Industrial equipment intelligent operation and maintenance method based on multi-source heterogeneous data dynamic acquisition and LSTM optimization

Abstract

The invention discloses an industrial equipment intelligent operation and maintenance method based on multi-source heterogeneous data dynamic collection and LSTM optimization, and relates to the technical field of industrial Internet of Things and industrial equipment intelligent operation and maintenance, and the method comprises the steps: collecting multi-source heterogeneous data, transmitting the multi-source heterogeneous data to an edge node, and carrying out the data preprocessing through a data preprocessing module; constructing a lightweight multi-modal LSTM model, and identifying and predicting the fault of the industrial equipment; through a self-adaptive threshold algorithm, differential weighted sampling and dynamic balance of energy consumption and precision, edge nodes realize real-time fault detection, and an operation and maintenance decision of local industrial equipment is generated; moreover, cross-factory cooperative training is carried out through a federal learning platform, and the generalization ability of the model is improved. Therefore, by the adoption of the industrial equipment intelligent operation and maintenance method based on multi-source heterogeneous data dynamic collection and LSTM optimization, the problem that response to hidden faults is lagged in a traditional method can be solved, optimal resource configuration in an industrial scene is achieved, and the operation and maintenance cost of an industrial Internet of Things terminal is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial Internet of Things and intelligent operation and maintenance of industrial equipment, and in particular to an intelligent operation and maintenance method for industrial equipment based on dynamic acquisition of multi-source heterogeneous data and LSTM optimization. Background Art

[0002] With the rapid development of industrial automation and informatization, the operation and maintenance management of industrial equipment faces increasingly high requirements. However, traditional operation and maintenance methods for industrial equipment mainly rely on manual inspections and threshold-triggered alarm mechanisms based on single-sensor data, making it difficult to meet the requirements of modern industrial production for equipment reliability and operation and maintenance efficiency.

[0003] Currently, the traditional operation and maintenance methods for industrial equipment have the following deficiencies:

[0004] First, the limitation of relying on single-sensor data. The triggering mechanism of traditional dynamic acquisition thresholds for multi-source heterogeneous data mostly relies on a single type of sensor (such as a vibration sensor). This technical solution believes that the operating state of the equipment depends on the change of a certain key physical quantity. However, some latent faults often do not show obvious abnormalities in a single parameter, but involve the combined action of multiple factors, such as changes in the electrical performance inside the equipment and damage to the microscopic structure of materials, which cannot be captured by a single type of sensor. Therefore, this method has defects in dealing with comprehensive data affected by multiple factors and is difficult to effectively and real-time respond to latent faults.

[0005] Second, there are drawbacks to the fixed sampling frequency. In order to ensure the consistency and regularity of data acquisition and facilitate subsequent data processing, traditional dynamic acquisition methods for multi-source heterogeneous data mostly adopt a sampling strategy with a fixed frequency. However, this method does not fully consider the dynamic characteristics of the equipment operating state in real time and fails to adjust the sampling frequency of the equipment in real time according to the actual state of the equipment, resulting in a very high data redundancy rate and huge acquisition energy consumption. In a typical industrial case, when the equipment is in the normal operation stage, many of the data generated by it change very little in a short period of time. However, the system will still collect a large number of similar data at a fixed frequency, which not only occupies a large amount of storage resources but also consumes the energy of the sensor and its transmission link.

[0006] Third, complex models are challenging in edge computing scenarios. Traditional methods use intelligent operation and maintenance models based on dynamic collection of multi-source heterogeneous data to comprehensively and deeply process multi-source data and mine potential patterns in the data. This type of intelligent model is usually complex in structure and contains a large number of parameters and computing nodes. However, in edge computing scenarios, due to the limited computing resources of edge devices, their performance cannot adapt to traditional intelligent models, resulting in long edge reasoning delays, inability to complete reasoning quickly, and inability to meet the needs of real-time response, resulting in the inability to capture early signs of equipment failure and respond in time. In actual industrial production, equipment failures often occur very quickly, and real-time response is crucial to taking timely measures to avoid serious losses.

