Self-adaptive molten steel underwater slag signal anomaly detection system and method

Abstract

The invention belongs to the technical field of detection, and particularly provides a self-adaptive molten steel underwater slag signal anomaly detection system and method, and the system comprises a data collection module, a data preprocessing module, a deep learning network prediction module, an anomaly detection module and an alarm module. The problems of high cost, low detection precision, poor anti-interference capability and low efficiency of roughing slag detection in the existing steel production process are solved, and the method is low in detection cost, high in detection precision, strong in anti-interference capability and high in efficiency.

Description

Technical Field

[0001] The invention belongs to the field of detection technology, and in particular relates to an adaptive abnormal detection system and method for molten steel slag signal. Background Art

[0002] In the steel production process, slag detection is an important link. Accurate slag detection can significantly improve the quality of steel, reduce production costs, and is also of great significance to environmental protection. Traditional slag detection methods mainly rely on manual observation and experience judgment, which is not only inefficient, but also prone to misjudgment, resulting in reduced production efficiency and waste of resources.

[0003] With the development of intelligent manufacturing technology, automated slag detection technology has gradually received attention. The current automated slag detection methods mainly include detection based on visual recognition, detection based on sound signals, and detection based on vibration signals. However, these methods still have certain limitations in practical applications. Visual recognition methods are easily affected by environmental factors such as light and dust; sound signal detection and vibration signal detection have high requirements on equipment and environment, and the detection accuracy is not high.

[0004] Defects and shortcomings of the existing technology:

[0005] 1. Defects of visual recognition methods:

[0006] Greatly affected by the environment: Visual recognition methods are highly sensitive to lighting conditions and dusty environments, and are prone to misjudgment.

[0007] High cost: High-precision visual recognition equipment is expensive and requires high maintenance costs.

[0008] After the molten steel in the converter is poured out from the outlet, the outlet is blocked by the slag blocking balls in the converter, so that the molten steel can only be poured out from the outlet. When the molten steel is poured out, some residue will flow out. The image of the molten steel pouring is imported into the industrial computer through the infrared imager, and the picture is processed by the industrial computer. By analyzing the picture, the residue in the molten steel is monitored in real time. However, the infrared imager has a single function and cannot filter out the residue in the molten steel after detecting it. The operator needs to operate the equipment separately to clean it, and the operation steps are cumbersome, which leads to the problem of low production efficiency and cannot meet the working requirements of the automatic detection device (see patent 202210146352.2).

[0009] 2. Defects of sound signal detection method:

[0010] An audio sensor is used to collect the corresponding sound signals when molten steel and slag flow out of the water outlet, and convert them into corresponding analog waveforms and output them to the audio signal preprocessing device; the audio signal preprocessing device is used to convert the analog waveforms corresponding to the sound signals of molten steel and slag into corresponding digital signals and output them to the main control device; and the main control device is used to analyze and identify the characteristic sound spectrum of slag from the digital signals corresponding to the sound signals of molten steel and slag, so as to determine whether the molten steel flowing into the ladle contains slag (see patent 201010601407.1).

[0011] Low detection accuracy: The sound signal detection method has a low signal-to-noise ratio and is easily interfered by environmental noise, resulting in inaccurate detection results.

[0012] High environmental requirements: The detection needs to be carried out in a low-noise environment, which limits its scope of application.

[0013] 3. Defects of vibration signal detection method:

[0014] At one end of the operating arm away from the long water nozzle, multiple vibration sensors are installed in different directions on its vertical cross section to collect vibration signals in multiple directions; the signals are amplified, filtered, sampled and A / D converted to form sampling signals; a large number of vibration signals collected are processed offline to form their typical feature vectors, and an effective vector quantization algorithm is used to train code books corresponding to each state as a code book knowledge base for each casting state during online recognition; the sampling signal is analyzed by wavelet packet method to obtain feature vectors, and by searching and calculating code words, it is determined which code book in the code book knowledge base the current state belongs to, thereby providing information on whether slag is to be laid (see patent 200510050311.X).

