A method for generating an aluminum leakage determination model, an aluminum leakage detection method, and a device

Abstract

The application relates to the technical field of aluminum processing supervision, in particular to a generation method of an aluminum leakage judgment model, an aluminum leakage detection method and equipment, which comprises the following steps: acquiring a first acoustic fingerprint signal and noise power spectral density under a casting state without aluminum leakage; acquiring a second acoustic fingerprint signal and environmental sound power spectral density under the casting state with aluminum leakage; calculating a noise suppression gain function; acquiring a training acoustic fingerprint signal and extracting casting acoustic fingerprint features based on the noise suppression gain function; classifying the casting acoustic fingerprint features to acquire non-aluminum leakage feature labels and aluminum leakage feature labels, and training to obtain an aluminum leakage judgment model; collecting real-time acoustic fingerprint signals under the casting state and obtaining real-time acoustic fingerprint features; inputting the real-time acoustic fingerprint features into the aluminum leakage judgment model and waiting for output judgment results; if the judgment results correspond to the aluminum leakage feature labels, it is determined that aluminum leakage is determined and alarm information is output. The application is used for improving aluminum leakage detection effects in a multi-noise environment.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of aluminum processing supervision, in particular to a generation method of an aluminum leakage judgment model, an aluminum leakage detection method and equipment. BACKGROUND

[0002] The aluminum processing industry is a high-risk industry, especially in the smelting workshop, there is often a great risk of explosion, and the safety hazards in the aluminum processing industry are often caused by aluminum leakage. The reasons for the leakage of molten aluminum usually include two factors: one is the leakage that is difficult to avoid in the production process itself, and the other is the leakage caused by improper behavior of workers in production operation.

[0003] Initially, the aluminum processing industry has a low degree of informatization in production safety monitoring, and production safety warning mainly relies on experience, that is, mainly relies on manual monitoring. The safety warning in this way is mainly dominated by experience and seriously depends on human monitoring and judgment.

[0004] In the current technology, with the continuous improvement of computer technology, the construction of workshop safety tends to be intelligent, and the aluminum processing industry has generally realized the man-machine co-monitoring mode. The workers monitor and inspect through experience, and the monitoring equipment determines whether there is aluminum leakage through hearing, such as obtaining the sound signal in the processing process to determine whether there is aluminum leakage sound. However, after the signal emitted by the leakage in the processing process environment sound, it will also contain environmental noise, such as water inlet and outlet noise, equipment noise, etc. The current technology cannot effectively accurately monitor the sound corresponding to the aluminum leakage in the numerous messy voiceprints during the actual sound signal acquisition process.

[0005] Therefore, in the current technology, the detection precision and detection effect of aluminum leakage detection are poor, and the aluminum leakage situation cannot be quickly and accurately detected. SUMMARY

[0006] In order to improve the aluminum leakage detection effect in a noisy environment, the application provides a generation method of an aluminum leakage judgment model, an aluminum leakage detection method and equipment.

[0007] In a first aspect, the application provides a generation method of an aluminum leakage judgment model, which adopts the following technical scheme:

[0008] A generation method of an aluminum leakage judgment model, comprising the following steps:

[0009] Obtain a first voiceprint signal under a casting state without aluminum leakage, and calculate the noise power spectral density corresponding to the first voiceprint signal;

[0010] Obtain a second voiceprint signal under a casting state with aluminum leakage, and calculate the environmental sound power spectral density corresponding to the second voiceprint signal;

[0011] calculating a noise suppression gain function based on the noise power spectral density and the ambient sound power spectral density;

[0012] obtaining a training voiceprint signal, and extracting a cast voiceprint feature in the training voiceprint signal based on the noise suppression gain function;

[0013] classifying the cast voiceprint feature to obtain an un-aluminum-leakage feature label and an aluminum-leakage feature label, and training the cast voiceprint feature based on the un-aluminum-leakage feature label and the aluminum-leakage feature label to obtain an aluminum-leakage determination model.

[0014] In some embodiments thereof, the noise power spectral density corresponding to the first voiceprint signal is calculated by:

[0015] performing Fourier transform on the first voiceprint signal;

[0016] calculating the power spectral density of the first voiceprint signal after Fourier transform and performing smoothing processing to obtain a theoretical noise power spectral density, and the specific formula is as follows:

[0017] ,

[0018] wherein, denotes the theoretical noise power spectral density after smoothing processing, and a denotes a smoothing coefficient, denotes the power spectral density of the frequency domain signal of the first voiceprint signal, and w denotes the frame number of the first voiceprint signal.