[0007] Fourth, there is the problem of balancing the generalization ability of the model and data privacy. Under the traditional centralized training model, the generalization ability of the industrial equipment operation and maintenance model is limited by the data distribution in a single factory, and it is difficult to adapt to the differences in equipment in different factories and the uneven distribution of data. In addition, the centralized training model also requires a large amount of sensitive data to be uploaded to the cloud for training, which faces the risk of data privacy leakage. This not only limits the performance improvement of the model, but also poses a potential threat to the data security of the enterprise.

[0008] In summary, the existing technology has many deficiencies in terms of multi-source heterogeneous data collection, lightweight and real-time intelligent operation and maintenance models, energy consumption optimization, and model generalization capabilities, and cannot meet the needs of modern industrial equipment for intelligent operation and maintenance with high precision, low latency, low energy consumption, and high adaptability. Therefore, a new intelligent operation and maintenance method for industrial equipment is urgently needed to overcome the defects of the existing technology and realize efficient and intelligent operation and maintenance of industrial equipment. Summary of the invention

[0009] The purpose of the present invention is to provide an intelligent operation and maintenance method for industrial equipment based on dynamic collection of multi-source heterogeneous data and LSTM optimization, which can utilize multi-source heterogeneous data to identify hidden equipment faults and reduce operation and maintenance costs; at the same time, it can effectively solve the problems of large differences in industrial scene equipment and uneven data distribution.

[0010] To achieve the above object, the present invention provides an industrial equipment intelligent operation and maintenance method based on dynamic collection of multi-source heterogeneous data and LSTM optimization, comprising the following steps:

[0011] S1. Collect multi-source heterogeneous data from various sensors and equipment maintenance logs deployed on industrial equipment, where the multi-source heterogeneous data has different modalities and communication protocols;

[0012] S2, transmitting the multi-source heterogeneous data collected in step S1 to the edge node, and then preprocessing the multi-source heterogeneous data through a data preprocessing module, including:

[0013] Automatically identify and standardize data of different communication protocols using protocol-independent interfaces, and then use the DSRN algorithm to normalize multi-rate data to obtain time-aligned equally-spaced sequence data;

[0014] S3. Use a pre-trained TS-GAN model for data augmentation to synthesize rare fault samples, expand the characteristic patterns of hidden faults in the training data, and perform deep feature extraction through a lightweight multi-modal LSTM model that integrates CNN, BERT, and LSTM to identify and predict faults in industrial equipment;

[0015] S4. According to the historical fault probability distribution, combined with the fault probability predicted in step S3, evaluate the operating risk level of industrial equipment, dynamically adjust the triggering conditions for industrial equipment sampling through an adaptive threshold algorithm, and use the attention mechanism to dynamically allocate weights to each sensor to achieve a differentiated sampling frequency to balance the energy consumption and detection accuracy of industrial equipment;

[0016] S5. Based on the fault data detected in real time by the edge node, directly make operation and maintenance decisions for local industrial equipment, and at the same time upload relevant fault data and local model update data to the cloud server for cross-plant collaborative training through the federated learning platform to improve the generalization ability of the model.

[0017] Preferably, in step S3, a lightweight multi-modal LSTM model is constructed through INT8 quantization and structured pruning.

[0018] Preferably, in step S3, a multi-modal LSTM model without pruning is pre-trained in the cloud using a large amount of industrial data. Key structural units are selected by absolute gradient or absolute weight, and channels, attention heads, or hidden units with low importance are deleted proportionally to obtain a structured-pruned multi-modal LSTM model; then, the weight parameters and activation values in the structured-pruned multi-modal LSTM model are converted from 32-bit floating-point type to 8-bit integer type, where the weight parameters use symmetric quantization and the activation values use asymmetric quantization, thus obtaining a lightweight multi-modal LSTM model.