[0015] Complex equipment: Vibration signal detection requires complex sensors and signal processing equipment, which increases the complexity and cost of the system.

[0016] Low efficiency: The real-time performance of vibration signal detection is poor, which makes it difficult to meet the needs of efficient production. Summary of the invention

[0017] The present invention provides an adaptive molten steel slag signal abnormality detection system and method, aiming to overcome the problems of high cost, low detection accuracy, poor anti-interference ability and low efficiency of slag detection in the steel production process in the prior art.

[0018] To this end, the present invention provides an adaptive molten steel slag signal abnormality detection system, comprising:

[0019] The data acquisition module is used to collect the time series data related to the slag feeding process in real time and transmit the collected time series data to the data preprocessing module;

[0020] A data preprocessing module is used to receive the time series data transmitted by the data preprocessing module, and perform normalization processing after filtering the received time series data;

[0021] A deep learning network prediction module is used to train the deep learning network prediction module using historical normal data, so that it can learn the normal mode in the slag laying process, and output and optimize the prediction results through the normal mode;

[0022] The anomaly detection module is used to input the data preprocessed by the data preprocessing module into the trained deep learning network prediction module in real time, output the prediction result, compare the prediction result with the threshold value, and determine whether there is an abnormal situation;

[0023] The alarm module is used to activate the alarm module when the abnormality detection module detects an abnormal situation, automatically close the water outlet, and remind the operator to take measures.

[0024] Preferably, the data acquisition module includes a plurality of sensors and coils.

[0025] Preferably, the deep learning network prediction module includes an encoder and a decoder, the encoder is used to process the input time series data through a normal mode and output a latent space representation; the decoder is used to use a structure of a linear layer plus a Sigmoid activation function to decode the latent space representation output by the encoder into a final prediction result.

[0026] Preferably, the deep learning network prediction module is trained using formula (1):

[0027]

[0028] in, is the prediction result at time t+k, x t is the input data at time t, θ is the parameter of the model; n and k are both time.

[0029] Preferably, the abnormality detection module includes an error calculation unit, an adaptive threshold judgment unit and an alarm trigger unit; the error calculation unit is used for prediction value processing and error calculation; the adaptive threshold judgment unit is used to set an adaptive threshold, compare the error with the adaptive threshold, and when the error is greater than the adaptive threshold, start the alarm trigger unit; the alarm trigger unit is used to start the alarm module.

[0030] Preferably, the error calculation unit is specifically: the trained deep learning network prediction module outputs multiple prediction values ​​within a time window Where k = 1, 2, ..., N; when calculating the error, the predicted value window is processed, and the processing formula is:

[0031]

[0032] The error calculation formula is:

[0033] Error=∣Processed Prediction-y t ∣ Formula (3)

[0034] Among them, y t is the actual value, and Processed Prediction is the predicted value after processing.

[0035] Preferably, the setting of the adaptive threshold includes adaptive threshold setting based on statistical analysis and adaptive threshold setting based on a sliding window.

[0036] Preferably, the adaptive threshold value based on statistical analysis is set as follows: by analyzing the error distribution in the historical data, calculating the mean μ and standard deviation σ of the error, and setting the adaptive threshold value:

[0037] Threshold=μ+k·σ Formula (4)

[0038] Among them, k is the adjustment coefficient and Threshold is the adaptive threshold.

[0039] Preferably, the adaptive threshold based on the sliding window is set as follows: in the real-time detection process, the sliding window technology is used to calculate the error mean μ in the recent period of time t and standard deviation σ t :

[0040] Threshold t =μ t +k·σ t Formula (5)

[0041] Where t is time, Threshold t is the average adaptive threshold value in the recent period.

