collected in real industrial scenarios, including normal operation data and labeled fault status data. The training data is first subjected to denoising and normalization processes consistent with the inference stage, and then the data for continuous time periods is sliced into fixed-length data segments by sliding a time window. Each segment forms an input sample. Each sample corresponds to a label indicating the operating state category of the device during that time period. During the model training process, the cross-entropy loss function is used to compare the model output with the labels, and the convolutional kernel parameters are continuously adjusted through the backpropagation algorithm, enabling the model to gradually learn the most relevant feature patterns for state classification from the fused data.

[0073] The structure of the feature extraction model includes an input layer, multiple convolutional hidden layers, and an output layer, forming an end-to-end deep network structure as a whole. In the input layer part, the system receives a structured fused data tensor. The dimension of this tensor is set as a matrix structure of the time step length multiplied by the number of channels, where the number of channels is equal to the three types of physical signals, namely temperature, vibration, and current. The model first performs dimension matching and normalization confirmation on the input tensor in the input layer to ensure that subsequent convolutional calculations meet the model's preset requirements both in terms of numerical scale and structural shape. This part of the work is automatically completed by the Tensor format manager for dimension verification, missing value filling, and data type conversion operations to ensure the integrity and processability of the tensor input.

[0074] After entering the convolutional layer, the input tensor is fed into multiple stacked convolutional modules for feature extraction. The convolutional modules adopt a multi-scale convolutional kernel structure, with three groups of convolutional kernels set respectively, whose receptive fields cover 3, 5, and 7 time step lengths, capable of capturing feature change patterns at different scales. Taking the layer with a convolutional kernel size of 3 as an example, its core objective is to identify the subtle change trends within a short time window, such as the slope structure formed by the continuous increase of temperature at 3 sampling points, or the local segment of sudden high-frequency fluctuations in the vibration signal. After each group of convolutional operations, a batch normalization layer and a ReLU activation function follow immediately to ensure stable feature distribution and smooth model gradient flow. The output of this level forms the first intermediate feature map of the model, used to capture the feature evolution trends at different time scales.

[0075] To further enhance the collaborative perception ability between channels, the model introduces a cross-channel convolution strategy in the second-layer convolutional structure. By fusing the features from the temperature, vibration, and current channels layer by layer, a multi-channel representation with cross-physical quantity feature coupling is constructed. This design enables the model to not only independently analyze the change trends of each signal channel but also capture the interaction relationships between them. For example, when temperature changes and current fluctuations occur simultaneously, cross-channel convolution can identify the potential fault causal relationships behind this synchronous variation. This part of the structure consists of multiple one-dimensional convolutional kernels, and the weight sharing mechanism ensures that the model has a unified discrimination criterion in different time windows, improving the stability of feature extraction.

[0076] In the third-layer convolutional structure, the model introduces a pooling layer for feature compression. The pooling operation uses the max-pooling method, and the maximum value is extracted every two time steps. The main function of this operation is to sample the features extracted from the previous layer, reducing the computational complexity while retaining the most representative feature response positions. In addition, max-pooling can also enhance the robustness of the model to local perturbations, enabling it to still extract stable fault symptom patterns when facing small perturbations in vibration signals or increasing volatility in temperature signals. The output of this layer constitutes the final intermediate feature map, whose size has been significantly compressed from the original input but the information density has been significantly enhanced.

[0077] As the feature expression result in the high-dimensional space, the intermediate feature map needs to be converted into a fixed-dimensional vector that can be used by the clustering analysis or classification model through the output layer. In the output layer part, the model uses a fully connected layer for feature dimensionality reduction and mapping. This process first flattens the intermediate feature map into a one-dimensional vector according to the channel and time dimensions, and then inputs it into a fully connected layer with 128 neurons for linear combination and weight transformation, and the output dimension is set to 64. This dimension setting is determined through experimental comparison, avoiding overfitting and wasting computational resources while maintaining sufficient feature expression ability. Finally, the output layer uses the Sigmoid function for normalization mapping to make the feature values fall within a fixed interval for subsequent direct processing by the clustering model and distance metric module.

[0078] Specifically, in an industrial site, when the motor is in a low-load high-frequency start-stop mode, the model extracts the local mutation patterns with frequent fluctuations in the vibration signal through the first convolutional layer, and at the same time identifies the impact characteristics of the current at the moment of startup in the second convolutional layer, and combines these two types of features to form a joint expression of abnormal start-stop behavior. After the pooling layer, these mutation points are condensed into key response peaks and mapped into feature vectors deviating from the normal working mode in the fully connected output. This feature vector is finally input into the clustering analysis module for distance comparison and classification with other operating modes to determine whether the current state belongs to a known fault mode or a potential risk state.

[0079] In summary, in step S13, through the structure standardization of the input layer, multi-scale feature extraction of the convolutional layer, cross-channel fusion and pooling compression, the dimensionality reduction and mapping of high-dimensional features are finally achieved in the output layer. This not only realizes the goal of vectorized representation of the fused data, but also provides a solid foundation for the accuracy of fault detection, the model generalization ability and the stable operation of the subsequent processing module.

[0080] In step S14, based on the K-means algorithm, dynamic operation recognition is performed on the feature vector set to obtain a dynamic operation mode baseline, including:

[0081] According to the feature vector set, clustering initialization processing is performed to obtain an initial clustering structure;

[0082] Based on the K-means algorithm, Euclidean distance measurement and intra-cluster minimization iterative operation are performed on the initial clustering structure to obtain multiple dynamic operation cluster results;

[0083] According to the multiple dynamic operation cluster results, central trajectory calculation is performed, and the feature distribution law is analyzed to obtain a dynamic operation mode baseline.

[0084] It should be noted that in step S14, the system performs dynamic operation recognition on the feature vector set output in the previous step based on the K-means algorithm, so as to obtain a dynamic operation mode baseline representing the current device operation state. The core purpose of this process is to divide the data points in the high-dimensional feature space into several operation mode clusters, so that the samples in the same class have similar operation behavior characteristics, while there are obvious differences between different classes. Through this process, the state distribution structure of the system operation can be effectively identified, and the central features of various operation states can be extracted, thus providing a basis for subsequent anomaly detection and fault judgment.

[0085] First, clustering initialization processing is performed according to the feature vector set. This feature vector set is output by the CNN feature extraction model and has unified dimensionality and normalization characteristics. In the clustering initialization stage, the number of clusters K is set, and K points are randomly selected in the feature space as the initial clustering centers to construct an initial clustering structure. Each center in this structure represents a potential operation state category, and its position determines the division boundary of the initial classification. During the initialization process, it is necessary to ensure the coverage of the sample distribution to avoid all clustering centers being concentrated in a certain area of the feature space, resulting in classification bias. In practice, the K-means++ strategy is used to optimize the initialization of the initial clustering centers, that is, the data points with the farthest distance are preferentially selected as the new clustering centers in the feature space to increase the diversity of the initial seeds.