[0019] Preferably, in step S3, CNN is used to process image data, BERT is used to process text data, and LSTM is used to process time-series data, and the features extracted by each part are used to identify the non-linear weak associations of multi-modal data through a cross-validation mechanism.

[0020] Preferably, in step S4, it includes collecting historical operation data and corresponding fault labels of industrial equipment to form a time-series data set, and using the sliding window method to calculate the historical fault probability to form a working condition-probability mapping table.

[0021] Preferably, according to the historical failure probability and the predicted failure probability in step S3, evaluate the risk level of the equipment state, and then set a threshold according to the risk level to dynamically adjust the triggering condition of industrial equipment sampling.

[0022] Preferably, the expression of the risk level is:

[0023] R(t) = αP pred (t) + (1 - α)P hist (t);

[0024] Wherein, P pred (t) is the predicted failure probability of the equipment at the current time t, P hist (t) is the historical failure probability, and α ∈ [0, 1] represents the real-time prediction weight;

[0025] θ(t) = θ base ·(1 - βR(t));

[0026] Wherein, θ base is the basic threshold, set based on the statistical mean of the corresponding characteristics during normal operation of the equipment; θ is the threshold adjustment coefficient.

[0027] Preferably, the triggering condition of industrial equipment sampling satisfies at least two modes exceeding the set threshold.

[0028] Preferably, in step S4, it includes assigning dynamic weights to each sensor through an attention mechanism according to the contribution degree of each sensor in historical failures, and dynamically adjusting the threshold adjustment coefficient of each sensor to achieve different sampling frequencies, as follows:

[0029] β i = βω i ;

[0030] Wherein, β i is the threshold adjustment coefficient of the i-th sensor, and ω i represents the weight of the i-th sensor.

[0031] Preferably, in step S4, by calculating the mutual information or correlation coefficient between the data of each sensor and the fault label, quantifying the prediction ability of the sensor data for faults, and then normalizing the mutual information of each sensor to obtain the weight of the corresponding sensor to measure the contribution degree of each sensor in historical failures, as follows:

[0032]

[0033] Wherein, MI(s i , y) represents the mutual information between the data s i of the i-th sensor and the fault label y, and n represents the total number of sensors.

[0034] Therefore, the present invention adopts the above-mentioned intelligent operation and maintenance method for industrial equipment based on dynamic acquisition of multi-source heterogeneous data and LSTM optimization, and has the following technical effects:

[0035] (1) The present invention maps multi-source heterogeneous data such as vibration, temperature, infrared images, and maintenance logs to a unified feature space through a cross-modal adapter, and combines a lightweight multi-modal LSTM model to achieve accurate identification of latent faults in equipment; compared with the traditional single-sensor threshold trigger mechanism, the present invention uses multi-dimensional feature cross-validation and deep feature extraction to improve the accuracy of fault detection, and successfully breaks through the technical bottleneck of the traditional method's lag in response to latent faults.

[0036] (2) The present invention realizes the efficient alignment of multi-rate data by combining three core algorithms: the DSRN dynamic sampling rate normalization algorithm, the adaptive threshold trigger mechanism, and the differential weighted sampling strategy, and retains high-frequency details and low-frequency event features; at the same time, based on the historical fault probability distribution and attention weights, the sampling conditions are dynamically adjusted to optimize the sensor energy consumption allocation. Under the synergistic effect of the triple mechanism, the data redundancy rate is reduced, the network bandwidth occupancy is reduced, and the optimal resource allocation for data acquisition and transmission in industrial scenarios is achieved while ensuring the integrity of key information.

[0037] (3) The present invention adopts a dual optimization mechanism of the edge node intelligent sleep strategy and model lightweight technology, dynamically regulates the sleep / working state of non-critical sensors through the risk level, reduces the energy consumption of edge nodes; compresses the volume of the multi-modal LSTM model through INT8 quantization and structured pruning, shortens the inference delay, meets the industrial real-time requirement (less than 20ms), and at the same time extends the battery life of edge devices by 3 times, significantly reducing the operation and maintenance costs of industrial Internet of Things terminals.