[0019] In some embodiments thereof, the noise power spectral density corresponding to the first voiceprint signal is calculated by further comprising the following steps:

[0020] setting the casting state as multi-furnace casting, and collecting the first voiceprint signal for a preset time;

[0021] calculating the noise power spectral density average value of each second after smoothing processing based on the theoretical noise power spectral density in the preset time, and taking the noise power spectral density average value as the noise power spectral density, and the specific formula is as follows:

[0022] ,

[0023] wherein, denotes the noise power spectral density average value, and V denotes the length of the preset time.

[0024] In some embodiments thereof, the ambient sound power spectral density corresponding to the second voiceprint signal is calculated by comprising the following steps:

[0025] performing Fourier transform on the second voiceprint signal to obtain:

[0026] ,

[0027] wherein, characterized as a frequency domain signal corresponding to the second voiceprint signal, characterized as a frequency domain signal corresponding to the aluminum leakage voiceprint signal, characterized as a frequency domain signal of the first voiceprint signal;

[0028] calculating the ambient sound power spectral density corresponding to the second voiceprint signal based on the following formula,

[0029] ,

[0030] .

[0031] In some embodiments thereof, a noise suppression gain function is calculated based on the noise power spectral density and the ambient sound power spectral density, including the following steps:

[0032] The noise suppression gain function is calculated by the following formula,

[0033] .

[0034] In some embodiments thereof, a training voiceprint signal is obtained, and a casting voiceprint feature in the training voiceprint signal is extracted based on the noise suppression gain function, including the following steps:

[0035] performing Fourier transform on the training voiceprint signal and calculating a training power spectral density corresponding thereto;

[0036] multiplying the training power spectral density by the noise suppression gain function to obtain the casting voiceprint feature corresponding to the training voiceprint signal.

[0037] In some embodiments thereof, the casting voiceprint feature is classified to obtain a non-aluminum leakage feature label and an aluminum leakage feature label, including the following steps:

[0038] setting the number of label classifications based on MobileNet2 algorithm to 2;

[0039] setting the non-aluminum leakage feature label corresponding to the casting voiceprint feature without aluminum leakage to 0;

[0040] setting the aluminum leakage feature label corresponding to the casting voiceprint feature with aluminum leakage to 1.

[0041] In a second aspect, the present application provides a method for detecting aluminum leakage, which adopts the following technical solution:

[0042] A method for detecting aluminum leakage, implemented based on an aluminum leakage determination model generated in the method described above, specifically comprising the following steps:

[0043] Collecting a real-time acoustic fingerprint signal in a casting state;

[0044] Calculating a real-time power spectral density corresponding to the real-time acoustic fingerprint signal and multiplying it by a noise suppression gain function to obtain a real-time acoustic fingerprint feature corresponding to the real-time acoustic fingerprint signal;

[0045] Inputting the real-time acoustic fingerprint feature into the aluminum leakage determination model and waiting for an output determination result;

[0046] If the determination result corresponds to a non-aluminum leakage feature label, it is determined that no aluminum leakage has occurred;

[0047] If the determination result corresponds to an aluminum leakage feature label, it is determined to be suspected aluminum leakage, and after the output result is maintained for a preset time, the suspected aluminum leakage is converted into a determined aluminum leakage and an alarm information is output.

[0048] In a third aspect, the present application provides an aluminum leakage detection device, which adopts the following technical solution:

[0049] An aluminum leakage detection device, comprising:

[0050] A hydrophone for collecting a real-time acoustic fingerprint signal in a casting state;

[0051] An acoustic fingerprint feature calculation module for calculating a real-time power spectral density corresponding to the real-time acoustic fingerprint signal and multiplying it by a noise suppression gain function to obtain a real-time acoustic fingerprint feature corresponding to the real-time acoustic fingerprint signal;

[0052] A deep learning module for generating and training an aluminum leakage determination model, and for obtaining the input real-time acoustic fingerprint feature and outputting a determination result;

[0053] An aluminum leakage determination module for detecting aluminum leakage based on the determination result, specifically, if the determination result corresponds to a non-aluminum leakage feature label, it is determined that no aluminum leakage has occurred; if the determination result corresponds to an aluminum leakage feature label, it is determined to be suspected aluminum leakage, and after the output result is maintained for a preset time, the suspected aluminum leakage is converted into a determined aluminum leakage and an alarm information is output.