[0086] After initialization, a standard clustering iteration process is performed based on the K-means algorithm. Specifically, for all samples in the feature vector set, the Euclidean distance between each sample and each cluster center is calculated, and they are assigned to the corresponding clusters according to the principle of the smallest distance, completing the first division of sample attribution. Subsequently, the mean positions of all samples in each cluster are statistically calculated to update the cluster centers. The process of sample division and center update is repeated until the offset of all cluster centers in two consecutive iterations is less than the set convergence threshold. The convergence result of this algorithm is multiple dynamically running clusters, and each cluster represents a typical device operation state mode. The samples within the cluster are highly aggregated in the feature expression space, indicating that their operating behaviors are similar.

[0087] To further extract representative operating baseline features from the clustering results, the system performs center trajectory calculation on all cluster results. This operation is based on the time-series evolution process of each cluster center, statistically calculates the average position and standard deviation of each cluster at different time periods to generate a baseline curve describing its state change trend. This curve not only reflects the static center position of this operating mode in the feature space but also captures the characteristic trend of its fluctuations on the time axis. For example, in a monitoring scenario of an industrial air compressor, the operating state of the device oscillates periodically with the change of load. For the "medium load" cluster obtained by clustering the feature vectors, its center trajectory shows a changing trend of stable temperature, gradually increasing current, and converging vibration fluctuations at different time periods, reflecting the mode of the device maintaining stable operation during the process of gradually increasing pressure.

[0088] Meanwhile, the system performs distribution law analysis on each cluster in the clustering results. This analysis is based on the dispersion of features within each cluster in each dimension to identify the fluctuation characteristics exhibited by this operating mode in different physical signal dimensions. If the standard deviation of the vibration dimension in a certain type of cluster is large, while the fluctuations in the temperature and current dimensions are small, it can be inferred that this type of operating state has the characteristic of "strong vibration interference". Such analysis results can not only be used for the semantic naming of the operating state but also provide a priori basis for subsequent abnormal boundary modeling and alarm strategy generation.

[0089] Finally, the system combines the center trajectories of multiple operating clusters with their distribution characteristics to form a dynamic operating mode baseline. The dynamic operating mode baseline comprehensively expresses the normal operating mode of the device and its change trend based on the center trajectories and feature distribution laws of various states.

[0090] In summary, under the support of the K-means algorithm in step S14, the feature vector set is used to complete the process of identifying dynamic operating states and extracting baselines. It not only realizes the identification of typical state modes from high-dimensional data but also constructs a baseline model with time and space structures, providing a stable and reliable operating reference for fault warning.

[0091] In step S15, the dynamic operation mode baseline is input into a pre-trained feature sequence model, and a precursor feature sequence is output, including:

[0092] The feature sequence model is obtained by training an LSTM neural network model;

[0093] Through the input layer of the feature sequence model, the dynamic operation mode baseline is unfolded in time steps and restructured in structure, and the feature evolution trajectory across time segments is extracted to obtain input units;

[0094] Through the memory layer of the feature sequence model, the input units are modeled with a phased dynamic state to obtain a trend recognition coding vector;

[0095] Through the mapping layer of the feature sequence model, the trend recognition coding vector is transformed into a precursor signal to obtain a precursor feature sequence.

[0096] It should be noted that in this step, the system performs a time series modeling process on the dynamic operation mode baseline based on the feature sequence model, so as to extract the precursor feature sequence. The feature sequence model uses a long short-term memory network (LSTM) as the core architecture, and its advantage lies in its ability to remember states across time periods and can effectively capture the dynamic evolution features at multiple time steps. To achieve the effective application of the model in a specific industrial scenario, systematic design is required in aspects such as data preparation, model construction, training process, and inference structure, to ensure that the model output result can truly reflect the trend change of the operation state and has good generalization ability.

[0097] First, in the model training stage, a representative time series training dataset needs to be constructed. This dataset is sourced from historical multi-source operation data under multiple working conditions. After fusion processing and state baseline extraction, the fusion processing integrates multiple data sources (such as temperature, vibration, current, etc.). First, denoising processing is performed to ensure data quality, and then standardization is carried out to make data with different dimensions have the same scale, ensuring consistency and accuracy. State baseline extraction is based on the fused data. Through methods such as the K-means clustering algorithm, the characteristics under normal working conditions are extracted to form a baseline, representing the normal operating state of the equipment and providing a reference for subsequent anomaly detection. This baseline reflects the normal behavior pattern of the equipment by extracting statistical features and establishing a feature space, thereby generating a series of time series samples with consistent lengths. Each group of samples consists of dynamic baseline vectors at several consecutive time steps and is accompanied by known labels. The known labels refer to when constructing the training dataset, a label is pre-assigned to each group of sample data. The content of the label can indicate whether there is a fault in the data or whether it belongs to the normal state. These labels are manually annotated. During the training process, the known labels serve as supervision information to help the model learn how to distinguish between the normal state and the abnormal state, thereby improving the accuracy and generalization ability of the model. The known labels are used to indicate whether there is a potential abnormal evolution trend in the sequence. To ensure the diversity and representativeness of the training set, the training data needs to cover normal state sequences, slightly fluctuating sequences, early fault evolution sequences, and mutating abnormal sequences. The time spans of various types of samples are controlled within the same interval. For example, each group of sequences uniformly contains 30 time steps, with a step interval of 1 minute, covering a total of 30 minutes of operation change process. On this basis, the model aims to minimize the prediction error and optimizes the parameters through a supervised training method, ultimately obtaining a set of LSTM network parameter sets that can generalize and identify multiple trend states.

[0098] In terms of the model structure, the feature sequence model consists of three main levels: the input layer, the memory layer, and the mapping layer. First, the input layer is responsible for receiving the time series data from the dynamic operation mode baseline and performing time step expansion and structural reorganization processing. Each dynamic baseline sample consists of operation state characteristics at multiple time points. For example, in the industrial motor scenario, each time point contains three-dimensional parameters such as temperature, current, and vibration, forming a three-dimensional vector. The system arranges these vectors in chronological order to form a time × feature input tensor, which is the input unit of the LSTM model. To make the dimensions of different batches of samples consistent, the system performs unified zero-mean normalization processing on all input samples and, if necessary, performs padding operations in the time dimension to ensure that the LSTM can perform batch training.

[0099] Next, the input unit enters the memory layer of the feature sequence model, which consists of one or more stacked LSTM units. Each LSTM unit sequentially receives the input vector at the current time step in the time dimension, and combines the hidden state and memory cell state passed from the previous time step to update the current output state. This mechanism enables the model to refer to the information of historical time steps in each step of prediction, realizing the memory modeling of the stage dynamic evolution process. Through the design of multi-layer stacking, the model can capture short-term trends from the bottom layer, identify mid-term transition states in the middle layer, and finally extract long-term trend patterns in the upper layer, thus realizing the hierarchical deconstruction and expression of complex evolution structures. During the training process, the model updates the weight matrices of all LSTM units through the backpropagation algorithm, gradually enhancing its sensitivity to precursor evolution features.

