Meat production line real-time quality monitoring method based on multi-sensor fusion

Abstract

The invention discloses a meat production line real-time quality monitoring method based on multi-sensor fusion. The method comprises the following steps: arranging multiple types of sensors, collecting multi-source original data, executing synchronous preprocessing, and generating a fusion feature vector; constructing an inversion reachable domain based on the fusion feature vector in combination with technological parameter constraints, and judging the quality legality; fusing the features and the process identifiers to construct a directed process diagram, forming a quality evolution path, and executing process consistency judgment; setting an excitation mark based on a process switching time point, collecting time window data before and after, and generating dynamic change information; constructing a mass resonance detector, performing multi-scale response decomposition, and extracting resonance characteristics; and fusing the three judgment results, generating a final quality judgment, and outputting a monitoring record, alarm information or a linkage control instruction. According to the invention, through multi-sensor fusion and process evolution dynamic analysis, real-time and reliable monitoring of the quality state of the meat production line is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of food safety detection and intelligent manufacturing, and particularly relates to a real-time quality monitoring method for a meat production line based on multi-sensor fusion. BACKGROUND

[0002] With the continuous improvement of the scale and automation level of the meat processing industry, higher requirements are put forward for the online monitoring and control of product quality in the meat production line. The existing meat quality detection technology mainly relies on manual sampling inspection, single sensor detection or offline laboratory analysis method to detect the appearance, temperature or microbial indicators of meat. Such methods have long detection cycles and limited coverage, and are difficult to meet the real-time and comprehensive quality monitoring needs of continuous production lines, and the detection results often lag behind the production process, and cannot intervene in abnormal situations in time.

[0003] To improve detection efficiency, some existing technologies begin to introduce multiple sensors to collect the state of the meat production process, such as using optical sensors to obtain surface characteristics, using temperature sensors to monitor the cooling process, or using gas sensors to detect volatile substances. However, most of such technologies still use static threshold judgment or single-process independent analysis methods, and lack systematic modeling of the correlation between different sensor data, which cannot determine whether the current meat quality state is truly derived from a legal production process, and also cannot identify hidden quality abnormalities that accumulate gradually in multiple processing procedures.

[0004] Existing technologies generally ignore the dynamic response characteristics of meat during the switching process of the production line. The switching of processing procedures is often accompanied by changes in temperature, environment and processing conditions, and the quality characteristics of meat will show a dynamic change rule with time sequence structure, while the existing detection methods are mostly based on single-time or short-time average for judgment, lacking analysis means for the consistency of quality evolution under process excitation, resulting in insufficient recognition ability for unknown pollution, process deviation or surface normal but internal abnormality.

[0005] Therefore, how to provide a real-time quality monitoring method for a meat production line based on multi-sensor fusion is a problem that those skilled in the art need to solve. SUMMARY

[0006] One purpose of the present application is to propose a meat production line real-time quality monitoring method based on multi-sensor fusion, which comprehensively utilizes multi-type sensor data fusion, production process information correlation analysis and dynamic response consistency detection technology to continuously monitor and comprehensively judge the quality state of meat in each processing procedure of the production line, realizes multi-dimensional verification of meat quality legitimacy, process consistency and process switching dynamic response by constructing quality inversion reachable domain, cross-process quality evolution path and quality evolution resonance detection mechanism, and has the advantages of strong monitoring real-time, high abnormality recognition ability, good adaptability to unknown pollution and process deviation and high traceability of quality judgment results.

[0007] The meat production line real-time quality monitoring method based on multi-sensor fusion according to the embodiment of the present application comprises: A plurality of types of sensors are arranged on the meat production line to collect multi-source sensor original data, the multi-source sensor original data is preprocessed to obtain a fusion quality feature vector of the meat sample; Based on the fusion quality feature vector, an inversion reachable domain of the meat quality state is constructed, it is judged whether the quality state of the current meat sample falls within the inversion reachable domain, and a quality legitimacy judgment result is obtained; Based on the fusion quality feature vector and corresponding process identification information, each processing procedure of the meat production line is constructed as a directed process graph to form a cross-process quality evolution path, a process quality conservation loop consistency judgment is performed, and a process consistency judgment result is obtained; Based on the process identification information, a process switching time point is taken as a process excitation marker, the fusion quality feature vector is continuously sampled within a preset time window before and after the process excitation marker, and quality feature dynamic change data are obtained; A quality evolution resonance detector is constructed, the quality feature dynamic change data are subjected to multi-scale process response decomposition, a quality evolution resonance feature is constructed, the quality evolution resonance feature is subjected to structure matching checking and timing consistency checking with a process response reference template, and a dynamic response consistency judgment result is outputted; Based on the quality legitimacy judgment result, the process consistency judgment result and the dynamic response consistency judgment result, a fusion judgment is performed on the quality state of the meat sample to generate a final quality judgment result, and a quality monitoring record, an abnormality alarm signal or a production line linkage control instruction is outputted.

[0008] Optionally, the plurality of types of sensors comprise optical sensors, temperature sensors, gas sensors and environmental parameter sensors.

[0009] Optionally, the multi-source sensor raw data comprises optical image data of the meat sample, temperature data of the meat sample and the environment, volatile gas concentration data related to the meat sample, and environmental parameter data reflecting the state of the meat production environment, the environmental parameter data comprising humidity data and air flow state data.

[0010] Optionally, the pre-processing of the multi-source sensor raw data to obtain the fusion quality feature vector of the meat sample comprises: performing denoising processing on the multi-source sensor raw data respectively to eliminate random noise and abnormal interference generated in the collection process; performing time alignment processing on the multi-source sensor raw data after denoising processing to make the data collected by different types of sensors correspond under the same time reference; performing scale normalization processing on the data after time alignment processing, and based on the multi-source sensor data after denoising processing, time alignment processing and scale normalization processing, performing feature combination and fusion to form a fusion quality feature vector representing the comprehensive quality state of the meat sample.