[0054] In some embodiments, a floating ball device is further included, which comprises a floating ball and a guide pipe, the floating ball and the hydrophone are slidingly arranged in the guide pipe, a lifting ring is arranged at the bottom end of the floating ball, a cable of the hydrophone passes through the lifting ring to achieve connection, and there is a preset length of spacing between the hydrophone and the floating ball.

[0055] Through the technical solutions provided by the embodiments of the present application, the following technical effects exist:

[0056] The suppression gain function is calculated based on the voiceprint signals detected in different aluminum leakage environments (aluminum leakage and no aluminum leakage) to estimate noise and pure voiceprint, and a large number of training voiceprint signals in the training samples are calculated based on the estimation results to extract corresponding casting voiceprint features in different aluminum leakage environments. After the casting voiceprint features are classified, sample sets of aluminum leakage and no aluminum leakage are generated, and a model for detecting aluminum leakage is trained using the sample sets. The deep learning algorithm is combined to realize the estimation and denoising of noise in a multi-noise environment to accurately and quickly detect aluminum leakage. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 is a step flowchart of a method for generating an aluminum leakage determination model provided by the embodiments of the present application.

[0058] Figure 2 is a step flowchart of an aluminum leakage detection method provided by the embodiments of the present application.

[0059] Figure 3 is a module connection diagram of an aluminum leakage detection device provided by the embodiments of the present application.

[0060] Figure 4 is a partial structure diagram of a module of an aluminum leakage detection device provided by the embodiments of the present application.

[0061] Reference signs: 1, hydrophone; 2, floating ball; 3, guide pipe; 4, lifting ring. DETAILED DESCRIPTION

[0062] To more clearly understand the purpose, technical solutions and advantages of the present application, the present application is described and explained below in conjunction with the drawings and embodiments. However, those of ordinary skill in the art should understand that the present application can be implemented without these details. In some cases, in order to avoid unnecessary description, the aspects of the present application become obscure. Well-known methods, processes, systems, components and / or circuits that have been described at a high level will not be described in detail. It is obvious to those of ordinary skill in the art that various changes can be made to the embodiments disclosed in the present application, and the general principles defined in the present application can be applied to other embodiments and application scenarios without deviating from the principles and scope of the present application. Therefore, the present application is not limited to the embodiments shown, but conforms to the broadest scope claimed in the present application.

[0063] It is to be noted that the description of the embodiments is intended for the purpose of aiding understanding of the present application and is not intended to be limited to the application described. Furthermore, the technical features involved in the respective embodiments of the present application described below can be combined with one another unless they contradict each other.

[0064] In the description of the present application, descriptions with reference to the terms "one embodiment", "some embodiments", "exemplary embodiment", "example", "specific example", or "some examples" mean that a specific feature, structure, material or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the present application. Descriptions of the above terms in the specification do not necessarily refer to the same embodiment or example. Also, the specific feature, structure, material, or characteristic described can be combined in any one or more embodiments or examples in a combined manner.

[0065] The embodiment of the present application discloses a method for generating a leak aluminum judgment model.

[0066] As shown in Figure 1 A method for generating a leak aluminum judgment model includes the following steps:

[0067] S100, a first acoustic signal under the casting state without aluminum leakage is obtained, and the noise power spectral density corresponding to the first acoustic signal is calculated.

[0068] First, it is determined that the casting environment is normal casting and there is no aluminum leakage, at this time the first acoustic signal is detected by the hydrophone, when there is no aluminum leakage, the currently detected acoustic signal only has noise in the environment, including machine working noise, environmental noise, etc.

[0069] The noise power spectral density corresponding to the first acoustic signal is calculated.

[0070] S200, a second acoustic signal under the casting state with aluminum leakage is obtained, and the environmental sound power spectral density corresponding to the second acoustic signal is calculated.

[0071] It is determined that the casting environment is normal casting and there is aluminum leakage, at this time the second acoustic signal is detected by the hydrophone 1, when there is aluminum leakage, the detected acoustic signal contains noise and aluminum leakage sound.

[0072] The environmental sound spectral density corresponding to the second acoustic signal is calculated.

[0073] S300, the noise suppression gain function is calculated based on the noise power spectral density and the environmental sound power spectral density.

[0074] The suppression gain function is calculated by the power spectral densities of two different state environments, and is mainly used to estimate the noise in the voiceprint information, and derive the estimation result of the pure voiceprint according to the estimation result of the noise and the real-time received voiceprint signal.

[0075] S400, obtain the training voiceprint signal, and extract the casting voiceprint feature in the training voiceprint signal based on the noise suppression gain function.