[0100] In practical applications, the output vector of the memory layer is regarded as the trend recognition coding vector, representing the comprehensive judgment result of the model on the evolution state of the current input sequence. For example, during the operation of a motor, if the temperature shows a linear increase, the current fluctuation intensifies, and the vibration has no obvious change in 15 consecutive time steps, the model will judge through the memory mechanism that this trend may point to the "abnormal load evolution" situation, and reflect the weight distribution of this trend signal in the output coding.

[0101] Finally, the trend recognition coding vector is passed into the mapping layer of the feature sequence model to complete the transformation process of the precursor signal. This layer consists of a fully connected neural network, whose function is to map the high-dimensional state coding output by the LSTM into a feature space with stronger semantic representation ability, and strengthen or suppress the response intensity of different dimensions through activation functions. During the mapping process, the system pre-defines a set of structural templates for precursor signals, such as the fluctuating type, continuously increasing type, mutation type, etc., and measures the similarity between the mapping result and these templates, so as to output the closest precursor type. Finally, the system constructs the precursor feature sequence according to the output precursor signal expression vector and uses it for subsequent anomaly detection and fault judgment modules.

[0102] Taking practical applications as an example, in a certain intelligent motor fault warning system, the system detected that the current value slowly rose from 5.2 amperes to 6.8 amperes within 30 minutes, while the temperature rose from 58 degrees to 65 degrees, and the vibration value remained within a stable range during the early operation stage. The system input this time series into the LSTM model for trend modeling. The model identified that this trend conforms to the "linear power increase" mode. Combining the historical sample matching results, the system judged that the current device is in a state of mild load deviation evolution. The output precursor feature sequence is marked as a medium-risk trend, and it is prompted that the monitoring frequency of power load regulation should be enhanced subsequently.

[0103] In summary, the feature sequence model is constructed based on the LSTM neural network structure, and uses its time memory ability to model the evolution trend of multi-dimensional time series data in the dynamic operation mode baseline. The training process of the model is constructed based on historical fault samples, and the structure consists of an input layer, a memory layer, and a mapping layer. Each layer cooperates to complete the transformation from the original operation baseline to the precursor feature sequence, and has good state recognition accuracy, trend judgment accuracy, and engineering feasibility.

[0104] In step S16, according to the precursor feature sequence, calculate the feature anomaly score to obtain the anomaly score, including:

[0105] According to the precursor feature sequence, perform outlier detection preprocessing to obtain an initial set of abnormal candidate segments;

[0106] Based on the Isolation Forest algorithm, calculate the anomaly score for the initial set of abnormal candidate segments to obtain the anomaly score.

[0107] It should be noted that in step S16, the Isolation Forest algorithm is an unsupervised learning algorithm based on a tree structure, which is particularly suitable for outlier detection of high-dimensional data. The core idea of this algorithm is to isolate data points to determine whether they are outlier data points. For each input sample, the Isolation Forest constructs a random tree to divide the sample into smaller and smaller subsets. Outliers are easier to isolate than normal points, so they have shorter paths in the tree.

[0108] First, the system performs outlier detection preprocessing according to the precursor feature sequence obtained in the previous step. The purpose of preprocessing is to remove noise and ensure the data quality input to the Isolation Forest algorithm. In practical applications, the precursor feature sequence consists of multiple sensor data, such as temperature, current, vibration and other signals. Since these signals may have errors or noise, the preprocessing steps include data denoising and standardization. Data denoising can remove irrelevant fluctuations through smoothing filtering methods, while standardization adjusts features with different dimensions to the same scale to ensure that each feature contributes equally to the result.

[0109] After preprocessing, the system uses the Isolation Forest algorithm to calculate the anomaly score. The basic principle of the Isolation Forest algorithm is to create a series of random trees and recursively partition the data to isolate the data samples. Each sample generates a score by calculating the path length by which it is isolated in each tree. Samples with lower scores are less likely to be isolated and are therefore more likely to be normal data, while samples with higher scores are easy to isolate and may be outlier data. In the case of the precursor feature sequence, points with higher scores will be marked as abnormal, indicating that the change trends of these data points deviate significantly from the normal pattern and may indicate the precursor of equipment failure.

[0110] Specifically, assume that when monitoring the temperature, current, and vibration data of a monitoring device, the system first generates a precursor feature sequence based on these sensor data. These data contain noise in their original state, so preprocessing is required, including data denoising and normalization. Assume that the temperature recorded by the sensor is [60, 61, 63, 64, 65], the current is [2.5, 2.6, 2.7, 2.8, 2.9], and the vibration is [0.05, 0.06, 0.06, 0.07, 0.08]. To remove sensor noise, the data is first smoothed using a filter to eliminate abnormal fluctuations. Then, the system performs normalization to adjust all data to the same dimension. For example, the temperature data is subtracted by its mean and divided by the standard deviation to obtain the normalized data.

[0111] After completing the data preprocessing, the system inputs this data into the Isolation Forest model for anomaly detection. The Isolation Forest constructs multiple random trees by recursively partitioning the feature vectors. Assume that the system uses 10 trees and sets the maximum depth of each tree to 10. In each tree, the algorithm randomly selects features and split points to continuously partition the data until all data is isolated. Each data point is evaluated as an anomaly point by calculating its path length in the tree (i.e., the distance from the root node to the leaf node). A shorter path indicates that the data point is easily isolated and may be an anomaly point; a longer path indicates that the data point is more difficult to isolate and may be a normal point.

[0112] For example, after applying the Isolation Forest algorithm to the above temperature, current, and vibration data, the system may calculate the score for each data point. For example, for the data point with a temperature of 65 degrees, the calculated score is 0.9, indicating that it may be an anomaly point, while for the data point with a temperature of 61 degrees, the score is 0.3, indicating that it is a normal point.

[0113] In step S17, if the anomaly score is greater than or equal to the preset threshold, it is marked as an abnormal feature; if the anomaly score is less than the preset threshold, it is marked as a normal feature, thus obtaining the anomaly detection result.

[0114] It should be noted that in step S17, based on the anomaly scores output by the Isolation Forest in the previous steps, the system judges the anomaly of each piece of precursor feature data and performs binary marking according to the set numerical threshold to determine the final anomaly detection result. The core of this process lies in distinguishing the feature segments that significantly deviate from the normal pattern by comparing the score results with the preset boundary, so as to effectively identify potential fault signals.

[0115] First, the calculation result of the anomaly score comes from the Isolation Forest processing of the precursor feature sequence in step S16. This score is a floating-point value between 0 and 1, used to characterize the isolation degree of data in the feature space. The higher the score, the easier it is for the data point to be isolated, and the shorter its average path length in the forest, which means it is in the marginal or sparse area of the feature distribution and has a strong possibility of being abnormal. Conversely, data with a lower score often lies inside the cluster or in a frequent area, is more representative, and shows a normal operating state.