[0011] Optionally, the obtaining of the quality legality determination result comprises: based on the fusion quality feature vector and the corresponding process identifier, establishing a legal process parameter range list, dividing the range of temperature, processing time, wind speed, relative humidity, spray flow, cooling medium temperature, conveying belt line speed and process residence time of each processing process item by item and setting a discrete step length to form a process parameter library; generating a physical constraint list according to the physical change constraint conditions in the meat processing process, the physical constraint list comprising consistency constraints of optical absorption and water content, monotonicity constraints of temperature change and cooling intensity, time sequence constraints of volatile gas release and residence time, and coupling consistency constraints between the three types of constraints, which are quantitatively described in the form of lower and upper interval and allowed deviation band; deducing the reachable upper bound, the reachable lower bound and the allowed deviation band of each component of the fusion quality feature vector by segmenting the interval endpoints and representative mid-values of the process parameter library, combining the physical constraint list, constructing the inverse reachable domain, the inverse reachable domain being composed of a kernel area, a buffer area and an outer envelope area, the kernel area corresponding to the inevitable combination, the buffer area corresponding to the combination that needs to be further checked, and the outer envelope area corresponding to the unreachable combination; comparing the current fusion quality feature vector with the inverse reachable domain, if all components fall into the kernel area, it is determined as reachable, if any component falls into the outer envelope area, it is determined as unreachable, if there is a component falling into the buffer area, dynamic consistency checking and process order checking are performed based on the near neighbor time window data and the data of the adjacent samples of the same batch, the final determination of reachable or unreachable is made, the quality legality determination result is output and the determination basis identifier is recorded.

[0012] Optionally, the obtaining the process consistency determination result comprises: Based on the process identifier and the timestamp, connecting the positions of the same meat sample at each process in the production line process sequence to form an original process path of the meat sample, taking the in-line process as the starting anchor point and the out-line process as the terminal anchor point; Based on the fused quality feature vector, generating a quality increment description item between adjacent processes in the original process path for each process pair, the quality increment description item including change direction, change amplitude, change rate and residence time, and being associated with the corresponding process pair; Generating one or more closed conservation loops between the starting anchor point and the terminal anchor point according to the process segmentation rule, the closed conservation loop taking heat-related dimension, water content state dimension and volatile precursor-related dimension as conservation dimension, inserting a loop segment or cross-segment connection in the original process path for the sample with rework or skip station to form a set of cross-process quality evolution paths; Performing conservation loop consistency determination on the set of quality evolution paths, using a two-level process of fixed tolerance band checking and adaptive tolerance band checking, wherein: The fixed tolerance band checking performs closedness check on each conservation dimension based on the process preset tolerance; The adaptive tolerance band checking forms a batch reference band based on the distribution range of the same batch of samples and performs closedness check on each conservation dimension; According to the process capability index, the quality increment description item is weighted and summarized to generate a closedness deviation quantization index and a suspected violation process list; Based on the closedness deviation quantization index and the suspected violation process list, a process consistency determination result is generated and output.

[0013] Optionally, the obtaining the quality feature dynamic change data comprises: Based on the process identifier information and the timestamp, obtaining the process switching time point and the process parameter change information corresponding to the meat sample from the production line control system, and recording the process switching time point as a process excitation marker; For each process excitation marker, set a first time window length and a second time window length, the first time window is located before the process excitation marker, and the second time window is located after the process excitation marker; In the first time window and the second time window, based on the fused quality feature vector, continuous sampling is performed at a preset sampling period to generate a pre-sampling sequence and a post-sampling sequence corresponding to the process excitation marker; The pre-sampling sequence and the post-sampling sequence are time-indexed, missing sampling is filled, and abnormal points are removed, maintaining a one-to-one correspondence with the sample identifier and the process identifier, forming quality feature dynamic change data; Packaging the process excitation mark, process parameter change information and quality feature dynamic change data into a process excitation data packet and outputting.

[0014] Optionally, the output dynamic response consistency determination result comprises: A quality evolution resonance detector is constructed, which is composed of an excitation analysis layer, a response decomposition layer and a structure checking layer. The excitation analysis layer receives the process excitation data packet and generates a standardized excitation sequence. The response decomposition layer receives the quality feature dynamic change data. The structure checking layer is used for matching and checking with the process response reference template library. The excitation analysis layer analyzes the process switching time point and the process parameter change information, generates an excitation index sequence corresponding to the sample identification, process identification and timestamp one by one, and establishes a scale index according to the process type and process parameter interval; The response decomposition layer performs multi-scale process response decomposition on the quality feature dynamic change data according to the excitation index sequence and the scale index, to obtain a response component set corresponding to different excitation intensity and duration. In each response component, the response starting time, peak time, decay termination time and cross-feature synchronization point are labeled. The response decomposition layer extracts quality evolution resonance features from the response component set, which include starting delay, peak sequence, peak amplitude interval label, decay duration interval length, cross-feature synchronization relationship table and stable segment proportion. The quality evolution resonance features are encoded according to the process type, process parameter interval and scale index, and the association with the sample identification, process identification and timestamp is maintained. The structure checking layer checks the quality evolution resonance features with the process response reference template library, and sequentially performs structure matching checking, timing consistency checking and cross-feature coupling consistency checking to generate a dynamic response consistency determination result.

[0015] Optionally, the multi-scale process response decomposition on the quality feature dynamic change data comprises: The basic scale, transient scale, delay scale and residual vibration scale are determined according to the process type, process parameter interval, conveyor belt line speed and process residence time, and the window length and step size are set for each scale. An anchor point slice is performed based on the process switching time point in the quality feature dynamic change data. Time index alignment, drift correction and missing data completion are completed by constructing the front and rear time windows on each scale to obtain the aligned data segments of each scale. The rising segment, peak segment, decay segment and stable segment are divided in the data segments of each scale, and the starting delay, peak sequence, decay duration interval and cross-feature synchronization relationship are labeled to generate the primary response component. The primary response components are structurally matched and temporally consistent with the process response reference template. Components that are not related to the process are removed, adjacent components of the same scale are merged, and the scale index, process index, intensity and duration are retained to output the response component set.