[0042] An adaptive method for detecting abnormality of molten steel slag signal comprises the following steps:

[0043] S1, data acquisition module collects time series data related to the slag discharging process in real time;

[0044] S2, the data preprocessing module filters the collected time series data and performs normalization processing;

[0045] S3. Use historical normal data to train the deep learning network prediction module so that it can learn the normal mode in the slag unloading process, output and optimize the prediction results through the normal mode, and complete the training of the deep learning network prediction module;

[0046] S4, input the data processed by the data preprocessing module into the trained deep learning network prediction module in real time, output the prediction result, compare the prediction result with the threshold value, and determine whether there is an abnormal situation;

[0047] S5. When the abnormality detection module detects an abnormal situation, the alarm module is activated to automatically close the water outlet and remind the operator to take measures.

[0048] Beneficial effects of the present invention:

[0049] 1. The adaptive molten steel slag signal anomaly detection system and method provided by the present invention uses historical normal data to train the deep learning network prediction module, inputs the data processed by the data preprocessing module into the trained deep learning network prediction module in real time, outputs the prediction result, compares the prediction result with the threshold, and determines whether there is an abnormal situation. It can effectively identify abnormal situations in the slag process, improve detection accuracy, and reduce costs. The deep learning network prediction module has strong anti-interference ability and can operate stably in a complex and changeable production environment, reducing false alarms and missed alarms; the automated detection method is combined with a real-time response mechanism, which greatly improves the efficiency of slag detection and meets the needs of efficient production.

[0050] 2. The adaptive abnormal slag signal detection system and method provided by the present invention comprises a data acquisition module including a plurality of sensors and coils. Data is collected through the sensors and coils without relying on high-precision visual equipment and complex vibration sensors, thereby significantly reducing the hardware cost of the system and making it more economical in practical applications.

[0051] 3. In the adaptive abnormal detection system and method for molten steel slag signal provided by the present invention, the adaptive threshold judgment unit is used to set the adaptive threshold, and the detection standard can be dynamically adjusted according to changes in the production environment and data characteristics, ensuring high-precision detection performance under different conditions without manual intervention, further improving the intelligence level of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The present invention will be further described in detail below with reference to the accompanying drawings.

[0053] Figure 1 It is a schematic diagram of an adaptive abnormal detection system for molten steel slag signal.

[0054] Figure 2 It is a schematic diagram of the deep learning network prediction module.

[0055] Figure 3 It is a schematic diagram of the anomaly detection module.

[0056] Figure 4 This is a schematic diagram of the alarm module.

[0057] Figure 5 It is a schematic diagram of adaptive threshold setting. DETAILED DESCRIPTION

[0058] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0059] Embodiment 1:

[0060] like Figure 1 As shown, an adaptive abnormal detection system for molten steel slag signal includes:

[0061] The data acquisition module is used to collect the time series data related to the slag feeding process in real time and transmit the collected time series data to the data preprocessing module;

[0062] A data preprocessing module is used to receive the time series data transmitted by the data preprocessing module, and perform normalization processing after filtering the received time series data;

[0063] A deep learning network prediction module is used to train the deep learning network prediction module using historical normal data, so that it can learn the normal mode in the slag laying process, and output and optimize the prediction results through the normal mode;

[0064] The anomaly detection module is used to input the data preprocessed by the data preprocessing module into the trained deep learning network prediction module in real time, output the prediction result, compare the prediction result with the threshold value, and determine whether there is an abnormal situation;

[0065] The alarm module is used to activate the alarm module when the abnormality detection module detects an abnormal situation, automatically close the water outlet, and remind the operator to take measures.

[0066] Specifically, the deep learning network prediction module is trained by using historical normal data, the data processed by the data preprocessing module is input into the trained deep learning network prediction module in real time, the prediction result is output, and the prediction result is compared with the threshold to determine whether there is an abnormal situation. This can effectively identify abnormal situations in the slag laying process, improve detection accuracy, and reduce costs. The deep learning network prediction module has strong anti-interference ability and can operate stably in a complex and changeable production environment, reducing false alarms and missed alarms. The automated detection method combined with the real-time response mechanism has greatly improved the efficiency of slag laying detection and met the needs of efficient production.