[0076] A certain amount of training voiceprint signal is input, which contains voiceprints with and without aluminum leakage, and the pure voiceprint in the voiceprint signal is estimated by the noise suppression gain function to extract the casting voiceprint feature in the training voiceprint signal, which is different in different aluminum leakage scenarios.

[0077] S500, classify the casting voiceprint feature to obtain the aluminum leakage feature label and the non-aluminum leakage feature label, and train the casting voiceprint feature based on the aluminum leakage feature label and the non-aluminum leakage feature label to obtain the aluminum leakage judgment model.

[0078] A large number of casting voiceprint features extracted from the training voiceprint signal are used as training samples, different labels are provided for different voiceprint features of aluminum leakage and non-aluminum leakage, and each label and the corresponding casting voiceprint feature are put into the model for training. The trained model is defined as the aluminum leakage judgment model, which can be input to the real-time acquired voiceprint signal in the subsequent detection process to obtain the corresponding aluminum leakage detection result.

[0079] Through the above method, the suppression gain function is calculated by the voiceprint signal detected in different aluminum leakage environments (aluminum leakage and non-aluminum leakage) to estimate the noise and the pure voiceprint, and a large number of training voiceprint signals in the training samples are calculated based on the estimation result to extract the corresponding casting voiceprint features in different aluminum leakage environments. After classifying these casting voiceprint features, sample sets of aluminum leakage and non-aluminum leakage are generated, and a model for detecting aluminum leakage is trained using the sample sets. The deep learning algorithm is combined to realize the estimation and denoising of noise in a multi-noise environment, so as to accurately and quickly detect aluminum leakage.

[0080] In some other embodiments, the noise power spectral density corresponding to the first voiceprint signal is calculated, including the following steps:

[0081] S110, Fourier transform is performed on the first voiceprint signal.

[0082] The first voiceprint information is collected when the aluminum leakage event does not occur. Since there is no aluminum leakage, the voiceprint signal collected per second at this time is a noise signal, that is:

[0083] .

[0084] characterizes the current collected voiceprint signal, that is, the first voiceprint information in the embodiment of the present application, and characterizes the noise signal.

[0085] After Fourier transform is performed on the first voiceprint signal, the following is obtained:

[0086] .

[0087] is the frequency domain signal of the current collected voiceprint signal, characterizes the frequency domain signal of the noise signal, and w represents the frame number corresponding to the first voiceprint signal (noise signal).

[0088] In S120, the power spectral density of the first voiceprint signal after Fourier transform is calculated and smoothed to obtain the theoretical noise power spectral density.

[0089] The specific formula is as follows:

[0090] ,

[0091] wherein, characterizes the smoothed theoretical noise power spectral density, and a represents a smoothing coefficient, characterizes the power spectral density of the frequency domain signal of the first voiceprint signal, and w represents the frame number of the first voiceprint signal.

[0092] Since the first voiceprint signal is a noise signal when there is no aluminum leakage, therefore = .

[0093] In some other embodiments, the calculation of the noise power spectral density corresponding to the first voiceprint signal further includes the following steps:

[0094] In S130, the casting state is set to multi-furnace casting, and the first voiceprint signal of a preset time is collected.

[0095] After the theoretical noise power spectral density is calculated, it is the power spectral density corresponding to the first voiceprint signal in an ideal case, and in order to improve the accuracy of subsequent training and calculation, the mean value of the noise power spectral density corresponding to multi-furnace casting also needs to be calculated.

[0096] The multi-furnace casting in the embodiment of the present application is specifically 10 furnaces.

[0097] In S140, the smoothed noise power spectral density mean value of each second in the preset time is calculated based on the theoretical noise power spectral density, and the noise power spectral density mean value is taken as the noise power spectral density.

[0098] The specific formula is as follows:

[0099] ,

[0100] Wherein, is represented as the average of the noise power spectrum density, and V is represented as the length of the preset time.

[0101] The calculated average of the noise power spectrum density is calculated on the basis of the relatively ideal theoretical noise power spectrum density in combination with the actual casting environment, and therefore has relatively good accuracy and universality, and therefore the average of the noise power spectrum density is taken as the final noise power spectrum density.

[0102] In other embodiments, the environmental sound power spectrum density corresponding to the second voiceprint signal is calculated, including the following steps:

[0103] S210, performing Fourier transform on the second voiceprint signal to obtain:

[0104] ,

[0105] Wherein, is represented as the frequency domain signal corresponding to the second voiceprint signal, is represented as the frequency domain signal corresponding to the aluminum leakage voiceprint signal, is represented as the frequency domain signal of the first voiceprint signal.