[0016] Optionally, the output quality monitoring records, abnormal alarm signals, or production line linkage control commands include: Receive the quality legality judgment results, process consistency judgment results, and dynamic response consistency judgment results, perform unified formatting on the judgment results, and form a set of judgment results that correspond one-to-one with the sample identifier, process identifier, and timestamp. A fusion weight and a judgment priority rule are set for the quality legality judgment result, the process consistency judgment result, and the dynamic response consistency judgment result, respectively. The fusion weight is used to reflect the degree of influence of each type of judgment result in the final quality judgment, and the judgment priority rule is used to determine the coverage order when there is a conflict in the judgment results. Based on the fusion weight, each judgment result is weighted and fused to generate a comprehensive judgment index. The consistency of each judgment result is checked in combination with the judgment priority rule to form the final quality judgment result. At the same time, the corresponding abnormality type identifier and suspected abnormal process list are generated. Output the final quality judgment result, record and store the final quality judgment result, the abnormality type identifier, the list of suspected abnormal processes and the corresponding judgment basis identifier, generate a quality monitoring record, and output the corresponding linkage control command to the production line control system when the final quality judgment result meets the preset abnormality triggering conditions.

[0017] The beneficial effects of this invention are: This invention introduces a multi-sensor fusion and process correlation analysis method into meat production lines, enabling real-time and continuous monitoring of meat quality status. Compared with existing technologies that rely on single sensors or static threshold judgments, this invention can comprehensively reflect the multi-dimensional quality characteristic changes of meat during processing, effectively improving the coverage and timeliness of quality monitoring. This allows the production line to obtain reliable quality status information in a timely manner during processing, reducing quality risks caused by detection delays.

[0018] This invention constructs an inversion reachable domain for meat quality status and a cross-process quality evolution path to determine whether meat quality truly originates from legitimate production processes and to identify hidden anomalies accumulated over multiple processes. This overcomes the shortcomings of existing technologies that struggle to identify surface-normal but inconsistent processes. From the perspectives of process consistency and the rationality of quality evolution, it improves the accuracy and reliability of quality judgment and effectively reduces the false positive and false negative rates.

[0019] This invention introduces a quality evolution resonance detector to analyze the dynamic response of quality characteristics during process switching, which makes up for the lack of monitoring capability of existing technology for dynamic quality changes under process excitation, enhances the ability to identify unknown contamination, process deviation and abnormal processing state, and improves the traceability of quality monitoring results and the level of automation control of production process by combining the recording of judgment results with linkage control mechanism. It has high practical value and promotion significance. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the real-time quality monitoring method for meat production lines based on multi-sensor fusion proposed in this invention; Figure 2 This is a schematic diagram illustrating the generation and inversion reachability domain construction process of the fusion quality feature vector of the real-time quality monitoring method for meat production lines based on multi-sensor fusion proposed in this invention. Figure 3 This is a schematic diagram of the structural composition and multi-scale process response decomposition of the quality evolution resonance detector in the real-time quality monitoring method for meat production lines based on multi-sensor fusion proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figure 1 , Figure 2 and Figure 3 A real-time quality monitoring method for meat production lines based on multi-sensor fusion includes: Multiple types of sensors are installed in the meat production line to collect raw data from multiple sources. The raw data from multiple sources is preprocessed to obtain the fusion quality feature vector of the meat sample. Based on the fused quality feature vector, the inversion reachability domain of meat quality status is constructed. It is then determined whether the quality status of the current meat sample falls within the inversion reachability domain to obtain the quality legality judgment result. Based on the fusion of quality feature vectors and corresponding process identification information, each processing process of the meat production line is constructed into a directed process graph, forming a cross-process quality evolution path, performing process quality conservation cycle consistency judgment, and obtaining process consistency judgment results. Based on the process identification information, the process switching time point is used as the process excitation mark. Within a preset time window before and after the process excitation mark, the fused quality feature vector is continuously sampled to obtain dynamic change data of quality features. A quality evolution resonance detector is constructed, and multi-scale process response decomposition is performed on the dynamic change data of quality characteristics. Quality evolution resonance features are constructed, and structural matching and temporal consistency checks are performed between the quality evolution resonance features and the process response reference template. The dynamic response consistency judgment result is output. Based on the results of quality legality assessment, process consistency assessment, and dynamic response consistency assessment, the quality status of meat samples is integrated and assessed to generate the final quality assessment result, and output quality monitoring records, abnormal alarm signals, or production line linkage control instructions.

[0023] In this embodiment, the multiple types of sensors include optical sensors, temperature sensors, gas sensors, and environmental parameter sensors.

[0024] In this embodiment, the raw data from the multi-source sensor includes optical image data of the meat sample, temperature data of the meat sample and its surrounding environment, volatile gas concentration data related to the meat sample, and environmental parameter data reflecting the state of the meat production environment. The environmental parameter data includes humidity data and airflow status data.

[0025] In this embodiment, the preprocessing of the raw data from the multi-source sensors to obtain the fused quality feature vector of the meat sample includes: The raw data from the multi-source sensors are subjected to noise reduction processing to eliminate random noise and abnormal interference generated during the acquisition process. Time alignment processing is performed on the denoised raw data from multiple sensors to ensure that data collected by different types of sensors correspond to each other under the same time reference. The data that has undergone time alignment processing is subjected to scale normalization processing. Based on the multi-source sensor data after denoising, time alignment, and scale normalization processing, feature combination and fusion are performed to form a fused quality feature vector that characterizes the overall quality status of meat samples.