[0067] Embodiment 2:

[0068] Based on Example 1, the data acquisition module includes multiple sensors and coils.

[0069] Specifically, the data acquisition module is mainly composed of multiple sensors, which are installed at key locations of the steel production line to collect real-time time series data related to the slag-feeding process, such as temperature, vibration, pressure, etc. In addition, the data acquisition module also integrates the coil detection slag-feeding signal of the front-end production line as one of the important data sources of the system.

[0070] The sensor and coil jointly transmit the collected data to the data preprocessing module in real time to ensure the continuity and integrity of the data, and provide comprehensive data information for accurate detection of the slag-laying process.

[0071] Embodiment 3:

[0072] Based on Example 2, the data preprocessing module includes a data filtering unit and a normalization unit.

[0073] Specifically, the filtering unit is used to filter the data transmitted by the sensor and the coil to remove noise and outliers; the normalization unit is used to normalize the data to eliminate the influence between different dimensions. Data preprocessing cleans and converts the data to meet the input requirements of the deep learning network prediction module, ensuring that the deep learning network prediction module can accurately learn the data features.

[0074] Embodiment 4:

[0075] On the basis of Example 3, Figure 2 As shown, the deep learning network prediction module includes an encoder and a decoder. The encoder is used to process the input time series data through a normal mode and output a latent space representation; the decoder is used to use a structure of a linear layer plus a Sigmoid activation function to decode the latent space representation output by the encoder into a final prediction result.

[0076] Specifically, the deep learning network prediction module consists of an encoder (LSTM) and a decoder (linear layer). The deep learning network prediction module aims to generate future prediction results by learning the patterns of time series data in order to identify possible abnormal situations in advance.

[0077] Encoder settings: The encoder uses a two-layer LSTM network to process the input time series data. The specific settings are as follows:

[0078] Input size: 2 (indicates the feature dimension at each time step, such as slag signal and weight).

[0079] Hidden layer size (hidden_size): 64 (the hidden state size of each layer of LSTM units).

[0080] Number of layers (num_layers): 2 (the encoder consists of 2 layers of LSTM, and the output of each layer serves as the input of the next layer).

[0081] batch_first: True (the input dimension order is [batch, sequence, feature]).

[0082] The role of LSTM is to map the input time series data into a low-dimensional latent space representation. In this process, each layer of LSTM updates its hidden state based on the output of the previous layer and the current input, thereby capturing the temporal dependencies and patterns in the data.

[0083] Latent space representation:

[0084] The latent space representation output by the encoder is a vector of dimension window*hidden_size, where window is the time window length of the input sequence (50 in this case).

[0085] The output of the LSTM encoder is reshaped from a three-dimensional tensor [batch_size, sequence_length, hidden_size] to a two-dimensional tensor [batch_size, window*hidden_size].

[0086] Decoder settings:

[0087] The decoder uses a linear layer plus a Sigmoid activation function structure to decode the latent space representation output by the encoder into the final prediction result. The specific settings are as follows:

[0088] Input size: window*hidden_size (ie 50*64=3200).

[0089] Output size (out_size): 1 (the prediction result is a scalar).

[0090] Activation function: Sigmoid (used to normalize the output value to the [0,1] interval).

[0091] The deep learning network prediction module generates future prediction results by encoding and decoding the input time series data. During the training process, the parameter θ is optimized by minimizing the error between the predicted value and the actual value. Finally, the deep learning network prediction module can accurately predict the future state, thereby determining whether it is necessary to trigger an alarm and close the water outlet.

[0092] During the training process, the deep learning network prediction module generates accurate future predictions by learning the normal patterns of time series data. Preferably, the deep learning network prediction module is trained using formula (1),

[0093]

[0094] in, is the prediction result at time t+k, x t is the input data at time t, and θ is the parameter of the model.