[0106] When the aluminum leakage event occurs, the second voiceprint signal collected contains the aluminum leakage voiceprint and the noise signal, that is,

[0107] .

[0108] The Fourier transform is performed on the second voiceprint signal to obtain the above formula.

[0109] S220, calculating the environmental sound power spectrum density corresponding to the second voiceprint signal based on the following formula,

[0110] ,

[0111] .

[0112] Since the aluminum leakage voiceprint and the noise signal belong to two independent events, the estimated aluminum leakage signal power spectrum density is actually the power spectrum density corresponding to the second voiceprint signal collected at the present moment minus the noise signal power spectrum density.

[0113] In this way, the value of the pure signal (aluminum leakage voiceprint) can be obtained.

[0114] In other embodiments, the noise suppression gain function is calculated based on the noise power spectrum density and the environmental sound power spectrum density, including the following steps:

[0115] S310, calculate the noise suppression gain function by the following formula:

[0116] .

[0117] In other embodiments, the training voiceprint signal is acquired, and the casting voiceprint features in the training voiceprint information are extracted based on the noise suppression gain function, including the following steps:

[0118] S410, perform Fourier transform on the training voiceprint signal and calculate the corresponding training power spectral density.

[0119] S420, multiply the training power spectral density by the noise suppression gain function to obtain the casting voiceprint features corresponding to the training voiceprint signal.

[0120] When the noise suppression gain function in the casting process is calculated, only the power spectral density corresponding to the pre-acquired training voiceprint signal needs to be multiplied by the noise suppression gain function to obtain the casting voiceprint features in the casting state. The casting voiceprint features are mainly used to display the characteristic representation of the aluminum leakage event and the non-aluminum leakage event in a voiceprint signal.

[0121] The specific formula is:

[0122] .

[0123] Y represents the casting voiceprint features, represents the training power spectral density corresponding to the training voiceprint signal.

[0124] When there is no aluminum leakage, because = = , therefore equals 0, and when there is aluminum leakage, does not equal 0, so according to the value of the casting voiceprint features, it can be judged whether the current aluminum leakage has occurred.

[0125] In other embodiments, the casting voiceprint features are classified to obtain non-aluminum leakage feature labels and aluminum leakage feature labels, including the following steps:

[0126] S510, set the label classification number to 2 based on the MobileNet2 algorithm.

[0127] S520, set the non-aluminum leakage feature label corresponding to the casting voiceprint features without aluminum leakage to 0.

[0128] S530, set the aluminum leakage feature label corresponding to the casting voiceprint features with aluminum leakage to 1.

[0129] After obtaining the non-leakage aluminum feature label and the leakage aluminum feature label, the cast state cast voiceprint features are trained by a MobileNet2 algorithm to obtain a leakage aluminum determination model.

[0130] As shown in Figure 2 The application also discloses a leakage aluminum detection method, which is realized based on the leakage aluminum determination model generated by the method, and specifically includes the following steps:

[0131] S600, collecting a real-time voiceprint signal in a cast state.

[0132] S610, calculating a real-time power spectral density corresponding to the real-time voiceprint signal and multiplying the real-time power spectral density by a noise suppression gain function to obtain a real-time voiceprint feature corresponding to the real-time voiceprint signal.

[0133] S620, inputting the real-time voiceprint feature into the leakage aluminum determination model and waiting for an output determination result.

[0134] S630, if the determination result corresponds to the non-leakage aluminum feature label, it is determined that no leakage aluminum occurs.

[0135] S640, if the determination result corresponds to the leakage aluminum feature label, it is determined that suspected leakage aluminum occurs, and the suspected leakage aluminum is converted into determined leakage aluminum after a preset time of output result maintenance, and an alarm information is output.

[0136] During the operation of the casting system, only the real-time voiceprint signal (which may be leakage aluminum or may be non-leakage aluminum) needs to be collected every second, Fourier transform and power spectral density calculation are performed on the real-time voiceprint signal respectively, and the calculated real-time power spectral density is multiplied by a noise suppression gain function to extract a voiceprint feature corresponding to the real-time voiceprint signal.

[0137] The extracted real-time voiceprint feature is input into the leakage aluminum determination model for determination, if the output result is 0, it corresponds to the non-leakage aluminum feature label, and it is determined that no leakage aluminum occurs; if the output result is 1, it corresponds to the leakage aluminum feature label, and the determination result is set to suspected leakage aluminum, in order to reduce the influence of detection errors and the like on the result, it is necessary to determine whether the result has been maintained for a certain time, when the result has been maintained for a certain time, it can be determined that leakage aluminum occurs, and an alarm is issued.