[0026] In this embodiment, obtaining the quality legality determination result includes: Based on the fusion quality feature vector and the corresponding process identifier, a list of legal process parameter ranges is established. For each processing step, the temperature, processing time, wind speed, relative humidity, spray flow rate, cooling medium temperature, conveyor belt speed and process dwell time are divided into ranges and the deviation length is set to form a process parameter library. Based on the physical constraints during meat processing, a list of physical constraints is generated. This list includes constraints on the consistency of optical absorption and moisture content, the monotonicity of temperature changes and cooling intensity, the temporal constraints on volatile gas release and residence time, and the coupling consistency constraints among these three types of constraints. These constraints are quantified in the form of upper and lower bounds and allowable deviation bands. Physical change constraints include: constraints on maintaining a consistent evolutionary relationship between the surface optical absorption characteristics of meat and changes in tissue water content during processing; The constraint that the temperature change of meat during the pre-cooling and cooling processes exhibits a monotonic relationship with increasing cooling intensity and cooling time; Constraints that establish a clear temporal correlation between the characteristics of volatile gas release from meat during processing and handling and the dwell time in each process; Coupling consistency constraints refer to the coordinated and non-contradictory joint changes among various physical characteristics related to meat quality in the same or adjacent processing steps of meat processing. They are used to constrain the consistency of optical absorption characteristics, changes in water content, temperature changes, and volatile gas release characteristics over time and in the process evolution. Based on the interval endpoints and representative medians of the process parameter library, a segmented deduction is performed. Combined with the physical constraint list, upper and lower reachable bounds and allowable deviation bands are generated for each component of the fused quality feature vector, constructing an inversion reachable domain. This inversion reachable domain consists of a kernel region, a buffer region, and an outer envelope region. The kernel region corresponds to combinations that are guaranteed to be reachable, the buffer region corresponds to combinations that require further verification, and the outer envelope region corresponds to unreachable combinations. Specifically, the segmented deduction based on the interval endpoints and representative medians of the process parameter library is as follows: Based on the range of process parameters corresponding to each processing step in the process parameter library, the lower endpoint, upper endpoint, and preset representative median of the range are selected as the deduction benchmark, corresponding to the minimum process conditions, maximum process conditions, and typical process conditions, respectively. According to the sequence of processing steps, taking a single step as the deduction unit, the deduction benchmark points are successively substituted into the physical constraint list to deduce the range of change of each component in the fused quality feature vector under the step, and the component change range corresponding to the step is obtained. After completing the deduction of a single process, the variation range of the component is used as the input boundary condition for the next process, and the process is deduced segment by segment until the complete processing process path is covered. During the step-by-step deduction process, component combinations that simultaneously satisfy all physical constraints are marked as necessarily reachable combinations, component combinations that only satisfy physical constraints under some process conditions are marked as combinations that need further verification, and component combinations that do not satisfy any physical constraints are marked as unreachable combinations, forming an inversion reachable domain that includes the kernel region, buffer region, and outer envelope region. The input to the deduction function includes the processing step identifier, the corresponding process parameter range, representative parameter values, and the current state of the fused quality feature vector. The processing involves segment-by-segment deduction based on the endpoints of the parameter range and the representative median, using the process parameter library. Under the premise of satisfying the physical constraint list, the change direction and change magnitude of each component of the fusion quality feature vector are processed by interval mapping, and the output is the lower reachable bound, upper reachable bound and allowable deviation band of each component of the fusion quality feature vector under the corresponding process, which is used to construct the inversion reachable domain. The current fused quality feature vector is compared with the inverted reachable region. If all components fall into the kernel region, it is determined to be reachable; if any component falls into the outer envelope region, it is determined to be unreachable. If any component falls into the buffer, dynamic consistency verification and process sequence verification are performed based on the data of the nearest time window and the data of adjacent samples in the same batch to make a final determination of reachability or unreachability. The quality legality judgment result is output and the judgment basis identifier is recorded. Specifically, the dynamic consistency verification and process sequence verification based on the data of the nearest time window and the data of adjacent samples in the same batch are performed as follows: Based on the timestamp information assigned to the raw data of multi-source sensors, taking the collection time point corresponding to the current meat sample as the center, the fused quality feature vector sequence continuously collected under the same processing procedure within a preset time range is selected as the nearest neighbor time window data. Based on the recorded sample batch identifier and process identifier, at least one meat sample that is located before and after the current sample in the time sequence and has undergone the same processing process is selected within the same production batch, and the corresponding fused quality feature vector is extracted as the data of adjacent samples in the same batch. Using the nearest neighbor time window data, the continuity of the quality characteristic changes of the current sample in the time dimension is checked to determine whether the trend of change is consistent with the continuous sampling characteristics under the same process. Using the data of adjacent samples in the same batch, the rationality of the quality characteristic evolution of the current sample in the process sequence dimension is checked, and it is determined whether it conforms to the overall change pattern of the batch samples in the cross-process evolution process. Based on the two types of verification results, a final determination of whether it is achievable or unachievable is made.

[0027] In this embodiment, obtaining the process consistency determination result includes: Based on process identifiers and timestamps, the positions of the same meat sample in each process are connected according to the sequence of production line processes to form the original process path of the meat sample, with the entry process as the starting anchor point and the exit process as the ending anchor point, wherein: The process identifier is a unique identifier assigned from the raw data of the multi-source sensors to characterize the current processing stage of the meat sample. The process identifier is used to distinguish different processing stages and indicate the process position of the meat sample in the production line. A timestamp is a time stamp information assigned to the raw data from multiple sensors to characterize the time of data acquisition. The timestamp is used to reflect the temporal sequence of meat samples in each processing step and to support temporal connections across processing paths. Based on the fused quality feature vector, a quality increment description item is generated for each process pair between adjacent processes in the original process path. The quality increment description item includes the direction of change, the magnitude of change, the rate of change, and the dwell time, and is associated with the corresponding process pair. One or more closed conservation loops are generated between the starting and ending anchor points according to the process segmentation rules. These closed conservation loops use heat-related dimensions, water content dimensions, and volatile precursor-related dimensions as conservation dimensions. For samples with rework or skipped stations, loop segments or cross-segment connections are inserted into the original process path to form a set of cross-process quality evolution paths. The process segmentation rules are as follows: Based on the main technological functions of each processing step in the meat production line, adjacent processes that maintain consistency in processing objectives and physical mechanisms are divided into the same process segment. The connection between process segments is established by using the locations where changes occur in the processing temperature range, processing method, or environmental control conditions between adjacent processes as process segment boundaries. When meat samples are reworked or skipped during the production process, corresponding loop segments or cross-segment connections are inserted into the original process path based on the process identifier and timestamp, so that the process segmentation structure can truly reflect the actual processing path of the sample. A conservation cycle consistency check is performed on the set of quality evolution paths, employing a two-level process of fixed tolerance band verification and adaptive tolerance band verification, wherein: The fixed tolerance band verification performs a closure check on each conservation dimension based on the pre-set tolerance of the process, specifically: Based on the process specification parameters, equipment control accuracy, and historical stable operation data corresponding to each processing step, preset tolerances are set for the heat-related dimension, moisture content dimension, and volatile precursor-related dimension. These preset tolerances reflect the allowable range of variation for each conserved dimension under normal process conditions. For heat-related dimensions, the preset tolerance is that the temperature change of meat samples at the start and end nodes of the closed conservation loop does not exceed ±1.0℃~±2.0℃; For the water-bearing state dimension, the preset tolerance is that the variation of the fusion quality characteristic component representing the water-bearing state at the start and end nodes of the closed conservation loop does not exceed the relative variation range of ±3% to ±6%. For the volatile precursor-related dimensions, the preset tolerance is that the change in the volatile precursor-related characteristics at the start and end nodes of the closed conservation loop does not exceed ±10% to ±20% of the relative change range. Using the start and end nodes of the closed conservation loop as comparison nodes, the characteristic changes of the same conservation dimension at the start and end nodes are compared to determine whether the changes fall within the preset tolerance range of the corresponding process. When the changes in all conserved dimensions within the closed conserved loop meet the corresponding process preset tolerance requirements, the closed conserved loop is determined to have passed the closure check. When the change in any conserved dimension exceeds the corresponding process preset tolerance range, the closed conserved loop is determined to have failed the closure check and is marked as a consistency risk. The adaptive tolerance band verification is based on the distribution range of samples in the same batch to form a batch reference band and performs closure checks on each conservation dimension. Specifically: Based on multiple meat samples from the same production batch that have passed quality legality assessment and have complete process paths, the characteristic change distribution of each conservation dimension at the start and end nodes of the corresponding closed conservation loop is statistically analyzed to form a batch reference band that reflects the normal processing fluctuation characteristics of the batch. Using the batch reference band as the adaptive tolerance range, the feature changes of each conservation dimension in the current closed conservation loop at the start node and the end node are compared to determine whether the change falls within the batch reference band of the corresponding conservation dimension. When the changes in all conserved dimensions within the closed conservation loop fall within the corresponding batch reference band, the closed conservation loop is determined to have passed the adaptive tolerance band check. When the change in any conserved dimension exceeds the corresponding batch reference band, the closed conservation loop is determined to have failed the adaptive tolerance band check and is marked as having a batch anomaly risk. The quality increment description items are weighted and summarized based on the process capability index to generate a quantitative index of closure deviation and a list of suspected violation processes. The process capability index is a quantitative index used to characterize the ability of each processing process to control changes in meat quality characteristics under normal production conditions. The process capability index is calculated based on historical stable operation data and reflects the fluctuation level and stability of the corresponding process in the dimensions of heat-related dimensions, moisture content dimensions, and volatile precursor-related dimensions. The higher the process capability index, the stronger the controllability of the process to quality changes and the lower the probability of abnormality. The lower the process capability index, the weaker the stability of the process to quality changes and the greater the contribution weight to closure deviation. Based on the closure deviation quantification index and the list of suspected violation processes, the process consistency judgment result is generated and output.