[0095] Embodiment 5:

[0096] On the basis of Example 4, Figure 3-Figure 5 As shown, the abnormality detection module includes an error calculation unit, an adaptive threshold judgment unit and an alarm trigger unit; the error calculation unit is used for prediction value processing and error calculation; the adaptive threshold judgment unit is used to set an adaptive threshold, compare the error with the adaptive threshold, and when the error is greater than the adaptive threshold, start the alarm trigger unit; the alarm trigger unit is used to start the alarm module.

[0097] Specifically, the anomaly detection module receives the output of the deep learning network prediction module in real time, compares the predicted value within a time window with the threshold calculated dynamically, and determines whether there is an abnormality. Once the error exceeds a certain multiple of the dynamic threshold, the system will immediately trigger an alarm signal to prompt the operator to take necessary measures.

[0098] Preferably, the error calculation unit is specifically: the trained deep learning network prediction module outputs multiple prediction values ​​within a time window Where k = 1, 2, ..., N; when calculating the error, the predicted value window is processed, and the processing formula is:

[0099]

[0100] The error calculation formula is:

[0101] Error =∣ Processed Prediction -y t ∣ Formula (3)

[0102] Among them, y t is the actual value, and Processed Prediction is the predicted value after processing.

[0103] Preferably, the setting of the adaptive threshold includes adaptive threshold setting based on statistical analysis and adaptive threshold setting based on a sliding window.

[0104] Preferably, the adaptive threshold value based on statistical analysis is set as follows: by analyzing the error distribution in the historical data, calculating the mean μ and standard deviation σ of the error, and setting the adaptive threshold value:

[0105] Threshold=μ+k·σ Formula (4)

[0106] Among them, k is the adjustment coefficient used to control the sensitivity of the threshold; Threshold is the adaptive threshold.

[0107] Preferably, the adaptive threshold based on the sliding window is set as follows: in the real-time detection process, the sliding window technology is used to calculate the error mean μ in the recent period of time t and standard deviation σ t :

[0108] Threshold t =μ t +k·σ t Formula (5)

[0109] Where, t is time; Threshold t It is the average of the adaptive thresholds in the recent period. The threshold will be adjusted over time and is suitable for scenarios with large error fluctuations.

[0110] Embodiment 6:

[0111] Based on Example 5, the alarm module consists of an audible and visual alarm, an information feedback unit and an actuator. The audible and visual alarm is used to sound an alarm, the information feedback unit is used to feed back abnormal information to the operator, and the actuator is used to perform related operations, including automatically closing the water outlet.

[0112] Specifically, when the abnormality detection module determines that an abnormal situation occurs, the alarm trigger unit starts the alarm module, and the alarm module is immediately started according to the set alarm conditions, and an alarm is sounded through the sound and light alarm, and the actuator automatically closes the water outlet. At the same time, the information feedback unit feeds back the abnormal information to the operator so that they can take further processing measures in time.

[0113] Preferably, the alarm module integrates two alarm modes: real-time response and delayed response.

[0114] Specifically, the two alarm modes ensure that abnormal situations can be quickly handled and fed back to operators in a timely manner when they occur.

[0115] Preferably, the alarm triggering condition is: after receiving the output of the abnormality detection module, the alarm module determines whether to trigger the alarm according to the following conditions:

[0116] (1) Immediate trigger:

[0117] When the calculated error value exceeds the multiple of the dynamic threshold (such as n·Threshold), the alarm is triggered immediately. In this case, the alarm module will respond quickly, close the water outlet and sound and light alarms to inform the operator of the abnormality. Among them, n is a preset multiple parameter that determines the alarm sensitivity of the alarm module.

[0118] if Error>n·Threshold

[0119] (2) Delayed trigger:

[0120] In some cases, the alarm module allows you to set a delay time. If the error exceeds the threshold multiple for the first time, the alarm module performs multiple tests within the set delay time and finds that the error continues to exceed the threshold multiple, then the alarm is triggered. This mode reduces false alarms caused by temporary fluctuations.