[0138] In some other embodiments, when detecting leakage aluminum, the following steps can also be included:

[0139] After the system is running, the visual feature acquisition module acquires video information in a working environment, and the visual feature acquisition module is generally a camera.

[0140] Different visual monitoring target areas are set, and the aluminum processing working environment state is determined according to the specific state of the workers and the work objects in the visual monitoring target area and in combination with the corresponding visual processing algorithm.

[0141] Specifically, the indicator light ROI area is set. The ROI area is also called the region of interest, and in machine vision and image processing, it is a region that needs to be processed and is outlined in a box, circle, ellipse, irregular polygon and the like from the processed image.

[0142] In the embodiment of the application, after the specific position of the visual feature acquisition module is determined, the image area collected by the visual feature acquisition module is acquired, and the corresponding indicator light ROI area is set according to the area where the device indicator light is located in the video image. Because the equipment used for casting is a large equipment, the position will not move for a long time, so after the indicator light ROI area is set, it is generally not necessary to revise and change it in the general situation.

[0143] The picture of the video frame is acquired from the video monitoring stream collected by the visual feature acquisition module, and the video frame is cropped. The size of the cropping is the size corresponding to the indicator light ROI area. After cropping, the picture containing the indicator light is obtained, and the MobileNetV2 algorithm is used to perform 2 classification on the indicator light picture. If the classification result is “light on”, it indicates that the indicator light of the equipment is currently in the on state, and if the classification result is “light off”, it indicates that the indicator light of the equipment is currently in the off state.

[0144] The current casting state is determined based on the lighting state, and the casting state is taken as the visual monitoring result. The casting state represents whether the current monitoring environment is in the casting working state. The subsequent visual monitoring and sound monitoring are combined to determine the aluminum leakage, which is based on the current casting state. If the current state is not in the casting state, the aluminum leakage monitoring is not needed.

[0145] The working area to be detected corresponding to the working environment is selected by a person, which is generally the area around the casting disc. The range is defined as the working range of the casting workers and the monitoring range of the personnel for the aluminum leakage.

[0146] The number of classes of the YOLOv5 algorithm is set to 2 to monitor the pedestrians (workers) and safety helmets in the video frame, so as to obtain the first detection box person_bbox containing the workers and the second detection box helmet_bbox containing the safety helmets.

[0147] The first center point corresponding to the first detection frame and the second center point corresponding to the second detection frame are calculated. First, it is judged whether the first center point exists in the work area to be detected. If the first center point exists, it indicates that there is a worker in the work range of the casting. Then, the number of the first center points existing in the work area to be detected is detected, which represents how many workers exist around the casting tray.

[0148] According to the requirement, the number of workers in the work area to be detected is set, for example, to 2. When there are more than two first center points in the work area to be detected, it is considered that the worker is not in the off-duty state in the work range, and when the number of the first center points in the work area to be detected is less than 2, it is considered that the off-duty state exists, and an alarm needs to be given.

[0149] Secondly, when the second center point exists in the first detection frame, it is considered that the worker wears a safety helmet when working, otherwise, if the second center point does not exist in the first detection frame, it is considered that the worker does not wear a safety helmet when working, and an alarm signal needs to be given at this time.

[0150] Through the above monitoring results, the worker behavior is monitored, and when the personnel behavior does not appear the alarm signal, it is considered that there is no aluminum leakage that cannot be monitored by the worker in time, and the aluminum leakage monitoring based on sound monitoring can be performed.

[0151] As shown in Figure 3 The aluminum leakage detection device disclosed in the embodiment of the present application comprises:

[0152] The hydrophone 1 is used to collect real-time voiceprint signals in the casting state.

[0153] The voiceprint feature calculation module is used to calculate the real-time power spectral density corresponding to the real-time voiceprint signal and multiply it by the noise suppression gain function to obtain the real-time voiceprint feature corresponding to the real-time voiceprint signal.

[0154] The deep learning module is used to generate and train the aluminum leakage judgment model, and is also used to obtain the input real-time voiceprint feature and output the judgment result.

[0155] The aluminum leakage judgment module is used to detect the aluminum leakage based on the judgment result. Specifically, if the judgment result corresponds to the no aluminum leakage feature label, it is judged that there is no aluminum leakage; if the judgment result corresponds to the aluminum leakage feature label, it is judged that there is suspected aluminum leakage, and after the output result is maintained for a preset time, the suspected aluminum leakage is converted into determined aluminum leakage and the alarm information is output.