[0028] In this embodiment, acquiring dynamic change data of quality characteristics includes: Based on process identification information and timestamps, the process switching time points and process parameter change information corresponding to meat samples are obtained from the production line control system, and the process switching time points are recorded as process excitation markers. For each process activation mark, a first time window length and a second time window length are set, with the first time window located before the process activation mark and the second time window located after the process activation mark; Within the first time window and the second time window, continuous sampling is performed based on the fusion quality feature vector at a preset sampling period to generate a pre-sampling sequence and a post-sampling sequence corresponding to the process excitation mark, wherein the preset sampling period is a fixed time interval between 1 second and 5 seconds. The pre-sampled sequence and the post-sampled sequence are aligned by time index, filled with missing samples and removed outliers, and a one-to-one correspondence with the sample identifier and process identifier is maintained to form dynamic change data of quality characteristics. The process excitation markers, process parameter change information, and quality characteristic dynamic change data are packaged into a process excitation data package and output.

[0029] In this embodiment, the output dynamic response consistency determination result includes: A quality evolution resonance detector is constructed, which consists of an excitation parsing layer, a response decomposition layer, and a structure verification layer. The excitation parsing layer receives process excitation data packets and generates standardized excitation sequences. The response decomposition layer receives dynamic change data of quality characteristics. The structure verification layer is used to match and verify the process response reference template library. The formation of the process response reference template library is as follows: Multiple historical meat samples that passed both the quality legality and process consistency assessments were selected and grouped according to process type, process parameter range, and process switching conditions. Multi-scale process response decomposition was performed on the dynamic change data of quality characteristics collected before and after the corresponding process switching for samples in each group, and quality evolution resonance features were extracted. The quality evolution resonance features in the same group were aggregated and screened to retain representative response structures and form process response reference templates for corresponding process types and process parameter ranges. The excitation parsing layer parses the process switching time point and process parameter change information, generates an excitation index sequence that corresponds one-to-one with the sample identifier, process identifier and timestamp, and establishes a scale index according to process type and process parameter range; The response decomposition layer performs multi-scale process response decomposition on the dynamic change data of quality characteristics based on the excitation index sequence and scale index, and obtains a set of response components corresponding to different excitation intensities and durations. The response start time, peak time, decay termination time and cross-feature synchronization point are marked in each response component. The response decomposition layer extracts quality evolution resonance features from the response component set. The quality evolution resonance features include start delay, peak order, peak amplitude interval label, attenuation duration interval length, cross-feature synchronization relationship table, and stable segment ratio. The quality evolution resonance features are encoded according to process type, process parameter interval, and scale index, and are associated with sample identifier, process identifier, and timestamp. The structural verification layer verifies the quality evolution resonance characteristics against the process response reference template library, sequentially performing structural matching verification, temporal consistency verification, and cross-feature coupling consistency verification to generate a dynamic response consistency judgment result, wherein: Structural matching verification refers to comparing the real-time extracted mass evolution resonance features with the corresponding template's response start, peak, decay, and stable segment structural labels at multiple scales to check whether the overall shape and segment order are consistent. Timing consistency verification refers to aligning the start time, duration and interval of each structural label with the time axis after structural matching is completed, to confirm that each timing parameter falls within the allowable deviation range of the template. Cross-feature coupling consistency verification refers to the joint comparison of the synchronization relationship, peak relative order and attenuation synchronization coefficient between different quality features to check whether the coupling mode between each feature is consistent with the predefined coupling rules in the template.