[0116] In this embodiment, to effectively distinguish abnormal and normal states under different working conditions, the system pre-sets the anomaly score threshold to 0.65. This value is determined through a large number of offline sample annotation experiments, and its selection basis lies in the balanced analysis of the anomaly detection rate and false alarm rate of the training set and the test set. In the monitoring scenario of an industrial motor system, if the threshold is set too high (such as above 0.8), although the false alarm rate can be reduced, it is easy to miss early abnormal features and reduce the early warning ability of the model. If it is set too low (such as below 0.5), a large number of normal features will be misjudged as abnormal, thus interfering with subsequent judgments and decisions. Therefore, setting the threshold to 0.65 is an empirical value that shows high accuracy and strong robustness in multiple actual measurements.

[0117] When executing this judgment logic, the system makes a separate judgment on the anomaly score of each precursor feature sample. If its score is greater than or equal to 0.65, it is considered that this segment has strong outlier features, and its feature structure is significantly different from the known normal state, then it is marked as an abnormal feature and retained as part of the potential fault signal. Conversely, if the score is less than 0.65, this segment is marked as a normal feature, and the system automatically eliminates the participation weight of this feature in subsequent early warning signal generation, risk level assessment, and fault cause analysis.

[0118] Taking a specific example, during the monitoring of an industrial motor, its precursor feature sequence includes a segment with rising vibration features, a segment with continuously high temperature, and a segment with slight current fluctuations. After calculation by the Isolation Forest, the vibration feature score is 0.71, the temperature feature score is 0.68, and the current feature score is 0.53. In this step, the system identifies the first two as abnormal features and retains them for subsequent risk identification processes, while the current feature is judged as normal state data and eliminated from the anomaly detection path. This process can effectively avoid the judgment interference caused by feature redundancy and focus on the core feature segments that actually show signs of fault evolution.

[0119] At the end of this step, the system integrates all the segments marked as abnormal and records their positions in the time series, score levels, and the information of the sensing channels involved to form complete anomaly detection result data.

[0120] In step S18, based on the abnormal detection result, a risk assessment is carried out to obtain a risk level.

[0121] It should be noted that in step S18, the system constructs a risk assessment model based on the abnormal detection result output in the previous stage S17, and accordingly evaluates the risk level of each feature in the current state. The core objective of this step is to form a set of quantifiable risk judgment logics through multi-dimensional indicators such as the number of abnormal features, the degree of abnormality, and the time distribution, providing a basic basis for subsequent fault cause analysis and early warning decision-making.

[0122] First of all, the risk assessment process takes the abnormal detection result as input, which contains parameter information such as the time position, feature dimension, and abnormal score of all identified abnormal features. The system first counts the number of abnormal features within a unit time window and classifies them into corresponding sensing channels, such as the temperature channel, vibration channel, and current channel. Then, combined with the abnormal score corresponding to each abnormal point, the system calculates the channel-level abnormal density index. This density value represents the concentration degree of high-score abnormal points per unit time under a certain sensing type, reflecting the intensity and frequency of the abnormality.

[0123] Next, the system compares the statistical result with the preset risk level standard. The risk level is divided into three levels, namely low risk, medium risk, and high risk. If the number of abnormal points in a certain channel exceeds 3 within the specified time window and the average abnormal score is greater than 0.75, it is classified as high risk; if the number of abnormal points is between 1 and 3 and the score is between 0.6 and 0.75, it is classified as medium risk; the rest are marked as low risk. This rule is obtained by training a large amount of industrial equipment operation historical data and can effectively cover various types of abnormal trends and precursor distribution states.

[0124] Taking an industrial motor as an example, the system detects 5 abnormal points in the temperature channel within a 10-minute monitoring cycle, with an average abnormal score of 0.78; there are 2 abnormal points in the vibration channel, with an average score of 0.66; and only 1 abnormal point in the current channel, with a score of 0.58. According to the risk level assessment standard, the temperature channel is rated as high risk, the vibration channel is rated as medium risk, and the current channel is rated as low risk. The system constructs a risk level vector under the current working condition based on this and uses it for subsequent fault cause reasoning and control strategy matching.

[0125] Finally, the system compares the risk level result with the historical risk evolution sequence to judge whether the current risk belongs to an upward trend, a fluctuating state, or a short-term concentrated type. If high risk levels appear in multiple consecutive monitoring cycles, the system will weight and enhance the cause analysis weight of the channel features in the subsequent steps.

[0126] In step S19, the risk level is input into a pre-trained fault analysis model to obtain a fault cause distribution, including:

[0127] The fault analysis model is trained by a Transformer neural network structure;

[0128] Through the encoding layer of the fault analysis model, positional encoding processing is performed on the risk level to obtain a risk embedding vector sequence

[0129] Through the self-attention layer of the fault analysis model, high-order correlation features are extracted from the risk embedding vector sequence to obtain an attention-weighted expression result;

[0130] Through the decoding layer of the fault analysis model, linear mapping and activation transformation are performed on the attention-weighted result to obtain a fault cause distribution.

[0131] It should be noted that in step S19, to achieve refined fault cause analysis of the identified risk level, the system takes this risk level as input and introduces it into a pre-trained fault analysis model, and finally outputs the distribution result of the fault cause. This model is constructed using a Transformer neural network structure, has good sequence modeling ability and the ability to mine correlations between multi-dimensional features, and is suitable for the dynamic mapping modeling requirements between risk levels and potential fault factors in complex industrial systems.

[0132] First, the risk level data is a set of level labels generated based on the multi-dimensional state feature anomaly detection results obtained in the previous steps, and each level label corresponds to the overall risk level within a working condition cycle. Before applying the model, it is necessary to vectorize this risk level data. To fully retain the position structure information of the risk level in the sequence, the system inputs this level data into the encoding layer of the Transformer model for positional encoding processing. The positional encoding uses a sine-cosine function embedding method to map the original risk level label into an embedding vector sequence with a sequence structure, forming a risk embedding vector. The role of this step is to provide an accurate time position information basis for the subsequent model to capture the risk fluctuation patterns at multiple moments, and to avoid affecting the analysis accuracy due to the lack of vector order.

[0133] Subsequently, the risk embedding vector sequence is fed into the self-attention layer in the fault analysis model. As the core structure of the Transformer, this layer identifies the high-order correlation relationship between the current risk level and the historical level state by calculating the attention weights between the position vectors in the embedding vector sequence. The self-attention mechanism not only considers the importance of each level point itself but also dynamically quantifies the potential interaction effects between different time steps, thereby constructing a more complete risk evolution trajectory. Specifically, the attention layer calculates the weighted expression results between the risk embedding vector at each moment and all other vectors, forming a weighted attention representation tensor. This tensor preserves the temporal relationship of multi-source context while aggregating information and is an important intermediate expression result for mining fault driving factors.