[0156] As shown in Figure 4In some embodiments, the water level sensor 5 is a float ball device, which includes a float ball 2 and a guide pipe 3, the float ball 2 and the hydrophone 1 are slidingly arranged in the guide pipe 3, the bottom end of the float ball 2 is provided with a lifting ring 4, and the cable of the hydrophone 1 passes through the lifting ring 4 to achieve connection, and there is a preset length gap between the hydrophone 1 and the float ball 2.

[0157] Because the distance between the hydrophone 1 and the water surface is different, the size and value of the acoustic fingerprint signal detected by the hydrophone 1 are different. In order to reduce the error of the acoustic fingerprint detection result caused by the change of the distance, the hydrophone 1 needs to maintain a certain distance from the water surface.

[0158] The float ball 2 is a stainless steel hollow float ball, with an outer diameter of 45mm, a length of 150mm, and a load of 15kg. A circular lifting ring 4 is fixed to the bottom end of the float ball 2.

[0159] The guide pipe 3 is a hollow long pipe made of stainless steel. The float ball 2 and the hydrophone 1 are slidingly arranged in the guide pipe 3.

[0160] The signal transmission line of the hydrophone 1 passes through the lifting ring 4 to connect with the float ball 2. The signal transmission line between the float ball 2 and the hydrophone 1 is wrapped with a certain hardness of wrapping material, and the length of the wrapping material is 1m. In this way, the interval distance of 1m between the hydrophone 1 and the float ball 2 can be ensured. The float ball 2 floats on the water surface by buoyancy, so that the hydrophone 1 can always maintain a 1m interval with the water surface, ensuring that the hydrophone 1 can stably collect acoustic fingerprint information during the casting process. The wrapping material can also protect the signal transmission line and fix the hydrophone 1. That is, when the hydrophone 1 moves up and down with the water level, it will not move left and right in the guide pipe 3.

[0161] The implementation principle is as follows:

[0162] The acoustic fingerprint signals detected under different aluminum leakage environments (aluminum leakage and no aluminum leakage) are used to calculate the suppression gain function to estimate noise and pure acoustic fingerprint. Based on the estimation result, a large number of training acoustic fingerprint signals in the training sample are calculated to extract the corresponding casting acoustic fingerprint features in different aluminum leakage environments. After classifying these casting acoustic fingerprint features, sample sets of aluminum leakage and no aluminum leakage are generated, and a model for detecting aluminum leakage is trained using the sample sets. The deep learning algorithm is combined to estimate and denoise the noise in the multi-noise environment to accurately and quickly detect the aluminum leakage.

[0163] It should be understood that although each step in the flowchart of the accompanying drawings is shown in sequence according to the direction of the arrow, these steps are not necessarily executed in sequence according to the direction of the arrow. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and they can be executed in other sequences.

[0164] The above are all preferred embodiments of the present application, and do not limit the protection scope of the present application, so that: all equivalent changes made according to the structure, shape, principle of the present application should be covered in the protection scope of the present application.

Claims

1. A method for generating an aluminum leakage determination model, characterized in that: The following steps are involved: Acquire a first soundprint signal in a casting state without aluminum leakage, and calculate the noise power spectrum density corresponding to the first soundprint signal; Acquire a second soundprint signal of aluminum leakage in a casting state, and calculate the ambient sound power spectrum density corresponding to the second soundprint signal; Calculating a noise suppression gain function based on the noise power spectrum density and the ambient sound power spectrum density; Acquire a training voiceprint signal, and extract the casting voiceprint feature in the training voiceprint signal based on the noise suppression gain function. Specifically, Performing Fourier transform on the training voiceprint signal and calculating its corresponding training power spectrum density; Multiplying the training power spectrum density by the noise suppression gain function to obtain the casting voiceprint feature corresponding to the training voiceprint signal, wherein the casting voiceprint feature is used to display the characteristic representation of aluminum leakage events and non-aluminum leakage events in a voiceprint signal; The casting voiceprint features are classified to obtain non-aluminum leakage feature labels and aluminum leakage feature labels, and the casting voiceprint features are trained based on the non-aluminum leakage feature labels and the aluminum leakage feature labels to obtain an aluminum leakage determination model.