[0038] (4) The present invention constructs an edge-cloud collaborative federated learning architecture, and through cross-plant equipment data collaborative training, improves the generalization ability of the model on the premise of ensuring data privacy; compared with the traditional centralized training mode, the technical solution of the present invention can effectively solve the problems of large equipment differences and uneven data distribution in industrial scenarios, realize the technical leap from single-plant to multi-plant collaborative operation and maintenance, and significantly improve the robustness and adaptability of the system in complex industrial environments.

[0039] The technical solution of the present invention will be further described in detail below through the drawings and embodiments. Description of the Drawings

[0040] Figure 1 is a flowchart of an intelligent operation and maintenance method for industrial equipment based on dynamic acquisition of multi-source heterogeneous data and LSTM optimization. Detailed Embodiments

[0041] The present invention can be explained in more detail by the following embodiments. The purpose of disclosing the present invention is to protect all changes and improvements within the scope of the present invention. The present invention is not limited to the following embodiments.

[0042] Please refer to Figure 1 , the present invention provides an intelligent operation and maintenance method for industrial equipment based on dynamic acquisition of multi-source heterogeneous data and LSTM optimization, including five key steps: multi-source heterogeneous data acquisition, data preprocessing and fusion, LSTM model prediction, dynamic adaptive data acquisition strategy adjustment, and edge-cloud collaboration, which are specifically as follows:

[0043] S1. Deploy various sensors (such as vibration sensors, temperature sensors, cameras for obtaining infrared images, etc.) on industrial equipment, and there is a storage device for recording maintenance logs. Multi-source heterogeneous data is collected through the deployed sensors and maintenance logs. These multi-source heterogeneous data have different modalities (such as time-series data, image data, text data) and communication protocols (such as Modbus, OPC UA, MQTT, etc.).

[0044] S2. Transmit the multi-source heterogeneous data collected in step S1 to the edge node, and then use the data preprocessing module of the edge node to preprocess the multi-source heterogeneous data. Among them, the data preprocessing module needs to have sufficient computing resources to run cross-modal adapters, protocol-independent interfaces, and dynamic sampling rate normalization algorithms (DSRN algorithms), specifically including:

[0045] (1) Standardize data with different communication protocols (such as Modbus, OPCUA, MQTT, etc.) through protocol-independent interfaces, and parse the original binary or text data into structured data in a unified format (such as JSON, CSV), including meta-information such as sensor ID, timestamp, value / text / image path, etc.

[0046] (2) Multi-rate data generated due to different sensors or data sources in industrial equipment sampling the same physical process or state at different frequencies according to different monitoring requirements, including high-frequency time-series data, low-frequency time-series data, fixed-frame-rate data, and non-uniformly sampled data. Among them, vibration sensors collect vibration data at a sampling rate of 10 kHz, which belongs to high-frequency time-series data and is used to capture the high-frequency details of the equipment vibration waveform; temperature sensors collect temperature data at a sampling rate of 1 Hz, which belongs to low-frequency time-series data and is used to monitor the slow change trend of temperature; the infrared images collected by cameras belong to periodic fixed-rate data; the equipment maintenance logs are recorded according to event triggers and have no fixed sampling frequency. Therefore, it is necessary to normalize the multi-rate data. In this embodiment, through the DSRN algorithm, the multi-rate data is converted into time-aligned equally spaced sequence data, so that there are corresponding vibration amplitudes, temperature values, image feature vectors, log event markers, etc. at each time point t, specifically as follows:

[0047] Time-axis alignment: Select a minimum time unit (such as 1 ms) as the benchmark, map all data to the same time axis, and make all data have a unified reference time interval. For example, for high-frequency vibration data, downsample according to the reference time interval to retain key waveform features; for low-frequency temperature data, fill in the missing intermediate points through linear interpolation or polynomial fitting to generate a continuous time series with the reference time interval; for fixed-rate image data, align it to the reference time axis according to the timestamp to ensure that each time point corresponds to the image features of the nearest frame.