[0134] Next, the model enters the decoding layer processing stage. The decoding layer structure includes multiple layers of linear mapping layers and activation function layers. Its main task is to map the attention-weighted result obtained in the previous step to the fault factor space. First, the attention representation is converted into an intermediate result with the same dimension as the fault factor through a linear transformation, and then the non-linear transformation is completed through the activation function to enhance the model's fitting ability for complex cause distributions. In the implementation process, the ReLU structure is used as the activation function to ensure that the output space has sufficient expressive power. Finally, the output of the decoding layer is a fault cause distribution vector, where each dimension corresponds to a potential fault factor, and the value in the vector represents the contribution level of this factor at the current risk level.

[0135] Taking an industrial cooling system as an example, when the system shows a high-risk level value of "3" during a certain period, the model embeds this value in vector form and inputs it into the Transformer model for analysis. After modeling the correlation between the historical and current risk levels through the self-attention layer, the model identifies that "coolant temperature rise" and "main pump current fluctuation" are typical accompanying factors under a similar level background. Finally, in the fault cause distribution vector output by the decoding layer, the contribution of abnormal coolant temperature is 0.72, the contribution of main pump current fluctuation is 0.64, and the contributions of other factors are all lower than 0.2. Based on this, the system determines that the main cause of the current risk level is the failure of the coolant control module, and the auxiliary factor is the power supply fluctuation of the pump station.

[0136] Through the processing of this step, not only is the in-depth semantic understanding and context information supplementation of the risk level achieved, but also an interpretable cause distribution result is successfully output, providing an important basis for subsequent mode adjustment and fault suppression strategy formulation.

[0137] In step S20, according to the fault cause distribution, determine the set of operation regulation parameters that match it to obtain the optimized operation mode parameters, including:

[0138] According to the distribution of the fault causes, domain features are extracted to obtain the basis for domain feature mapping;

[0139] According to the basis for domain feature mapping, unified mapping of the feature space is performed to obtain an adapted operating feature structure;

[0140] According to the adapted operating feature structure, the parameter dimensions that need to be key-regulated in the current operating state are determined to obtain optimized operating mode parameters.

[0141] It should be noted that this step involves multiple operation links, including fault cause analysis, feature mapping, and feature space optimization. The system analyzes the fault causes in detail, extracts domain features related to the current operating state, and adapts these features to form an optimized set of operating mode parameters.

[0142] First, based on the fault cause analysis, the system identifies the key factors that may lead to equipment failures. These factors may include temperature fluctuations, current anomalies, vibration frequency changes, etc. during equipment operation. To ensure the accuracy of the fault cause analysis, the system collects real-time data from multiple sensors and compares this data with historical data to infer the current possible operating state of the equipment and whether this state is related to faults.

[0143] Next, based on the obtained distribution of fault causes, the system extracts the operating data through a domain feature extraction algorithm. The domain feature extraction algorithm is an algorithm that extracts features related to a specific domain from the original data. It extracts important features that can effectively characterize the domain characteristics by analyzing the laws, trends, and patterns in the data. In the equipment monitoring and fault warning system, this algorithm is used to identify key features related to equipment operation from multi-source heterogeneous data, help the system accurately understand the current state of the equipment, provide basic data support for subsequent analysis and decision-making, and thus improve the accuracy of fault diagnosis and warning.

[0144] After completing the extraction of domain features, the system performs spatial mapping processing based on the extracted features. The purpose of feature mapping is to map features from different domains into a unified feature space to avoid the influence of differences between different data sources on subsequent analysis. For this purpose, the system uses a domain-adapted feature mapping method, which ensures that all features can be correctly aligned within the same feature space through an optimization algorithm. For example, after features of two different physical quantities, such as temperature and vibration, are mapped into the same space, the system can more easily analyze the correlation between them to identify potential abnormal patterns during equipment operation.

[0145] The optimization of the feature space is achieved through a multi-dimensional optimization algorithm. The multi-dimensional optimization algorithm is an algorithm that makes each dimension in the feature space more accurately reflect the importance of the current problem by optimizing the feature weights of each dimension. Its role is to adjust the influence of different features on the model through weighted processing to ensure that the correlation between features is effectively expressed. Thus, weighted processing is performed on the features of each dimension so that the influence of each dimension can reflect the actual operating state of the current device. The optimized feature space will provide a basis for subsequent adjustment of the operating mode to ensure that the system can accurately identify the parameter dimensions that need to be key-regulated for the current device.

[0146] Finally, the system dynamically adjusts the operating strategy of the device according to the operating mode parameters generated from the optimized feature space. For example, if the device is currently in a high-load state, the system will adjust the operating mode according to the optimized parameters and focus on regulating the change ranges of temperature and current to prevent overheating or overloading. This process enables the device to always maintain the best operating state under different operating conditions, thereby improving the operating efficiency of the device and extending its service life.

[0147] In a specific application case, in a certain intelligent air-conditioning system, the system analyzes the operating state of the device and finds that there is a strong correlation between temperature fluctuations and current changes. Based on this analysis, the system maps the temperature and current features to the same feature space. The optimized operating mode parameters show that the load state of the air conditioner should be adjusted by increasing the wind speed to regulate the temperature, thereby avoiding energy consumption waste caused by high temperature. At this time, the system adjusts the operating strategy according to the optimized parameters, successfully preventing the overload risk of the device and ensuring the stable operation of the device.

[0148] In step S21, deviation analysis is performed on the multi-source heterogeneous data and the optimized operating mode parameters to obtain a deviation value. If the deviation value is greater than or equal to a preset threshold, a warning signal is obtained, including:

[0149] According to the multi-source heterogeneous data, time alignment and normalization processing are performed to obtain a data input vector;

[0150] According to the data input vector and the optimized operating mode parameters, dimension comparison analysis is performed, and a real-time deviation value is calculated to obtain a deviation sequence;

[0151] If the deviation value of any point in the deviation sequence is greater than the preset threshold, it is marked as an abnormal point, thereby obtaining a warning signal.

[0152] It should be noted that in step S21, the system completes the deviation analysis by comparing the multi-source heterogeneous data with the optimized operation mode parameters, and determines whether to trigger a warning signal based on this. The core of this step is to clarify the deviation degree between the current real-time operation state and the predicted safety baseline, so as to identify potential abnormal trends. The whole process involves multiple refinement steps such as data synchronization, normalization processing, dimension alignment and deviation calculation, with a clear logical path and practical feasibility.

[0153] First of all, the system needs to perform time synchronization and normalization processing on the original multi-source heterogeneous data, aiming to enable data with different dimensions and units to be compared in a unified feature space. This processing process is carried out in accordance with the sampling frequency alignment method, and uniformly maps multi-dimensional data such as temperature, current and vibration into a time series structure with a fixed interval. For example, in practical applications, if the sampling frequency is set to once per minute, all the acquisition points in the last ten minutes are aligned respectively, and the missing data is filled. Then the system performs normalization calculation on each data channel, so that the data distribution of each dimension falls into the numerical space with zero mean and unit variance, thus avoiding the suppression effect of the numerical magnitude of temperature on the vibration amplitude or current fluctuation. After the above processing, the data input vector can be obtained as the unified expression of the real-time operation state.