[0121] (3) Multiple over-threshold confirmations:

[0122] The alarm module can set a continuous detection strategy, which triggers an alarm only when the error exceeds the threshold multiple times in multiple consecutive detections. This strategy is used to further reduce the false alarm rate.

[0123] Response of the alarm module:

[0124] When any of the above trigger conditions is met, the alarm module starts immediately:

[0125] Sound and light alarm: emits obvious sound and light signals to alert operators to abnormal situations.

[0126] Actuator automatically closes the sprue: In high-risk situations, the actuator will automatically close the sprue to prevent further abnormal spread during slag feeding.

[0127] Information feedback unit: Feedback abnormal information to operators in real time so that they can take timely measures, such as adjusting process parameters or further checking system status.

[0128] Post-alarm measures:

[0129] The system provides two processing strategies: immediate shutdown and delayed shutdown. Users can choose the appropriate strategy based on actual conditions:

[0130] Close the water outlet immediately: Applicable to scenarios where serious abnormalities are detected. The system will immediately perform the shutdown operation when the alarm is triggered.

[0131] Delayed closing of water outlet: Suitable for scenarios where short-term fluctuations are allowed. The system will perform the closing operation according to the set delay time after the alarm is triggered.

[0132] Embodiment 7:

[0133] Based on Example 2, an adaptive method for detecting abnormal slag signal under molten steel comprises the following steps:

[0134] S1, data acquisition module collects time series data related to the slag discharging process in real time;

[0135] S2, the data preprocessing module filters the collected time series data and performs normalization processing;

[0136] S3. Use historical normal data to train the deep learning network prediction module so that it can learn the normal mode in the slag unloading process, output and optimize the prediction results through the normal mode, and complete the training of the deep learning network prediction module;

[0137] S4, input the data processed by the data preprocessing module into the trained deep learning network prediction module in real time, output the prediction result, compare the prediction result with the threshold value, and determine whether there is an abnormal situation;

[0138] S5. When the abnormality detection module detects an abnormal situation, the alarm module is activated to automatically close the water outlet and remind the operator to take measures.

[0139] The present invention has the following technical advantages:

[0140] 1. Adaptive detection, no need to manually set thresholds

[0141] Technical improvement: The present invention uses a deep learning network prediction module for anomaly detection. The model can automatically adapt to the normal mode of the data through unsupervised learning, and automatically determine anomalies based on the comparison of the prediction results with the dynamic threshold, without manually setting a fixed threshold. This adaptive detection method greatly improves the flexibility and accuracy of the system.

[0142] Advantages analysis:

[0143] 1) Reduce human intervention:

[0144] Traditional anomaly detection methods rely on manually set thresholds, which are usually based on experience or test results in a specific environment. However, when the production environment or data characteristics change, the fixed threshold may fail, resulting in a decrease in detection accuracy. At this time, the operator needs to adjust the threshold frequently, which increases the workload and the risk of error. In contrast, the deep learning network prediction module used in the present invention can automatically adjust the detection criteria through adaptive learning without human intervention, thereby avoiding detection errors caused by improper threshold setting.

[0145] 2) Dynamically adjust the threshold:

[0146] The deep learning network prediction module can continuously learn and dynamically adjust its threshold settings according to changes in the production environment and data characteristics. Even when the data characteristics change significantly, the model can automatically adapt and maintain a high level of detection accuracy. This dynamic adjustment mechanism enables the system to operate stably in complex and changing production environments, overcoming the shortcomings of traditional methods where fixed thresholds cannot adapt to changes.

[0147] 3) Enhanced detection flexibility:

[0148] Since the deep learning network prediction module does not rely on fixed threshold settings, but adaptively adjusts the detection criteria according to real-time data and environmental changes, it has greater detection flexibility. Whether it is a slowly changing signal or rapidly fluctuating data, the model can effectively detect abnormal situations. This flexibility enables the present invention to be widely used in different slag detection scenarios, improving the versatility and practicality of the system.