[0048] Feature retention strategy: For high-frequency data such as vibration signals, anti-aliasing filtering (low-pass filter) needs to be adopted to avoid signal distortion caused by downsampling and retain key features such as vibration amplitude and frequency components. For low-frequency data such as temperature and equipment maintenance logs, it is necessary to give priority to retaining extreme points (such as temperature mutation points and maintenance event time points) during interpolation to avoid losing important event information due to smoothing processing.

[0049] S3. In the edge node, it is necessary to load the pre-trained TS-GAN model to perform data augmentation on the preprocessed data, synthesize rare fault samples, expand the feature patterns of hidden faults in the training data, and enable the model to learn the weak associations of multi-modal features in the early stage of faults. For example, there is no significant change in the vibration amplitude at the initial stage of a crack, but the local temperature gradient of the infrared image is abnormal, so as to solve the small-sample learning problem.

[0050] At the same time, based on INT8 quantization and structured pruning, a lightweight multi-modal LSTM model integrating CNN, BERT, and LSTM is constructed, specifically as follows:

[0051] (1) Use a large amount of industrial data in the cloud to pre-train an unpruned multi-modal model. Screen key structural units through the absolute value of the gradient or the absolute value of the weights, delete channels, heads or units with low importance in proportion, keep the model structure regularized for easy hardware acceleration, and fine-tune the pruned model with local data at the edge node to compensate for the performance loss caused by structural simplification. Among them, for CNN, prune the channels of the convolutional layer, calculate the sum of the absolute values of the weights of each channel, and delete the 20% channels with the lowest contribution according to the threshold (for example, prune 64 channels to 51 channels) to retain the high-frequency feature extraction ability; for the BERT module, prune the attention heads and layers. For the multi-layer Transformer encoder, delete the redundant attention heads at the bottom or top layers and retain the core semantic encoding layer to focus on the extraction of key events in the maintenance log; for the LSTM module, prune the hidden layer units. By calculating the gradient contribution of each hidden unit to the fault prediction, delete the 30% low-contribution units to reduce the redundant calculation of time series modeling.

[0052] (2) After structurally pruning the model to delete redundant structures, convert the weight parameter Weight and the activation value Activation from 32-bit floating-point type (FP32) to 8-bit integer type (INT8), so that the lightweight multi-modal LSTM model meets the computing power limit of the edge node (CPU single-core computing power ≤ 1GHz), while retaining the multi-dimensional feature cross-validation ability (the fusion inference of CNN image features and LSTM time series features).

[0053] Among them, the weight matrix adopts a symmetric quantization strategy, that is, assuming that the weight distribution is symmetric about 0, determine the quantization scale factor Scale by calculating the maximum absolute value. The formula is:

[0054]

[0055] Then calculate the quantized weight value w, and the value range is [-127, 127]. The expression is:

[0056]

[0057] In the formula, round() represents the rounding function.

[0058] The activation value (such as the CNN output feature map, the BERT hidden layer output) then adopts asymmetric quantization. Considering the non-zero mean distribution, determine the quantization scale factor Scale and the zero point Zero Point by calculating the minimum value min and the maximum value max. At this time, the formula is:

[0059]

[0060] Then calculate the quantized activation value Among them, the clip() function means to control the activation value range within [0, 255].

[0061] Through the above two improvements, the integrated CNN processes image data to extract spatial features such as the surface temperature distribution of the device and the location of abnormal hot spots, which are used to identify physical damage on the device surface, such as local overheating caused by bearing wear, to supplement static or quasi-static abnormalities that cannot be captured by time-series data (such as visual features of bolt loosening), and to avoid missed detection by a single sensor; BERT processes text data, converts text data such as maintenance logs into Token sequences, and then uses the BERT model to extract semantic features such as fault keywords and historical maintenance event associations, to mine the historical laws of hidden faults, such as the causal relationship between "motor overload" and "insufficient bearing lubrication", and then combines timestamps to convert the text into event sequence features (such as the time interval between maintenance operations and the recurrence frequency of the same type of fault), to assist the time-series model in predicting the fault cycle; LSTM processes time-series data to capture dynamic dependencies in the time series, such as the periodic change of vibration amplitude over time, to identify the time-series pattern of fault development, and to model the long-term dependencies of multi-rate time-series data (such as the performance degradation trend after the device has run for thousands of hours), to make up for the defect that traditional threshold alarms cannot predict progressive faults.