[0154] On this basis, the system calls the "optimized operation mode parameter set" generated according to the cause analysis and domain adaptation mechanism before, and performs the same normalization processing on this set as the real-time data to ensure the relative consistency of the comparison. In the optimized parameter set here, it includes the target values, optimal regulation values and error tolerance bands for the three types of signals of temperature, current and vibration. For example, the optimization goal of a certain motor is that the temperature is not higher than 62 degrees, the current does not exceed 13 amperes, and the vibration amplitude is controlled within 0.26g. This parameter set represents the ideal regulation baseline that the equipment should reach under the current operation scenario.

[0155] The system performs a dimension-by-dimension comparison and analysis of the real-time data vector with the above-mentioned reference parameter vector. The comparison method is based on Euclidean distance and relative error, calculates the deviation degree between the data of each dimension respectively, and constructs a deviation sequence. Taking the data of a certain minute as an example, if the currently measured temperature is 66 degrees while the optimized parameter target is 62 degrees, then the temperature deviation is 4 degrees. This value will be included in the deviation sequence and marked as an abnormal fluctuation in the temperature dimension. Similarly, if the target value of the current is 13 amperes and the currently measured value is 12.6 amperes, then the deviation value is negative (-0.4 amperes). The system will judge whether the deviation value exceeds the preset threshold. If the deviation value is lower than the set threshold, for example, the current deviation is -0.4 amperes and the threshold is 0.7 amperes, the system considers that this dimension is in a normal state. Only when the deviation value exceeds the threshold will it be marked as an abnormal point and trigger a warning signal. Therefore, a negative deviation value is not considered abnormal if it does not exceed the threshold. Finally, the system forms a multi-dimensional deviation sequence, reflecting the item-by-item differences between the monitored parameters and the target reference within the current time window.

[0156] Next, the system makes a point-by-point threshold judgment on the deviation sequence. The system sets the abnormal judgment threshold for each monitoring dimension, and this threshold is derived from historical statistical analysis and summary of operation experience. For example, for temperature, exceeding the baseline value by 3 degrees can be regarded as the warning critical value; if the current exceeds 0.7 amperes, it is a mild abnormality; once the vibration amplitude exceeds 0.05g, it enters a highly sensitive state. In the above example, the temperature deviation of 4 degrees has exceeded the threshold and is therefore marked as an abnormal point. If the deviation value of any dimension in the deviation sequence exceeds the corresponding threshold, the system judges that there is a potential risk in the current state and marks this time point as an abnormal sample.

[0157] Based on the abnormal point marking results, the system generates corresponding warning signals. The warning signal is not just a binary flag, but an information output with a hierarchical structure, including fields such as warning level, abnormal parameter type, fluctuation direction, and duration. Still taking the aforementioned motor as an example, if the temperature deviation of the system continuously exceeds 4 degrees for three consecutive minutes within a certain time period and shows a continuous upward trend, the system will output a "medium-level temperature rising warning" and attach relevant statistical information. Such warning signals can be sent to the control module to realize real-time correction of the equipment operation strategy, such as reducing the load, switching the working mode, or requesting manual intervention.

[0158] The implementation of this step not only depends on mathematical methods, but also combines the operation laws of industrial equipment, fault triggering characteristics, and engineering experience. Therefore, it has good adaptability and execution effect in actual deployment.

[0159] The working process of the present invention is described below by taking a relatively common scenario as an example. Please also refer to Figure 2 , which is Figure 1 a schematic diagram of the working scenario of the method.

[0160] In an industrial motor operation monitoring scenario, the system is deployed in a factory workshop equipped with multi-sensor devices, and is used to identify and warn in advance of abnormal states or potential faults that may occur during the long-term operation of the equipment. In this working scenario, through the processing, modeling, and identification of multi-source heterogeneous data collected in real time, it is possible to achieve step-by-step fault analysis and prediction from the signal layer to the decision layer. The following combines Figure 1 with Figure 2 the structure and module functions to describe in detail the working process of the present invention.

[0161] In step S11, the system collects multi-source heterogeneous data during the operation of the motor through the data acquisition module. Taking the factory motor as an example, three key physical quantities, namely real-time temperature, vibration, and current, are synchronously obtained during its operation. The temperature sensor uploads the winding temperature data at regular intervals. The vibration sensor is arranged on the surface of the machine shell to collect vibration intensity data. The current transformer is located at the incoming line port to record the real-time working current. These data are all packed and uploaded to the system end in the form of a unified timestamp to form an original data input sequence. Considering that the data update frequencies and noise characteristics of each sensor are different, unified acquisition scheduling management is required to ensure data time synchronization to ensure the accuracy of subsequent processing.

[0162] In step S12, the system performs denoising and normalization processing on the collected multi-source heterogeneous data through the fusion data module. Taking the temperature data as an example, first, the local average filtering method is used to remove the short-period jitter interference, and then it is linearly normalized according to the historical maximum and minimum ranges so that it participates in subsequent calculations on the same scale. The vibration signal uses the wavelet denoising method to eliminate high-frequency interference and adjusts its distribution mean through Z-score normalization. The current signal uses the sliding window method to remove the outliers after removing the power frequency interference. Finally, the three signal channels are all processed into standardized data streams, forming a fusion data tensor and inputting it into the feature extraction module.

[0163] In step S13, the fusion data is input into the CNN neural network feature extraction model built in the feature vector module. This model is trained by a large amount of pre-labeled industrial operation data and has good feature perception ability. The system first passes through the input layer of the model, converts the three-channel fusion data into a three-dimensional tensor and completes dimension alignment. Subsequently, it enters the convolutional hidden layer of the model, uses multi-scale convolutional kernels to extract features from the time series of different data channels, and at the same time uses the channel attention mechanism to enhance the response of the key feature dimensions. After multiple layers of convolution, an intermediate feature map is obtained, and then through the output layer, full connection mapping and feature dimensionality reduction are completed, and finally a set of feature vectors describing the current operating state is output.

[0164] In step S14, the system performs cluster analysis on the extracted feature vector set through the K-means algorithm module to realize dynamic operation mode recognition. First, perform initial clustering processing on the input feature vector, and set the number of clusters k to 3, corresponding to the three types of operation states: normal, slightly abnormal, and severely abnormal. Then calculate the Euclidean distance between samples, and continuously iterate and optimize the clustering results by minimizing the intra-cluster square error criterion. Finally, multiple cluster centers representing the operation state are formed, and these cluster centers are used as the benchmark to count the trend of the center trajectory change of each type of feature sample to construct a dynamic operation mode baseline.

[0165] In step S15, the system inputs the dynamic operation mode baseline into the feature sequence model constructed by LSTM to mine the time evolution law. The model is pre-trained with a large amount of long-term equipment operation data and can effectively learn the trend structure across time segments. The model first expands the operation mode sequence in the input layer, extracts the feature evolution trajectory in the time window, and forms an input unit. After entering the memory layer, the LSTM structure performs phased modeling on the input unit through the gating mechanism, identifies the stable change segment and the violent fluctuation segment, and outputs the trend recognition code. Finally, the model maps the trend code to the precursor feature space through the mapping layer to obtain the precursor feature sequence.