[0149] 4) Improve detection accuracy and robustness:

[0150] By learning the features of a large amount of normal data, the deep learning network prediction module can achieve high-precision anomaly detection. Compared with the traditional fixed threshold method, the adaptive detection method of the present invention can more accurately identify anomalies and reduce the possibility of missed detection and false alarms. At the same time, the dynamically adjusted threshold enhances the robustness of the system, enabling it to cope with various uncertainties and environmental changes, and ensure the stability and reliability of the detection process.

[0151] 2. Improve detection accuracy

[0152] Technical improvements: Use the deep learning network prediction module to detect anomalies in time series data, and train the deep learning network to predict the characteristic patterns of the learning data, so as to accurately predict future values ​​during the detection process.

[0153] Advantages analysis:

[0154] 1) High precision: When training data, the deep learning network can accurately capture the correlation features before and after the time series data, so as to accurately identify any data that deviates from the normal pattern and then identify anomalies. Compared with traditional visual recognition and sound detection methods, this method has higher accuracy and lower error rate.

[0155] 2) Real-time: In actual production, real-time data is input into the model for prediction, which can identify and alarm before abnormal signals occur, ensuring the continuity and safety of production.

[0156] 3. Reduce costs

[0157] Technical improvements: Use deep learning network prediction modules for slag detection without relying on high-precision visual equipment and complex vibration sensors.

[0158] Advantages analysis:

[0159] 1) Low equipment cost: Deep learning network prediction is based on software implementation and only requires conventional computers and data acquisition equipment, avoiding high-cost visual recognition equipment and complex vibration sensors.

[0160] 2) Low maintenance cost: The hardware requirements of the deep learning network prediction system are simple and the maintenance cost is low, which reduces the operating cost of the enterprise.

[0161] 4. Strong anti-interference ability

[0162] Technical improvements: The deep learning network can extract effective features by learning a large amount of time series data and has strong anti-interference ability.

[0163] Advantages analysis:

[0164] 1) Adapt to complex environments: The deep learning network prediction module can work effectively in complex environments and is not affected by external environmental factors such as light and noise. Compared with visual recognition and sound detection methods, it is more stable and reliable.

[0165] 2) Strong robustness: The feature extraction capability of deep learning network prediction enables it to maintain high detection accuracy and stability in different environments and conditions.

[0166] 5. Improve detection efficiency

[0167] Technical improvements: The use of automated detection methods, real-time prediction and processing of time series data through deep learning networks, improves the real-time and efficiency of detection.

[0168] Advantages analysis:

[0169] 1) Efficient production: Automated detection methods can achieve real-time monitoring and rapid response, avoiding the delay of manual detection and improving production efficiency.

[0170] 2) Reduce the manpower burden: reduce the reliance on manual observation and experience judgment, reduce the workload of operators, and improve the level of production automation.

[0171] 6. Simple structure and easy to use

[0172] Technical improvements: The system has a simple structure, is easy to deploy and use, and does not require complicated settings and debugging.

[0173] Advantages analysis:

[0174] 1) Easy to deploy: The hardware requirements of the deep learning network prediction system are simple, requiring only conventional sensors and computing equipment, and the deployment process is quick and convenient.

[0175] 2) User-friendly: The system is easy to operate and operators only need simple training to become proficient in using it, which improves user experience and operational efficiency.

[0176] In the description of the present invention, it should be understood that if there are terms indicating an orientation or positional relationship, it is based on the orientation or positional relationship shown in the drawings, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the terms describing the positional relationship in the drawings are only used for illustrative purposes and cannot be understood as a limitation on the present invention.

[0177] The above examples are merely illustrative of the present invention and do not constitute a limitation on the protection scope of the present invention. All designs that are the same or similar to the present invention fall within the protection scope of the present invention.