[0062] Then, the three types of extracted features are concatenated along the dimension to form multi-modal fusion features, which are used as the input of the subsequent prediction layer to predict the fault probability of industrial equipment and output the fault category at the current moment (such as normal, minor fault, serious fault). During the detection process, a cross-validation mechanism is also set up to identify non-linear weak associations in multi-modal data. For example, there is no significant change in the vibration signal, but the temperature sensor shows a slight temperature rise, and there is a partial discharge spot in the infrared image, and the maintenance log shows that "the insulation detection period has expired". The multi-modal data comprehensively shows that there is a hidden fault of insulation layer aging inside the device, and this abnormality is not obvious in the single-sensor data. It can be seen that the method of this embodiment can improve the accuracy of fault detection and at the same time improve the detection efficiency of the system, and successfully breaks through the technical bottleneck of the traditional method's lag in response to hidden faults.

[0063] S4. Collect historical operation data of industrial equipment (including time-series, image, and text multi-modal data) and corresponding fault labels to form a time-series data set, and use the sliding window method to calculate the historical fault probability P hist (t), which reflects the frequency of frequent faults of the device in the near future. Combine the device working condition data (such as load, speed), classify and count the fault probabilities under different working conditions (such as the fault probability increases significantly under high load) to form a working condition-probability mapping table. Combine the fault probability P pred (t) of the device predicted at the current moment in step S3 to evaluate the risk level of the current device state as follows:

[0064] R(t) = αP pred (t) + (1 - α)P hist (t);

[0065] Where α ∈ [0, 1] represents the real-time prediction weight, ensuring the leading role of recent prediction results in threshold adjustment, and taking 0.7 - 0.9 in high real-time scenarios.

[0066] When R(t) ≥ 0.7, it is regarded as a high risk, triggering the high-precision mode, with the sensor sampling at full frequency and disabling the sleep mode to ensure real-time acquisition of all-modal data; when 0.3 ≤ R(t) < 0.7, it is regarded as a medium risk, maintaining the conventional sampling strategy; when R(t) < 0.3, it is regarded as a low risk, entering the low-power consumption mode, and non-critical sensors reduce frequency or go to sleep.

[0067] This embodiment also designs an adaptive threshold algorithm, setting the threshold according to the risk level and dynamically adjusting the trigger conditions for industrial equipment sampling. Among them, the threshold θ(t) changes inversely with the risk level R(t), as follows:

[0068] θ(t) = θ base ·(1 - βR(t));

[0069] Where θ base is the basic threshold, set based on the characteristic statistical mean when the equipment is running normally; β is the threshold adjustment coefficient, and the set range is 0.5 - 0.8 according to experience.

[0070] In high-risk scenarios, the threshold θ(t) decreases, and the trigger condition is more easily met, that is, the sensor samples at a higher frequency; while in low-risk scenarios, the threshold θ(t) increases, and the trigger condition is more difficult to meet, and the sampling frequency of non-critical sensors decreases. In addition, the trigger condition needs to simultaneously meet that at least two modal characteristics exceed the threshold to avoid false triggering caused by single-modal noise.