[0166] In step S16, the system first preprocesses the precursor sequence, identifies abnormal change segments and generates an initial set of candidate segments. Then, the set is input into an isolation forest consisting of multiple random subtrees, and the degree of isolation of each segment in the tree structure is calculated by path length to obtain an abnormality score. The higher the score, the more the segment deviates from the overall pattern structure and the higher the possibility of abnormality.

[0167] Then in step S17, the system classifies the feature segments according to the set anomaly score threshold. The threshold is determined by cross-validation during the model training phase and is set to 0.62. In practical applications, if the anomaly score of a precursor segment is greater than or equal to 0.62, it is marked as an abnormal feature, otherwise it is marked as normal. After judging one by one, an anomaly detection result set is formed to provide a basis for subsequent risk assessment.

[0168] In step S18, the system divides the risk level according to the abnormal detection results obtained in the previous step. The risk level division is based on the three dimensions of the density, persistence and relevant signal deviation value of the abnormal points in time distribution. Taking the current motor example, if the abnormal fragment is mainly concentrated in the vibration signal, the duration exceeds 10 minutes, and the current deviation increases significantly, the risk level is assessed as high risk. The final output results are low, medium and high, which are specifically used for risk decision-making and resource scheduling priority sorting.

[0169] In step S19, the system inputs the risk level into a pre-trained fault analysis model. This model is built based on the Transformer neural network structure. Through its encoding layer, the risk level is positionally encoded and transformed into an embedding vector, and then the self-attention layer captures the non-linear dependence relationships between key features. Finally, a linear transformation and activation function output are completed in the decoding layer to obtain a probability distribution representing the fault impact factors. Taking the input of a high risk level in this case as an example, in the output result, the weight of the cause distribution of the temperature feature is as high as 0.74, and the vibration and current are 0.15 and 0.11 respectively, indicating that this risk is mainly caused by abnormal temperature rise.

[0170] In step S20, the system performs an adaptation modeling on the fault cause distribution and adjusts the operating mode to obtain more appropriate control parameters. First, the feature mapping basis of the dominant factor in the fault distribution is extracted, and then it is uniformly projected in the target operating environment through a spatial mapping method. This process uses the maximum mean discrepancy method to align the features of the source domain and the target domain. Taking the current case as an example, it is finally obtained that the temperature control parameter needs to be adjusted down from the set 60 degrees to 58 degrees, and the current setting value is synchronously reduced by 2%, forming new operating mode parameters.

[0171] Finally, in step S21, the system compares the optimized operating mode parameters with the current multi-source heterogeneous data in real time for deviation analysis. First, the three types of real-time signals are time-aligned and normalized to form data vectors, and then a dimension-by-dimension difference calculation is performed with the operating mode parameters to obtain a deviation sequence. If the deviation amount exceeds the threshold setting in any dimension, a corresponding warning signal is generated. In the current scenario, the temperature deviation amount reaches 3.2 degrees, exceeding the set threshold of 2 degrees, and the system issues a "high temperature risk" signal to prompt on-site personnel to conduct inspections and handling.

[0172] Through the above complete workflow, the present invention realizes a full-chain intelligent processing method from signal acquisition, feature extraction, precursor analysis to warning output, significantly improving the timeliness and accuracy of fault identification. Combined with Figure 2 the system architecture shown, each module can flexibly adjust parameters and model structures in actual deployment to adapt to the complex operating requirements under different industrial scenarios.

[0173] Referring to Figure 2 , the second embodiment of the present invention provides a fault warning system based on an AI large model, including:

[0174] A data acquisition module for acquiring multi-source heterogeneous data; wherein, the multi-source heterogeneous data includes: real-time temperature, real-time vibration and real-time current signals;

[0175] A fused data module for performing data denoising and normalization processing on the multi-source heterogeneous data to obtain fused data;

[0176] A feature vector module, configured to input the fusion data into a pre-trained feature extraction model and output a set of feature vectors;

[0177] A K-means algorithm module, configured to perform dynamic operation recognition on the set of feature vectors based on the K-means algorithm to obtain a dynamic operation mode baseline;

[0178] A precursor feature sequence module, configured to input the dynamic operation mode baseline into a pre-trained feature sequence model and output a precursor feature sequence;

[0179] An anomaly score module, configured to calculate a feature anomaly score based on the precursor feature sequence to obtain an anomaly score;

[0180] An anomaly detection result module, configured to mark as an abnormal feature if the anomaly score is greater than or equal to a preset threshold, and mark as a normal feature if the anomaly score is less than the preset threshold, thereby obtaining an anomaly detection result;

[0181] A risk level module, configured to perform risk assessment based on the anomaly detection result to obtain a risk level;

[0182] A fault cause module, configured to input the risk level into a pre-trained fault analysis model to obtain a fault cause distribution;

[0183] An operation model parameter module, configured to determine a set of operation regulation parameters matching the fault cause distribution to obtain optimized operation mode parameters;

[0184] An early warning signal module, configured to perform deviation analysis on the multi-source heterogeneous data and the optimized operation mode parameters to obtain a deviation value, and obtain an early warning signal if the deviation value is greater than or equal to a preset threshold.

[0185] It should be noted that a fault early warning system based on an AI large model provided in an embodiment of the present invention is used to execute all process steps of a fault early warning method based on an AI large model in the above embodiment. The working principles and beneficial effects of the two correspond one by one, and thus will not be elaborated herein.

[0186] An embodiment of the present invention further provides an electronic device. The electronic device includes: a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a fault early warning program based on an AI large model. When the processor executes the computer program, the steps in the above embodiments of the fault early warning method based on an AI large model are implemented, such as Figure 1 the step S11 shown. Alternatively, when the processor executes the computer program, the functions of each module / unit in the above device embodiments are implemented, such as the early warning signal module.

[0187] Exemplarily, the computer program may be divided into one or more modules / units, and the one or more modules / units are stored in the memory and executed by the processor to implement the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing specific functions, and these instruction segments are used to describe the execution process of the computer program in the electronic device.

[0188] The electronic device may be a computing device such as a desktop computer, a notebook, a palm computer, and a smart tablet. The electronic device may include, but is not limited to, a processor and a memory. Those skilled in the art can understand that the above components are only examples of the electronic device and do not constitute a limitation on the electronic device. It may include more or fewer components than the above, or combine certain components, or different components. For example, the electronic device may further include input / output devices, network access devices, a bus, etc.