Claims

1. An adaptive abnormal detection system for molten steel slag signal, characterized in that: include The data acquisition module is used to collect the time series data related to the slag feeding process in real time and transmit the collected time series data to the data preprocessing module; A data preprocessing module is used to receive the time series data transmitted by the data preprocessing module, and perform normalization processing after filtering the received time series data; A deep learning network prediction module is used to train the deep learning network prediction module using historical normal data, so that it can learn the normal mode in the slag laying process, and output and optimize the prediction results through the normal mode; The anomaly detection module is used to input the data preprocessed by the data preprocessing module into the trained deep learning network prediction module in real time, output the prediction result, compare the prediction result with the threshold value, and determine whether there is an abnormal situation; The alarm module is used to activate the alarm module when the abnormality detection module detects an abnormal situation, automatically close the water outlet, and remind the operator to take measures.

2. The adaptive abnormal detection system for molten steel slag signal according to claim 1, characterized in that: The data acquisition module includes a plurality of sensors and coils.

3. The adaptive abnormal detection system for molten steel slag signal according to claim 1, characterized in that: The deep learning network prediction module includes an encoder and a decoder, the encoder is used to process the input time series data through a normal mode and output a latent space representation; The decoder is used to decode the latent space representation output by the encoder into the final prediction result using a linear layer plus a Sigmoid activation function structure.

4. The adaptive abnormal detection method of molten steel slag signal according to claim 3, characterized in that: The deep learning network prediction module is trained using formula (1). in, is the prediction result at time t+k, x t is the input data at time t, θ is the parameter of the model; n and k are both time.

5. The adaptive abnormal detection system for molten steel slag signal according to claim 1, characterized in that: The abnormality detection module includes an error calculation unit, an adaptive threshold judgment unit and an alarm trigger unit; the error calculation unit is used for prediction value processing and error calculation; The adaptive threshold judgment unit is used to set the adaptive threshold, compare the error with the adaptive threshold, and start the alarm trigger unit when the error is greater than the adaptive threshold; the alarm trigger unit is used to start the alarm module.

6. The adaptive abnormal detection system for molten steel slag signal according to claim 5, characterized in that: The error calculation unit is specifically: the trained deep learning network prediction module outputs multiple prediction values ​​within a time window Where k = 1, 2, ..., N; when calculating the error, the predicted value window is processed, and the processing formula is: The error calculation formula is: Error =∣ Processed Prediction - y t ∣ Equation (3) Among them, y t is the actual value, and Processed Prediction is the predicted value after processing.

7. The adaptive abnormal detection system for molten steel slag signal according to claim 6, characterized in that: The setting of the adaptive threshold includes the setting of the adaptive threshold based on statistical analysis and the setting of the adaptive threshold based on a sliding window.

8. The adaptive abnormal detection system for molten steel slag signal according to claim 7, characterized in that: The adaptive threshold value based on statistical analysis is set as follows: by analyzing the error distribution in the historical data, calculating the mean μ and standard deviation σ of the error, and setting the adaptive threshold value: Threshold=μ+k·σ Formula (4) Among them, k is the adjustment coefficient and Threshold is the adaptive threshold.

9. The adaptive abnormal detection system for molten steel slag signal according to claim 8, characterized in that: The sliding window-based adaptive threshold is set as follows: In the real-time detection process, the sliding window technology is used to calculate the error mean μ in the recent period of time. t and standard deviation σ t : Threshold t = μ t + k·σ t Equation (5) Where t is time, Threshold t is the average adaptive threshold value in the recent period.

10. An adaptive method for detecting abnormality of molten steel slag signal, characterized in that: The steps include: S1, data acquisition module collects time series data related to the slag discharging process in real time; S2, the data preprocessing module filters the collected time series data and performs normalization processing; S3. Use historical normal data to train the deep learning network prediction module so that it can learn the normal mode in the slag unloading process. Through the normal mode, output and optimize the prediction results to complete the training of the deep learning network prediction module. S4, input the data processed by the data preprocessing module into the trained deep learning network prediction module in real time, output the prediction result, compare the prediction result with the threshold value, and determine whether there is an abnormal situation; S5. When the abnormality detection module detects an abnormal situation, the alarm module is activated to automatically close the water outlet and remind the operator to take measures.

Images