[0030] In this embodiment, the step of performing multi-scale process response decomposition on the dynamic change data of quality characteristics includes: Based on the process type, process parameter range, conveyor belt speed, and process dwell time, the basic scale, transient scale, delayed scale, and residual vibration scale are determined, and the window length and step size are set for each scale, wherein: The basic scale refers to the time scale used to describe the stable quality change process of meat samples under the influence of the main process in a single processing step. The scale covers the main interval of the process dwell time and is used to characterize the overall change trend of quality characteristics and the normal evolution characteristics within the process. Transient scales refer to the time scales used to capture the rapid quality response of meat samples at the moment of process switching or in the early stage of changes in process parameters. The scales mainly cover the short time interval before and after the process excitation mark, and are used to characterize the initial response, rapid rise or abrupt behavior of quality characteristics. The lag scale refers to the time scale used to characterize the delayed response characteristics of meat samples after process excitation due to the lag of physical or chemical processes. The scale covers the period from the start of process excitation to before significant changes in quality characteristics and its initial evolution stage, and is used to reflect the lag in response initiation and slow change process. The residual scale refers to the time scale used to describe the decline, oscillation, or gradual stabilization behavior of the quality characteristics of meat samples after the main effects of the process have been completed. The scale covers the attenuation and stabilization range after the peak and is used to characterize the post-peak evolution of quality characteristics and the residual effects of the process. Anchor point slices are made based on the process switching time points in the dynamic change data of quality characteristics. Time windows before and after are constructed at each scale. Time index alignment, drift correction and missing measurement filling are completed to obtain aligned data segments at each scale. In data segments of each scale, the rising segment, peak segment, decay segment and stable segment are divided, and the start delay, peak order, decay duration interval and cross-feature synchronization relationship are labeled to generate primary response components. The primary response components are structurally matched and temporally consistent with the process response reference template. Components that are not related to the process are removed, adjacent components of the same scale are merged, and the scale index, process index, intensity and duration are retained to output the response component set.

[0031] In this embodiment, the output of quality monitoring records, abnormal alarm signals, or production line linkage control commands includes: Receive the quality legality judgment results, process consistency judgment results, and dynamic response consistency judgment results, perform unified formatting on the judgment results, and form a set of judgment results that correspond one-to-one with the sample identifier, process identifier, and timestamp. A fusion weight and a judgment priority rule are set for the quality legality judgment result, the process consistency judgment result, and the dynamic response consistency judgment result, respectively. The fusion weight is used to reflect the degree of influence of each type of judgment result in the final quality judgment, and the judgment priority rule is used to determine the coverage order when there is a conflict in the judgment results, wherein: The fusion weight is a pre-set weight parameter based on the technical reliability and anomaly indication intensity corresponding to each judgment result. It is used to characterize the relative influence of the quality legality judgment result, process consistency judgment result, and dynamic response consistency judgment result in the final quality judgment. The fusion weight is configured according to historical production data and process stability during system initialization and remains unchanged within the same production cycle. The priority rule is that when there is a conflict between the judgment results, the judgment results are overridden in the following order: the quality legality judgment result takes precedence over the process consistency judgment result, and the process consistency judgment result takes precedence over the dynamic response consistency judgment result. When the high priority judgment result indicates an anomaly, it is directly used as the final quality judgment basis without waiting for the low priority judgment result to confirm. Based on the fusion weight, each judgment result is weighted and fused to generate a comprehensive judgment index. The consistency of each judgment result is checked in combination with the judgment priority rule to form the final quality judgment result. At the same time, the corresponding abnormality type identifier and suspected abnormal process list are generated. Output the final quality judgment result, record and store the final quality judgment result, the abnormality type identifier, the list of suspected abnormal processes and the corresponding judgment basis identifier, generate a quality monitoring record, and output the corresponding linkage control command to the production line control system when the final quality judgment result meets the preset abnormality triggering conditions.

[0032] Example 1:

[0033] To verify the feasibility of this invention in practice, it was applied to a pork production line in a chilled meat processing enterprise. This production line is mainly used for the large-scale processing of chilled pork, with a production cycle of approximately 450 pigs per hour. The production line includes pre-cooling, cutting, cleaning, cooling, weighing, and packaging processes. Due to the high continuity of production and frequent process changes, meat is easily affected by environmental changes, equipment fluctuations, and human operational deviations during processing. Traditional methods relying on manual inspection and single-point temperature monitoring are insufficient to detect quality abnormalities in a timely manner, especially for latent abnormalities that accumulate gradually across processes and abnormal responses during process switching, resulting in problems of identification lag and misjudgment.

[0034] This invention applies the real-time quality monitoring method for meat production lines based on multi-sensor fusion. Multiple types of sensors are deployed at key processing nodes to collect optical, temperature, volatile gas, and environmental parameters of meat samples during processing. All collected data are uniformly timestamped and process-identified, enabling continuous tracking of the state changes of the same meat sample across different processes. The system performs denoising, time alignment, and scale normalization on the raw data from the multi-source sensors to form a fused quality feature vector, which characterizes the overall quality status of the meat at the current process stage.

[0035] During production, the system constructs an inversion reachable domain for meat quality status based on the fused quality feature vector and the existing process parameter specifications of the production line. This domain describes the reasonable range of variation in meat quality characteristics within the legal process parameter range. The system continuously determines whether the current fused quality characteristics of a meat sample fall within this inversion reachable domain, thereby identifying anomalies where, although a single indicator does not exceed the limit, the overall quality status cannot be generated by a legal process. During continuous operation, the system successfully identified a quality status shift caused by a short-term spike in washing water temperature, an anomaly that would not have been immediately detected by traditional manual inspection.

[0036] The system constructs a cross-process quality evolution path based on the changes in the fusion quality characteristics of the same meat sample across various processing steps, and continuously analyzes the quality increments. Through consistency determination of the process quality conservation loop, the system can identify quality anomalies that gradually accumulate across multiple processes. For example, in a certain batch of production, the system found that the quality evolution path of the sample between the cutting and cooling processes deviated from the existing process rules, further pinpointing it as a problem of reduced cooling efficiency due to insufficient cooling air velocity, providing a basis for equipment adjustment.

[0037] During the process switching phase, the system invokes a quality evolution resonance detector to analyze the dynamic changes in meat quality characteristics before and after the process switching. Based on the process switching stimulus, the quality evolution resonance detector performs multi-scale process response decomposition on the dynamic change data of quality characteristics, extracts response structural features at different time scales, and matches them with the reference template corresponding to the process. The system determines whether the meat's response to the process change during the process switching conforms to normal processing patterns. In this embodiment, the system identifies an abnormal response caused by a fluctuation in refrigeration efficiency approximately 4 minutes after the cooling process switching, about 10 minutes earlier than manual detection.