[0189] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. In addition, in the drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that they have a communication connection, which can be specifically implemented as one or more communication buses or signal lines. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0190] The specific embodiments described above further elaborate on the purpose, technical solutions, and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the protection scope of the present invention. It is particularly pointed out that for those skilled in the art, any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A fault warning method based on an AI large model, characterized in that, Including: Obtain multi-source heterogeneous data; among them, the multi-source heterogeneous data includes: real-time temperature, real-time vibration, and real-time current signals; According to the multi-source heterogeneous data, perform data denoising and normalization processing to obtain fused data; Input the fused data into a pre-trained feature extraction model to output a set of feature vectors; Based on the K-means algorithm, perform dynamic operation recognition on the set of feature vectors to obtain a dynamic operation mode baseline; Input the dynamic operation mode baseline into a pre-trained feature sequence model to output a precursor feature sequence; According to the precursor feature sequence, calculate the feature anomaly score to obtain an anomaly score; If the anomaly score is greater than or equal to a preset threshold, it is marked as an abnormal feature. If the anomaly score is less than the preset threshold, it is marked as a normal feature, thereby obtaining an anomaly detection result; According to the anomaly detection result, perform risk assessment to obtain a risk level; Input the risk level into a pre-trained fault analysis model to obtain a fault cause distribution; According to the fault cause distribution, determine a set of operation regulation parameters that match it to obtain optimized operation mode parameters; Perform deviation analysis on the multi-source heterogeneous data and the optimized operation mode parameters to obtain a deviation value. If the deviation value is greater than or equal to the preset threshold, obtain a warning signal.

2. The fault warning method based on the AI large model according to claim 1, characterized in that The step of inputting the fused data into a pre-trained feature extraction model to output a set of feature vectors includes: The feature extraction model is trained by a CNN neural network model; Through the input layer of the feature extraction model, perform dimension matching on the fused data to obtain an input tensor; Through the hidden layer of the feature extraction model, perform multi-scale convolution operations and channel information extraction on the input tensor to obtain an intermediate feature map; Through the output layer of the feature extraction model, perform fully connected mapping and dimensionality reduction processing on the intermediate feature map to obtain a set of feature vectors.

3. The fault warning method based on the AI large model according to claim 1, wherein The step of performing dynamic operation recognition on the set of feature vectors based on the K-means algorithm to obtain a dynamic operation mode baseline includes: According to the set of feature vectors, perform clustering initialization processing to obtain an initial clustering structure; Based on the K-means algorithm, perform Euclidean distance measurement and intra-cluster minimization iterative operations on the initial clustering structure to obtain multiple dynamic operation cluster results; According to the multiple dynamic operation cluster results, perform central trajectory calculation and analyze the feature distribution law to obtain a dynamic operation mode baseline.

4. The fault warning method based on the AI large model according to claim 1, characterized in that, The step of inputting the dynamic operation mode baseline into a pre-trained feature sequence model to output a precursor feature sequence includes: The feature sequence model is trained by an LSTM neural network model; Through the input layer of the feature sequence model, perform time step expansion and structure reorganization on the dynamic operation mode baseline, and extract the feature evolution trajectory across time segments to obtain input units; Through the memory layer of the feature sequence model, perform phased dynamic state modeling on the input units to obtain a trend recognition coding vector; Through the mapping layer of the feature sequence model, perform precursor signal transformation on the trend recognition coding vector to obtain a precursor feature sequence.

5. The fault warning method based on the AI large model according to claim 1, characterized in that Performing feature anomaly score calculation based on the precursor feature sequence to obtain an anomaly score, including: Performing outlier detection preprocessing on the precursor feature sequence to obtain an initial set of anomaly candidate segments; Based on the Isolation Forest algorithm, calculating the anomaly score for the initial set of anomaly candidate segments to obtain the anomaly score.

6. The fault warning method based on the AI large model according to claim 1, wherein Inputting the risk level into a pre-trained fault analysis model to obtain the fault cause distribution, including: The fault analysis model is trained by a Transformer neural network structure; Through the encoding layer of the fault analysis model, performing position encoding processing on the risk level to obtain a risk embedding vector sequence Through the self-attention layer of the fault analysis model, performing high-order correlation feature extraction on the risk embedding vector sequence to obtain an attention-weighted expression result; Through the decoding layer of the fault analysis model, performing linear mapping and activation transformation on the attention-weighted result to obtain the fault cause distribution.

7. The fault warning method based on the AI large model according to claim 1, wherein Determining a set of operation regulation parameters that match the fault cause distribution to obtain optimized operation mode parameters, including: Performing domain feature extraction based on the fault cause distribution to obtain a domain feature mapping basis; Performing unified mapping of the feature space based on the domain feature mapping basis to obtain an adapted operation feature structure; Based on the adapted operation feature structure, determining the parameter dimensions that need to be key-regulated in the current operation state to obtain optimized operation mode parameters.

8. The fault warning method based on the AI large model according to claim 1, characterized in that Performing deviation analysis on the multi-source heterogeneous data and the optimized operation mode parameters to obtain a deviation value. If the deviation value is greater than or equal to a preset threshold, obtaining a warning signal, including: Performing time alignment and normalization processing on the multi-source heterogeneous data to obtain a data input vector; Performing dimensional comparison analysis on the data input vector and the optimized operation mode parameters and calculating the real-time deviation value to obtain a deviation sequence; If the deviation value at any point in the deviation sequence is greater than the preset threshold, marking it as an abnormal point to obtain a warning signal.

9. A fault warning system based on an AI large model, characterized in that, Including: A data acquisition module for acquiring multi-source heterogeneous data; among them, the multi-source heterogeneous data includes: real-time temperature, real-time vibration, and real-time current signals; A fusion data module for performing data denoising and standardization processing on the multi-source heterogeneous data to obtain fusion data; A feature vector module for inputting the fusion data into a pre-trained feature extraction model and outputting a set of feature vectors; A K-means algorithm module for performing dynamic operation recognition on the set of feature vectors based on the K-means algorithm to obtain a dynamic operation mode baseline; A precursor feature sequence module for inputting the dynamic operation mode baseline into a pre-trained feature sequence model and outputting a precursor feature sequence; An anomaly score module for performing feature anomaly score calculation based on the precursor feature sequence to obtain the anomaly score; Anomaly detection result module, which is used to mark as an abnormal feature if the anomaly score is greater than or equal to a preset threshold, and mark as a normal feature if the anomaly score is less than the preset threshold, so as to obtain the anomaly detection result; Risk level module, which is used to conduct risk assessment according to the anomaly detection result to obtain the risk level; Fault cause module, which is used to input the risk level into a pre-trained fault analysis model to obtain the fault cause distribution; Operating model parameter module, which is used to determine a set of operation regulation parameters matching the fault cause distribution according to the fault cause distribution to obtain the optimized operation mode parameters; Early warning signal module, which conducts deviation analysis on the multi-source heterogeneous data and the optimized operation mode parameters to obtain a deviation value, and if the deviation value is greater than or equal to a preset threshold, an early warning signal is obtained.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein when the computer program runs, it controls the device where the computer-readable storage medium is located to execute the fault early warning method based on the AI large model according to any one of claims 1 to 8.

Images