[0071] At the same time, according to the contribution degree of each sensor in historical faults, dynamic weights are assigned to the sensors through the attention mechanism. The threshold adjustment of high-weight sensors is more sensitive, so as to achieve different sampling frequencies for each sensor, and be able to dynamically balance energy consumption and accuracy according to the running risk degree of industrial equipment, and reduce the sampling frequency of non-critical sensors or put non-critical sensors to sleep in low-risk situations. In this process, the contribution degree of each sensor is calculated by the mutual information or correlation coefficient between the sensor data and the fault label, quantifying the prediction ability of the sensor data for faults, and then normalizing the mutual information of each sensor to obtain the weight of the corresponding sensor, as follows:

[0072]

[0073] Where ω idenotes the weight of the i-th sensor, and MI(s i , y) represents the data s of the i-th sensor i and the mutual information with the fault label y, where n represents the total number of sensors.

[0074] During the training process of the lightweight multi-modal LSTM model, the influence degree of each sensor feature on the fault prediction result is calculated through gradient backpropagation or feature perturbation method. For example, when perturbing the input feature of a certain sensor, if the model prediction accuracy drops significantly, the weight of this sensor is relatively high. In addition, according to the real-time fault detection results and newly collected data, the weights of the corresponding sensors need to be dynamically updated (recalculated once a day / week) to adapt to the drift of sensor importance caused by equipment aging, working condition changes, etc.

[0075] S5. Establish a stable network connection between the edge node and the cloud server, and a federated learning platform needs to be built on the cloud server. On the one hand, the fault data detected in real time by the edge node is directly used for the operation and maintenance decision-making of local industrial equipment. On the other hand, the fault data detected in real time by the edge node (such as the predicted fault label and probability, the multi-modal fusion feature vector at the time of fault occurrence) and the updated information (gradient parameters or model weight increments) trained locally are uploaded to the cloud server, and cross-plant collaborative training is carried out through the federated learning platform, which can not only protect data privacy but also improve the generalization ability of the model.

[0076] Therefore, the present invention adopts the above-mentioned industrial equipment intelligent operation and maintenance method based on dynamic acquisition of multi-source heterogeneous data and LSTM optimization. Through the multi-modal data fusion architecture, the fault detection accuracy is improved; at the same time, through the efficient data processing and transmission process, the data redundancy rate is reduced, the network bandwidth occupation is reduced, the integrity of key information is guaranteed, the optimal resource allocation of data acquisition and transmission in the industrial scenario is realized, and the operation and maintenance cost of industrial Internet of Things terminals is significantly reduced.

[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An intelligent operation and maintenance method for industrial equipment based on dynamic acquisition of multi-source heterogeneous data and LSTM optimization, characterized in that, Including the following steps: S1. Collect multi-source heterogeneous data from various sensors deployed on industrial equipment and equipment maintenance logs. The multi-source heterogeneous data has different modalities and communication protocols. S2. Transmit the multi-source heterogeneous data collected in step S1 to the edge node, and then preprocess the multi-source heterogeneous data through a data preprocessing module, including: Automatically identify and standardize data with different communication protocols using a protocol-independent interface, and then normalize multi-rate data using the DSRN algorithm to obtain time-aligned equally-spaced sequence data. S3. Use a pre-trained TS-GAN model for data augmentation to synthesize rare fault samples, expand the feature patterns of latent faults in the training data, and perform deep feature extraction through a lightweight multi-modal LSTM model integrating CNN, BERT, and LSTM to identify and predict faults in industrial equipment. S4. According to the historical fault probability distribution, combined with the fault probability predicted in step S3, evaluate the operating risk level of industrial equipment, dynamically adjust the trigger conditions for industrial equipment sampling through an adaptive threshold algorithm, and use an attention mechanism to dynamically allocate weights to each sensor to achieve a differentiated sampling frequency to balance the energy consumption and detection accuracy of industrial equipment. S5. Based on the fault data detected in real time by the edge node, directly make operation and maintenance decisions for local industrial equipment, and at the same time upload relevant fault data and local model update data to the cloud server for cross-plant collaborative training through a federated learning platform to improve the generalization ability of the model.