[0038] The system ultimately integrates the quality legality judgment results, process consistency judgment results, and dynamic response consistency judgment results to output the final quality judgment result. It also links with the production line control system to mark abnormal batches and provide manual review prompts.

[0039] Table 1. Comparison of monitoring effects of the method of the present invention and the traditional method in actual production lines.

[0040] With a consistent daily sample size, the data in Table 1 objectively reflects the performance differences between the two monitoring methods. The method of this invention improves the detection rate of quality anomalies from 81.6% to 92.8% compared to the traditional method. By integrating multi-sensor data with process information for comprehensive analysis, it more comprehensively identifies anomalies occurring during meat processing, rather than relying solely on a single indicator or static threshold.

[0041] The method of this invention demonstrates a more significant advantage in identifying latent anomalies and process switching anomalies. The latent anomaly identification rate increased from 58.9% to 86.4%, indicating that by inverting the reachability domain and analyzing cross-process quality evolution paths, it can effectively detect quality deviations that accumulate over multiple processes but are not easily detected in a single process. The process switching anomaly detection rate increased from 62.3% to 90.1%, and the average anomaly detection time decreased from 13.2 minutes to 4.6 minutes, reflecting that after introducing dynamic response consistency analysis during the process switching phase, the system responds more promptly to abnormal conditions, which is beneficial for reducing the spread of quality risks.

[0042] In terms of stability and traceability, the method of this invention is also superior to traditional monitoring methods. The false alarm rate decreased from 5.8% to 3.1%, indicating that the fusion of multi-dimensional judgment results effectively reduced false alarms; the quality traceability integrity rate increased from 87.4% to 98.2%, demonstrating that this invention can more completely record the quality evolution process of meat samples in the production line, providing a more reliable data foundation for quality traceability and process optimization.

[0043] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for real-time quality monitoring of meat production lines based on multi-sensor fusion, characterized in that, include: Multiple types of sensors are installed in the meat production line to collect raw data from multiple sources. The raw data from multiple sources is preprocessed to obtain the fusion quality feature vector of the meat sample. Based on the fused quality feature vector, the inversion reachability domain of meat quality status is constructed. It is then determined whether the quality status of the current meat sample falls within the inversion reachability domain to obtain the quality legality judgment result. Based on the fusion of quality feature vectors and corresponding process identification information, each processing process of the meat production line is constructed into a directed process graph, forming a cross-process quality evolution path, performing process quality conservation cycle consistency judgment, and obtaining process consistency judgment results. Based on the process identification information, the process switching time point is used as the process excitation mark. Within a preset time window before and after the process excitation mark, the fused quality feature vector is continuously sampled to obtain dynamic change data of quality features. A quality evolution resonance detector is constructed, and multi-scale process response decomposition is performed on the dynamic change data of quality characteristics. Quality evolution resonance features are constructed, and structural matching and temporal consistency checks are performed between the quality evolution resonance features and the process response reference template. The dynamic response consistency judgment result is output. Based on the results of quality legality assessment, process consistency assessment, and dynamic response consistency assessment, the quality status of meat samples is integrated and assessed to generate the final quality assessment result, and output quality monitoring records, abnormal alarm signals, or production line linkage control instructions.

2. The real-time quality monitoring method for meat production lines based on multi-sensor fusion according to claim 1, characterized in that, The various types of sensors include optical sensors, temperature sensors, gas sensors, and environmental parameter sensors.

3. The real-time quality monitoring method for meat production lines based on multi-sensor fusion according to claim 1, characterized in that, The raw data from the multi-source sensors includes optical image data of the meat sample, temperature data of the meat sample and its surrounding environment, volatile gas concentration data related to the meat sample, and environmental parameter data reflecting the state of the meat production environment, including humidity data and airflow status data.

4. The real-time quality monitoring method for meat production lines based on multi-sensor fusion according to claim 1, characterized in that, The preprocessing of raw data from multiple sensors to obtain the fused quality feature vector of the meat sample includes: Denoising processing is performed on the raw data from the multi-source sensors to eliminate random noise and abnormal interference generated during the acquisition process; Time alignment processing is performed on the denoised raw data from multiple sensors to ensure that data collected by different types of sensors correspond to each other under the same time reference. The data that has undergone time alignment processing is subjected to scale normalization processing. Based on the multi-source sensor data after denoising, time alignment, and scale normalization processing, feature combination and fusion are performed to form a fused quality feature vector that characterizes the overall quality status of meat samples.

5. The method for real-time quality monitoring of meat production lines based on multi-sensor fusion according to claim 1, characterized in that, The process of obtaining the quality legality determination result includes: Based on the fusion quality feature vector and the corresponding process identifier, a list of legal process parameter ranges is established. For each processing step, the temperature, processing time, wind speed, relative humidity, spray flow rate, cooling medium temperature, conveyor belt speed and process dwell time are divided into ranges and the deviation length is set to form a process parameter library. Based on the physical constraints of the meat processing process, a physical constraint list is generated. The physical constraint list includes the consistency constraints of optical absorption and water content, the monotonicity constraints of temperature change and cooling intensity, the temporal constraints of volatile gas release and residence time, and the coupling consistency constraints between the three types of constraints. The list is quantitatively described in the form of upper and lower bound intervals and allowable deviation bands. Based on the interval endpoints and representative medians of the process parameter library, segment-by-segment deduction is performed. Combined with the physical constraint list, an upper reachable bound, a lower reachable bound, and an allowable deviation band are generated for each component of the fused quality feature vector. An inversion reachable domain is constructed. The inversion reachable domain consists of a kernel region, a buffer region, and an outer envelope region. The kernel region corresponds to combinations that are necessarily reachable, the buffer region corresponds to combinations that need further verification, and the outer envelope region corresponds to combinations that are not reachable. The current fusion quality feature vector is compared with the inverted reachable region. If all components fall into the kernel region, it is determined to be reachable. If any component falls into the outer envelope region, it is determined to be unreachable. If any component falls into the buffer, dynamic consistency verification and process sequence verification are performed based on the data of the nearest time window and the data of adjacent samples in the same batch. The final determination of reachability or unreachability is made, the quality legality determination result is output, and the determination basis identifier is recorded.