2. The method for generating an aluminum leakage determination model according to claim 1, characterized in that: Calculating the noise power spectral density corresponding to the first voiceprint signal includes the following steps: performing Fourier transform on the first voiceprint signal; The power spectrum density of the first voiceprint signal after Fourier transformation is calculated and smoothed to obtain the theoretical noise power spectrum density. The specific formula is as follows: , in, It is represented by the theoretical noise power spectrum density after smoothing, a is represented by the smoothing coefficient, It is represented by the power spectrum density of the frequency domain signal of the first voiceprint signal, and w is represented by the number of frames of the first voiceprint signal.

3. The method for generating an aluminum leakage determination model according to claim 2, wherein: Calculating the noise power spectral density corresponding to the first voiceprint signal further includes the following steps: Setting the casting state to multi-furnace casting, and collecting the first voiceprint signal at a preset time; The smoothed noise power spectrum density average value per second in the preset time is calculated based on the theoretical noise power spectrum density, and the noise power spectrum density average value is used as the noise power spectrum density. The specific formula is as follows: ， in, It is represented by the mean value of the noise power spectrum density, and V is represented by the length of the preset time.

4. The method for generating an aluminum leakage determination model according to claim 3, wherein: Calculating the ambient sound power spectrum density corresponding to the second voiceprint signal includes the following steps: Perform Fourier transform on the second voiceprint signal to obtain: ， in, Characterized as the frequency domain signal corresponding to the second voiceprint signal, It is characterized by the frequency domain signal corresponding to the aluminum leakage soundprint signal. a frequency domain signal characterized as the first voiceprint signal; The ambient sound power spectrum density corresponding to the second voiceprint signal is calculated based on the following formula: ， 。 5. The method for generating an aluminum leakage determination model according to claim 4, characterized in that: Calculating a noise suppression gain function based on the noise power spectrum density and the ambient sound power spectrum density comprises the following steps: The noise suppression gain function is calculated by the following formula, 。 6. The method for generating an aluminum leakage determination model according to claim 1, wherein: Classifying the casting voiceprint features to obtain non-aluminum leakage feature labels and aluminum leakage feature labels includes the following steps: Set the number of label classifications to 2 based on the MobileNet2 algorithm; The non-aluminum leakage feature label corresponding to the casting voiceprint feature where no aluminum leakage occurs is set to 0; The aluminum leakage feature label corresponding to the casting voiceprint feature where aluminum leakage occurs is set to 1.

7. A method for detecting aluminum leakage, characterized in that: The aluminum leakage determination model generated based on the method according to any one of claims 1 to 6 is implemented, specifically comprising the following steps: Collect real-time voiceprint signals in the casting state; Calculating the real-time power spectrum density corresponding to the real-time voiceprint signal and multiplying it by the noise suppression gain function to obtain the real-time voiceprint feature corresponding to the real-time voiceprint signal; Inputting the real-time voiceprint feature into the aluminum leakage determination model and waiting for output of the determination result; If the determination result corresponds to the non-aluminum leakage characteristic label, it is determined that no aluminum leakage occurs; If the determination result corresponds to an aluminum leakage feature tag, it is determined to be suspected aluminum leakage, and after the determination result is maintained for a preset time, the suspected aluminum leakage is converted into confirmed aluminum leakage and an alarm message is output.

8. An aluminum leakage detection device, characterized in that: The aluminum leakage determination model generated based on the method according to any one of claims 1 to 6 is implemented, comprising: A hydrophone (1) for collecting real-time soundprint signals in a casting state; a voiceprint feature calculation module, configured to calculate the real-time power spectrum density corresponding to the real-time voiceprint signal and multiply the power spectrum density by the noise suppression gain function to obtain the real-time voiceprint feature corresponding to the real-time voiceprint signal; A deep learning module, used to generate and train an aluminum leakage determination model, and also used to obtain the input real-time voiceprint features and output a determination result; The aluminum leakage determination module is used to detect aluminum leakage based on the determination result. Specifically, if the determination result corresponds to a non-aluminum leakage feature label, it is determined that no aluminum leakage has occurred; if the determination result corresponds to an aluminum leakage feature label, it is determined to be suspected aluminum leakage, and after the determination result is maintained for a preset time, the suspected aluminum leakage is converted into confirmed aluminum leakage and an alarm message is output.

9. The aluminum leakage detection equipment according to claim 8, characterized in that: The invention also includes a float device, which includes a float (2) and a guide tube (3). The float (2) and the hydrophone (1) are slidably arranged in the guide tube (3). A hanging ring (4) is provided at the bottom end of the float (2). The cable of the hydrophone (1) passes through the hanging ring (4) to achieve connection, and a preset length of interval exists between the hydrophone (1) and the float (2).

Images