2. The intelligent operation and maintenance method for industrial equipment based on dynamic acquisition of multi-source heterogeneous data and LSTM optimization according to claim 1, characterized in that, In step S3, a lightweight multi-modal LSTM model is constructed through INT8 quantization and structured pruning.

3. An intelligent operation and maintenance method for industrial equipment based on dynamic acquisition of multi-source heterogeneous data and LSTM optimization according to claim 2, characterized in that In step S3, a multi-modal LSTM model without pruning is pre-trained in the cloud using a large amount of industrial data. Key structural units are selected through the absolute value of the gradient or the absolute value of the weight, and channels, attention heads, or hidden units with low importance are deleted proportionally to obtain a structured-pruned multi-modal LSTM model. Then, the weight parameters and activation values in the structured-pruned multi-modal LSTM model are converted from 32-bit floating-point type to 8-bit integer type. The weight parameters use symmetric quantization, and the activation values use asymmetric quantization, thereby obtaining a lightweight multi-modal LSTM model.

4. The intelligent operation and maintenance method for industrial equipment based on dynamic acquisition of multi-source heterogeneous data and LSTM optimization according to claim 3, characterized in that In step S3, CNN is used to process image data, BERT is used to process text data, and LSTM is used to process time-series data. The features extracted by each part identify the non-linear weak associations of multi-modal data through a cross-validation mechanism.

5. The intelligent operation and maintenance method for industrial equipment based on dynamic acquisition of multi-source heterogeneous data and LSTM optimization according to claim 1, characterized in that, In step S4, it includes collecting historical operation data and corresponding fault labels of industrial equipment to form a time-series data set, and using the sliding window method to calculate the historical fault probability to form a working condition-probability mapping table.

6. The intelligent operation and maintenance method of industrial equipment based on dynamic acquisition of multi-source heterogeneous data and LSTM optimization according to claim 5, characterized in that, Based on the historical fault probability and the fault probability predicted in step S3, evaluate the risk level of the equipment state, and then set a threshold according to the risk level to dynamically adjust the trigger conditions for industrial equipment sampling.

7. An intelligent operation and maintenance method for industrial equipment based on dynamic acquisition of multi-source heterogeneous data and LSTM optimization according to claim 6, characterized in that, The expression for the risk level is: R(t) = αP pred (t) + (1 - α)P hist (t); where, P pred (t) is the predicted failure probability of the device at the current moment t, and P hist (t) is the historical failure probability, and α ∈ [0, 1] represents the real-time prediction weight; θ(t) = θ base ·(1 - βR(t)); where θ base is the base threshold, set based on the statistical mean of the corresponding feature during normal operation of the device; θ is the threshold adjustment coefficient.

8. An intelligent operation and maintenance method for industrial equipment based on dynamic acquisition of multi-source heterogeneous data and LSTM optimization according to claim 6, characterized in that, The trigger condition for industrial equipment sampling should at least satisfy that two modalities exceed the set threshold.

9. An intelligent operation and maintenance method for industrial equipment based on dynamic acquisition of multi-source heterogeneous data and LSTM optimization according to claim 7, characterized in that, In step S4, it includes allocating dynamic weights to each sensor through an attention mechanism according to the contribution degrees of the sensors in historical faults, dynamically adjusting the threshold adjustment coefficients of each sensor, and achieving a differentiated sampling frequency, as follows: β i = βω i ; where β i is the threshold adjustment coefficient of the i-th sensor, and ω i represents the weight of the i-th sensor.

10. A method for intelligent operation and maintenance of industrial equipment based on dynamic acquisition of multi-source heterogeneous data and LSTM optimization according to claim 9, characterized in that, In step S4, the prediction ability of sensor data for faults is quantified by calculating the mutual information or correlation coefficient between the data of each sensor and the fault label, and then the mutual information of each sensor is normalized to obtain the weight of the corresponding sensor to measure the contribution degrees of the sensors in historical faults, as follows: where MI(s i , y) represents the mutual information between the data s i of the i-th sensor and the fault label y, and n represents the total number of sensors.

Images