6. The real-time quality monitoring method for meat production lines based on multi-sensor fusion according to claim 1, characterized in that, The process consistency determination result includes: Based on process identifiers and timestamps, the positions of the same meat sample in each process are connected according to the sequence of the production line processes to form the original process path of the meat sample, with the entry process as the starting anchor point and the exit process as the ending anchor point. Based on the fused quality feature vector, a quality increment description item is generated for each process pair between adjacent processes in the original process path. The quality increment description item includes the direction of change, the magnitude of change, the rate of change, and the dwell time, and is associated with the corresponding process pair. One or more closed conservation loops are generated between the starting anchor point and the ending anchor point according to the process segmentation rules. The closed conservation loops use the heat-related dimension, the water content dimension, and the volatile precursor-related dimension as conservation dimensions. For samples with rework or skipping stations, loop segments or cross-segment connections are inserted into the original process path to form a set of cross-process quality evolution paths. A conservation cycle consistency check is performed on the set of quality evolution paths, employing a two-level process of fixed tolerance band verification and adaptive tolerance band verification, wherein: Fixed tolerance band verification performs closure checks on each conservation dimension based on the process preset tolerance; Adaptive tolerance band verification forms a batch reference band based on the distribution range of samples in the same batch and performs closure checks on each conservation dimension; Based on the process capability index, the quality increment description items are weighted and summarized to generate a closed deviation quantitative index and a list of suspected violation processes; Based on the closure deviation quantification index and the list of suspected violation processes, the process consistency judgment result is generated and output.

7. The method for real-time quality monitoring of meat production lines based on multi-sensor fusion according to claim 1, characterized in that, The acquisition of dynamic change data of quality characteristics includes: Based on process identification information and timestamps, the process switching time points and process parameter change information corresponding to meat samples are obtained from the production line control system, and the process switching time points are recorded as process excitation markers. For each process activation mark, a first time window length and a second time window length are set, with the first time window located before the process activation mark and the second time window located after the process activation mark; Within the first and second time windows, continuous sampling is performed based on the fused quality feature vector at a preset sampling period to generate a pre-sampling sequence and a post-sampling sequence corresponding to the process excitation marker. The pre-sampled sequence and the post-sampled sequence are aligned by time index, filled with missing samples and removed outliers, and a one-to-one correspondence with the sample identifier and process identifier is maintained to form dynamic change data of quality characteristics. The process excitation markers, process parameter change information, and quality characteristic dynamic change data are packaged into a process excitation data package and output.

8. The method for real-time quality monitoring of meat production lines based on multi-sensor fusion according to claim 1, characterized in that, The output dynamic response consistency determination result includes: A quality evolution resonance detector is constructed, which consists of an excitation parsing layer, a response decomposition layer and a structure verification layer. The excitation parsing layer receives process excitation data packets and generates standardized excitation sequences. The response decomposition layer receives dynamic change data of quality characteristics. The structure verification layer is used to match and verify with the process response reference template library. The excitation parsing layer parses the process switching time point and process parameter change information, generates an excitation index sequence that corresponds one-to-one with the sample identifier, process identifier and timestamp, and establishes a scale index according to process type and process parameter range; The response decomposition layer performs multi-scale process response decomposition on the dynamic change data of quality characteristics based on the excitation index sequence and scale index, and obtains a set of response components corresponding to different excitation intensities and durations. The response start time, peak time, decay termination time and cross-feature synchronization point are marked in each response component. The response decomposition layer extracts quality evolution resonance features from the response component set. The quality evolution resonance features include start delay, peak order, peak amplitude interval label, attenuation duration interval length, cross-feature synchronization relationship table, and stable segment ratio. The quality evolution resonance features are encoded according to process type, process parameter range, and scale index, and are associated with sample identifier, process identifier, and timestamp. The structural verification layer verifies the quality evolution resonance characteristics against the process response reference template library, and sequentially performs structural matching verification, temporal consistency verification, and cross-feature coupling consistency verification to generate dynamic response consistency judgment results.

9. The method for real-time quality monitoring of meat production lines based on multi-sensor fusion according to claim 8, characterized in that, The process of performing multi-scale process response decomposition on the dynamic change data of quality characteristics includes: Based on the process type, process parameter range, conveyor belt speed and process dwell time, determine the basic scale, transient scale, delay scale and residual vibration scale, and set the window length and step size for each scale; Anchor point slices are made based on the process switching time points in the dynamic change data of quality characteristics. Time windows before and after are constructed at each scale. Time index alignment, drift correction and missing measurement filling are completed to obtain aligned data segments at each scale. In data segments of each scale, the rising segment, peak segment, decay segment and stable segment are divided, and the start delay, peak order, decay duration interval and cross-feature synchronization relationship are labeled to generate primary response components. The primary response components are structurally matched and temporally consistent with the process response reference template. Components that are not related to the process are removed, adjacent components of the same scale are merged, and the scale index, process index, intensity and duration are retained to output the response component set.

10. The method for real-time quality monitoring of meat production lines based on multi-sensor fusion according to claim 1, characterized in that, The output quality monitoring records, abnormal alarm signals, or production line linkage control commands include: Receive the quality legality judgment results, process consistency judgment results, and dynamic response consistency judgment results, perform unified formatting on the judgment results, and form a set of judgment results that correspond one-to-one with the sample identifier, process identifier, and timestamp. A fusion weight and a judgment priority rule are set for the quality legality judgment result, the process consistency judgment result, and the dynamic response consistency judgment result, respectively. The fusion weight is used to reflect the degree of influence of each type of judgment result in the final quality judgment, and the judgment priority rule is used to determine the coverage order when there is a conflict in the judgment results. Based on the fusion weight, each judgment result is weighted and fused to generate a comprehensive judgment index. The consistency of each judgment result is checked in combination with the judgment priority rule to form the final quality judgment result. At the same time, the corresponding abnormality type identifier and suspected abnormal process list are generated. Output the final quality judgment result, record and store the final quality judgment result, the abnormality type identifier, the list of suspected abnormal processes and the corresponding judgment basis identifier, generate a quality monitoring record, and output the corresponding linkage control command to the production line control system when the final quality judgment result meets the preset abnormality triggering conditions.

Images